A Model Based On Support Vector Machine With Simulated Annealing Algorithms For Early Warning Financial Crisis

At present, the most modeling research still remain in the traditional discriminated analysis, the development of this method is relatively mature, the calculation is relatively simple, but the recognition rate is limited and poor portabilitied. With the development of economy and the accumulation of data and artificial intelligence, more and more machine learning methods are introduced to the field of financial crisis early warning. Support vector machine (Support Vector Machine, referred to as SVM) is a new generation machine learning technique based on statistical learning theory, has its advantages in solving the small sample, nonlinear, high dimension problems, widely applied to studies on financial crisis early warning.This paper selects the support vector machine for the early warning of enterprise financial crisis. At present, most of the literature focus on the early warning model establishment of SVM, ignores the method of selecting model parameters, and proper selection of parameters is directly related to the prediction accuracy and generalization ability. In view of this situation, this paper selects the simulated annealing algorithm (Simulated Annealing, referred to as SA) to find the optimal parameters of support vector machine, this method has been proved to be a global optimization algorithm to the global optimal solution with probability 1 convergence.This paper establishes the traditional model of SVM which to find the optimal parameters by using the simulated annealing algorithm on the basis of listing corporation financial data from 2009 to 2012 of the public offering and obtain the training accuracy, training time and the parameter values, then recognizes the test set and obtain the recognition rate.Then using the grid search algorithm to search the optimal parameters in the modeling process, through recognizes the same test set to obtain the recognition rate. The comparison of the results shows that the accuracy rate and the training time of the former is greater than the latter.