Support vector machine with adaptive parameters in financial time series forecasting.

Area of Science:

• Computational intelligence
• Machine learning applications
• Financial econometrics

Background:

• Support vector machines (SVM) are advanced learning machines with notable generalization performance.
• SVMs are increasingly applied beyond pattern recognition to areas like regression estimation.
• Financial time series forecasting presents unique challenges due to nonstationarity.

Purpose of the Study:

• To evaluate the efficacy of SVM for financial time series forecasting.
• To compare SVM performance against established neural network models.
• To investigate the impact of SVM parameters and propose adaptive solutions.

Main Methods:

• Comparative analysis of SVM, multilayer back-propagation (BP) neural networks, and regularized radial basis function (RBF) neural networks.
• Experimental investigation of SVM free parameter sensitivity.
• Development and application of adaptive SVM parameters incorporating time series nonstationarity.
• Utilized futures contract data from the Chicago Mercantile Market.

Main Results:

• SVM demonstrated superior performance compared to BP neural networks in financial forecasting.
• SVM exhibited comparable generalization performance to regularized RBF neural networks.
• SVM's generalization performance is significantly influenced by its free parameters.
• Adaptive SVM parameters led to improved generalization and reduced support vector usage.

Conclusions:

• SVM is a viable and effective tool for financial time series forecasting.
• Adaptive parameterization is crucial for optimizing SVM in nonstationary financial markets.
• The proposed adaptive SVM approach offers enhanced predictive accuracy and efficiency.