Financial time series prediction using least squares support vector machines within the evidence framework

Area of Science:

• Computational finance
• Statistical modeling
• Machine learning

Background:

• Financial time series analysis often requires models that capture time-varying volatility.
• Least Squares Support Vector Machine (LS-SVM) regression is a powerful tool for nonlinear modeling.
• The Bayesian evidence framework offers a principled approach to model inference and selection.

Purpose of the Study:

• To integrate the Bayesian evidence framework with LS-SVM regression for financial time series and volatility prediction.
• To develop a method for inferring nonlinear models that account for time-varying market volatility.
• To automatically tune model parameters and select relevant inputs using model comparison.

Main Methods:

• Application of the Bayesian evidence framework to LS-SVM regression.
• Multi-level inference for hyper-parameter estimation and volatility modeling.
• Utilizing Mercer's theorem to derive practical expressions in the dual space.
• Model comparison for kernel parameter tuning and input selection.

Main Results:

• Successful inference of nonlinear models for financial time series and volatility.
• Construction of a volatility model within the Bayesian evidence framework.
• Demonstration of significant out-of-sample one-step-ahead prediction performance for T-bill rates and DAX30 closing prices.
• Validation using the Pesaran-Timmerman test statistic.

Conclusions:

• The Bayesian evidence framework effectively enhances LS-SVM regression for financial forecasting.
• The proposed method accurately predicts financial time series and their associated volatility.
• The approach provides a robust framework for model selection and parameter tuning in financial applications.