Comparação de diferentes métodos de previsão em séries temporais com valores discrepantes

Abstract

Forecats are important when one is faced with phenomena whereupon uncertainty is present. There are various methods of forecasting and each one has its applicability, depending on the kind of series that one wants to study. In time series the sample is a set of observations ordered in time, and the order wherein the measurements are made cannot be ignored. The focus of this work is to study different methods of forecasting in time series where there are outliers. In this context, we propose a method of forecasting based on Moving Median, for constant level series, based on the fact that the median is not influenced by outliers. The performance of this method was compared to that of the Simple Exponential Smoothing methods, Simple Moving Average and a linear combination of Moving Average and Moving Median, through simulations on simulated data sets and real data. The parameter estimation was made using the function optimize() and implementing the function find_n in the software R. The residues were analyzed and compared, in order to seek the method that minimizes the error, in other words, the most appropriate method for the studied series.