Study On Application Of Support Vector Machines In Financial Problems

Based on the knowledge-driven datamining, this paper proposes methods for the modification of support vector machines (SVM) to acclimatize them to the characteristics of financial data and then explores applications of the modified SVMs to financial problems, with the essence narrated as follows:First, the theory of support vector machines and their limitations in applications to financial data mining are expounded. Except for the definition, modelling principles, categories and corresponding arithmetics, the structural risk minimization (SRM) principle with its intrinsic nature and advantages is specially emphasized. But, conventional support vector machines still have the existing defects mainly being short of adapation to the characteristics of financial data, such as nonlinearity, instability and high noise, etc., which leads to the lack of adaptability and robustness of models to samples. Of course, all of these limitations will constitute next objectives and tasks for modification of support vector machines.Secondly, from the point view of model selection for support vector machines, the existing approaches to modifying the importance of sample and feature are innovated respectively and simultaneously for the meta-parameters, C andε, with reasonable weighting methods correspondingly investigated, to intensivieren deeply the models considered in the adapation to financial data. Based on the discounted least squares (DLS) ideology, the probility estimation approach considering the characteristics and contents of the information is introduced to modify the exponential weighting function in order that appropriate weights can be assigned to samples. At the same time, another weighting ideology, so called the derivative-based saliency analysis (DB-SA), is used to assign features of input vectors according to their correlation to the decision function. Meantime, all of the two ideologies discussed above are deeply integrated to assign not only the samples but also the features to obtain better model robustness and generation capacity, with the proposed model called the hybridly weigthied support vector machine (HW-SVM).Thirdly, support vector machines are outside integrated to other machine learning methods and statistical approaches to take advantage of all of their capabilities and seize more unique information in financial data. These methods include copula functions and Wavelet neural networks (WNN), with mixed nonlinear systems constructed. All of these integration methods will enable support vector machines outside to mine information from various angles and beddingplanes and to depict the essential characteristics and feature variables of the financial market, with continuously evoluting complex mechanism caught.In order to test the modified support vector machines proposed, several empirical analysis concerning their applications in financial problems are conducted, such as the early-warning of corporate financial distress, the pricing of options and convertible coporate bonds, the determination of hedging ratio and the analysis of morphostructure and approaches of the connection between financial markets. All the results of the empirical analysis demonstrate that the models proposed are satisfactory in their generation performance.This work is sponsored by the national natural science fund (NO.70971079).