Uma avaliação de métodos de previsão aplicados à grandes quantidades de séries temporais univariadas

Abstract

Time series forecasting is probably one of the most primordial interests on economics and econometrics, and the literature on this subject is extremely vast. Due to technological growth in recent decades, large amounts of time series are daily collected; which, in a first moment, it requires forecasts according a fixed horizon; and on the second moment the forecasts must be constantly updated, making it impractical to human interaction. Towards this direction, computational procedures that are able to model and return accurate forecasts are required in several research areas. The search for models with high predictive power is an issue that has resulted in a large number of publications in the area of forecasting models. We propose to do a theorical and applied study of forecasting methods applied to multiple univariate time series. The study was based on exponential smoothing via state space approach, automatic ARIMA methods and the generalized Theta method. Each model and method were applied in large data bases of univariate time series and the forecast errors were evaluated. We also propose an approach to estimate the Theta coefficients, as well as a redefinition of the method regarding the number of decomposition lines, extrapolation methods and a combining approach.