Time-Series Prediction with Neural Networks: Combinatorial versus Sequential Approach

Abstract

In this paper, two different approaches of time-series prediction with neural networks are presented. The first is called combinatorial because it deals with a finite set of classes, obtained from the differences between several consequent function values. It is implemented through a modular neural network. The second describes time-series with interval functions or sequences of successive function values and is therefore a sequential approach, employing Kaiman neural gas networks. In the first case the future value (prediction) of an input vector depends on the classes (from input vectors, possibly together with the next values) obtained from learning the history of a time-series. In the second, based on the sequence of last input vector(s), the closest covering neuron (interval function) is defined, and is responsible for a future value calculation. A linear autoregressive method (AR) and multilayer perceptron (MLP) are used as references, and with the help of three different time-series, the efficiency of the suggested methods is given.