Time series forecasting using wavelet and support vector machine

Abstract

This thesis examines the use of multi-scaling capabilities of wavelet and support vector machine (SVM), for forecasting of time series. The basic idea revolves around using wavelet transform to decompose the original time series into separate series at different time scales before using them as inputs to SVM to improve forecasting. Redundant Haar wavelet transform was used due to its property of shift invariance, accurate alignment of events between scales and minimum boundary distortions, which have profound influence in forecasting performance. Two schemes were devised to integrate both wavelet and SVM into a system for forecasting. Tests were carried out to evaluate the performance of this system and compared with other models such as neural networks. It is found that wavelet pre-processing generally improves the performance of the both neural network and SVM. Both inputs characterisation and types of neural estimator has profound effect on the performance of time series forecasting.