Study On Regularized Machine Learning Algorithms And Their Application Of Financial Early-warning

Machine learning algorithms had been widely applied in building financial crisisearly-warning model. However, with the development of studies on machine learningalgorithms, the models that were constructed by machine learning algorithms had over-fittingand non-sparse coefficients problems. Therefore, the regularization technique of statisticallearning theory was introduced into the machine learning algorithms to build financialearly-warning models of regularized machine learning algorithms. A variety of improvedforms of Logistic Regression (LR) and Support Vector Machine (SVM) were put forward, andthe corresponding solution algorithms were presented. And the improved algorithms wereused for constructing failure prediction models.To begin with, the paper introduced the research background and significance of thestudy, and simply summed the research actuality of early-warning modeling methods at homeand abroad. Next, LR algorithm and SVM algorithm were systematically introduced, and78A-share manufacturing listed companies from the year2010to2012were selected as thesample companies. And,29financial indicators were selected and years of the sample datawere determined.Then, aiming at the LR model that had over-fitting and non-sparse coefficients problems,the financial early-warning model using LR with Smoothly Clipped Absolute Deviation(SCAD) penalty was proposed, and the solution algorithm was presented to solve the model.And using the financial samples data demonstrated that SCAD-LR model has moreadvantages. Moreover, aiming at the intractability of L1regularized LR problem, an efficientalgorithm based on interior-point method is designed and presented, and the results of thesimulation experiments showed that L1-LR model has better sparseness and betterclassification effect and the designed interior-point method is more superior.Furthermore, because the correlation between variables could not be well described byusing traditional SVM, so q-Gaussian function that was a parametric generalization ofclassical Gaussian function was put forward as the kernel function of SVM, and the financialcrisis early-warning model based on SVM with q-Gaussian kernel was proposed. To delete thenon-significant indicator variables, the significance test was conducted by using the samplesdata before the experiments. Next, the data of significant variables was used for the contrastexperiments, and the results showed that SVM with q-Gaussian kernel had much higheraccuracy of prediction and lower two types of errors than SVM with Gaussian kernel.Finally, to overcome the problems that traditional SVM was sensitive to outliers and had the large number of Support Vectors (SV) and the parameter of its separating hyperplane wasnot sparse, the financial early-warning model based on truncated hinge loss SVM with SCADpenalty was put forward. And an iterative updating algorithm was designed to solve thismodel. The experiments are implemented by using the financial samples data, and the resultsshowed that the model constructed by SCAD-TSVM algorithm was superior to the models ofother improved algorithms of SVM in the sparseness, the accuracy of prediction etc.