Métodos de seleção automática de omdelos ARIMA no software R: uma comparação dos algoritmos do pacote forecast

Abstract

In time series analysis, forecasting is one of the most important applications. In forecasting, the accuracy of the method used is one of the critical factors in its adoption. One of the main available forecasting methods is ARIMA model, whose downside is the complexity to specify the most adequate model. Currently, with the increase in computational power and in the availability of statistical software, it is already possible to find fully automated solutions, such as the “forecast” package written in the R programming language. The package has two algorithms for specifying models: exhaustive and stepwise search. The latter is an option to shorten computation time. The present work aimed to test whether there is a significant difference between the accuracy of point forecasts generated by both algorithms. For that, 2,829 annual, quarterly and monthly series of the M3 Competition were selected, predictions with the two methods were executed and their accuracies were calculated. The conclusion was that for the annual and quarterly series, the methods had on average the same degree of precision and for the monthly series, the exhaustive procedure was on average slightly better.