PREDICTING FINANCIAL DISTRESS USING REAL-VALUED GENETIC ALGORITHM WITH SUPPORT VECTOR MACHINE: THE CASE OF US CORPORATIONS

• Main Authors: TANG,JYUN-PING, 湯竣評
• Other Authors: LYINN CHUNG
• Format: Others
• Language: en_US
• Published: 2006
• Online Access: http://ndltd.ncl.edu.tw/handle/59442085707122749212

碩士 === 國立臺北大學 === 統計學系 === 94 === Due to the finance liberalization and internationalization, the financial institutions must be very competitive in order to survive these days. There are many factors to cause enterprises to have financial distress and go past a time limit to load money; i.e. wrong to make policy and wrong investment environment during the economy depression in recent years. Establishing financial warning system becomes important topics for discussion before financial institutions to make loads due to the announcement of the new Basel Accord. Therefore, the aim of this study is to apply the commendable idea that combines the real-valued genetic algorithm(RGA) with support vector machine(SVM) to establish and predict the model(RGA-SVM) of the financial distress in North America. This study also uses the SVM method, traditional artificial intelligent method(Neural network) and statistical method（Logit regression） to compare with the RGA-SVM method. The experimental results of this study show that the SVM model unquestionably needs an optimum method（e.g.,RGA） to optimize its parameters and to stabilize its model. The comparisons of predictive accuracy reveal that the new AI method (SVM and RGA-SVM) has certainly superiority than the traditional AI method(NN) and the statistical method(Logit). In order to examine the influence of the variable selection to the financial distress prediction, a full model was compared with a reduced model. The results reveal that the predictive accuracies of the models with original variables or selected variables were similar. Finally, this study applies a proposed method of variable contribution measures to measure the variable contribution of a full model and a reduced model based on the NN model respectively. The experimental results indicate that the model with selected significant variables may evaluate the variable contribution more positive than the models with excess input variables（original variables）.