TIME SERIES PREDICTION

Abstract

The aim of time series prediction is to accurately predict the short-term evolution of the system. How to reach this aim receives a lot of attention recently. Many different types of prediction methods have been proposed and put into use. Among them, the method using artificial intelligence and machine learning techniques is no doubt one of the most promising methods. In this thesis, we present an approach to predict time series, which combines the concepts of genetic algorithms and competitive learning theories. Genetic algorithms are powerful and robust search processes. They can find good solutions to hard problem in a complex search space in a reasonable time. Competitive learning is an unsupervised learning paradigm. It provides a way to discover the general features of the input data, these features then can be used to classify or categorize the input data. In this way, the approach we proposed will be able to get one of the best classifying results for the input data and then used to predict the time series. To illustrate our work, we have established a prototype system to predict the time series generated by the daily closing bids of Singapore Unit Trust-The Commerce. It also adopts other methods such as adaptive local linear mapping, coefficient seeking algorithm and correlation seeking algorithm to improve the prediction accuracy. The experimental result indicates that this prototype system is effective to predict the chaotic time series (Singapore Unit Trust-The Commerce).