Modelagem e simulação computacional da combinação de preditores de séries temporais por meio de cópulas

Abstract

This dissertation disusses the problem of combining models for time series forecasting. In this context, the main characteristics and basic properties of combined models are presented. Some of the main methods of time series forecasting present in the literature are described. Computational experiments compare diverse copulas-based combined models. First of all, an algorithm is presented to combine predictions of the predictive models via Gumbel-Hougaard copula. In the second case, it is proposed a combined estimator constructed via non parametric Cacoullos multivariate functions. In the third and nal case of study, the main results of this dissertation are presented, in which an experiment that compares combined estimators constructed taking into account thousands of time series and numerous forecasting models were simulated. Thus, computational experiments show that the combined estimator constructed via copula obtained better results compared with the individual models and the linear combination method.