'Enforced-Denial' Support Vector Machines for Noisy Data with Applications to Financial Time Series Forecasting
by
David Edelman
University of Wollongong

The Support-Vector Machine (SVM) approach of Vapnik has received much attention recently as being a simple and effective approach for pattern recognition and classification in problems with low-to-moderate noise component. While the strength of SVM's lies in extracting definition in 'hard-to-classify' cases at the 'boundaries' between categories, it has been noted that their error-penalty structure, which sacrifices modelling suitability in favour of computational tractability, is not ideal, and indeed tends to lead to inferior models in problems of high noise component. In the present paper, we explore a method for improving the error-penalty character of SVM's in the presence of noisy data, while preserving the computational tractability of the SVM approach. The method is demonstrated on a Time Series dataset of the Australian All-Ordinaries Index, with the results then compared to those of other traditional and nonlinear methods applied to the same dataset.

Date received: November 12, 2001

Copyright © 2001 by the author(s). The author(s) of this document and the organizers of the conference have granted their consent to include this abstract in Atlas Conferences Inc. Document # caid-72.