Feature Weighting of Support Vector Machines Based on Derivative Saliency Analysis and Its Application to Financial Data Mining

Abstract This paper proposes a novel feature weighting approach based on derivative saliency analysis, which can specifically display to what extent the output of support vector regression machines varies with the features (i.e. the components of the input vector). The empirical analysis of its application to option pricing demonstrates that the methodology proposed enables relevant features to be assigned right weights under given generation performance error criterion. At the same time, a comparison with some other methods frequently used for option pricing, such as the Black-Scholes equations and the traditional SVM models, is done. The results show that the feature weighting ideology proposed has advantages in not only generation accuracy but also robustness and stability and enhances the capacity of addressing the nonlinearity and nonstationarity of the financial time series.