Recurrent neural networks for prediction of short and long memory time series

Author(s)

Daniel Takata Gomes

• Document type: Master's Dissertation
• Press: Campinas, SP.
• Institution: Universidade Estadual de Campinas (UNICAMP). Instituto de Matemática, Estatística e Computação Científica
• Defense date: 2005-11-18

Examining board members

Emanuel Pimentel Barbosa; Clélia Maria de Castro Toloi; Marcelo Cunha Medeiros; Peter Sussner

• Advisor: Emanuel Pimentel Barbosa

Abstract

Forecasting of time series is a topic of great interest nowadays. To do so, the data generating process needs to be estimated with a good degree of accuracy. In the last years, artificial neural networks are becoming more important in the statistical community. The more basic structure of a neural network, the feedforward neural nets, without feedback, can be a profitable alternative, in some cases, comparing to linear traditional models. However, some time series present characteristics that allow to introduce some kind of feedback in the network. Such network is called recurrent neural network, a tool not so popular in the statistical community.Two of the main concerns of this work are the study of recurrent neural networks for prediction of time series (theoretical fundamental, main architectures and learning algorithms) and the comparative study of the predictive performance of these networks for short and long memory time series. Spatial series (solo science data) and time series (inflation data) are used. The results show that the networks have good prediction performance comparing to linear models in several series. A new model (neural networks with varying coefficients) and its estimation algorithm are proposed, based in the data characteristics (AU)