A neural network based forecasting system

Abstract

The Neural Network (NN) arena has suffered in the past years a remarkable development as one of the novel fields for Artificial Intelligence (AI). Time Series Analysis (TSA) based on models of the variability of observations by postulating trends and cyclic effects, with a view to understand the cause of variation and to improve forecasting, suggests the use of NNs that do something like, what is called in statistics, Principal Component Analysis (PCA). The purpose of this work is to present a logical based NN system, along with: (i) Time Series Forecasting (TSF), with its characteristics of strong noise component and non-linearity in data, showing itself as a field in which the use of NN's stuff is particularly advisable; (ii) PCA rules, organized in a default hierarchy as logical theories, competing with one another for the right to represent a particular situation or to predict its successors; i.e., assisting in the process of choosing the best network to forecast each series. Some trials will be conducted, and the basic performance measures used as baselines for comparison with other methods.