Duas propostas para solucionar o problema referente a médias móveis nulas na normalização adaptativa

Abstract

This dissertation presents two new proposals to solve the problem of null moving averages for Adaptive Normalization, used as part of the pre-processing phase of a machine learning method for time series forecasting. They are: Adaptive Normalization by Compensation and Adaptive Normalization by Subtraction. It compares and analyzes the results obtained with these two new proposals, applied to 5 different datasets, with the Original Adaptive Normalization, as well as with other normalization methods present in the literature and the baseline ARIMA. It is concluded that, the Adaptive Normalization by Subtraction, proposed in this work, surpasses all other methods when it is applied to non-stationary heteroscedastic time series, also correcting the problem related to null moving averages of the Original Adaptive Normalization. For time series with high degree of stationarity, all the Adaptive Normalization methods are not satisfatory.