Research On Financial Distress Prediction Using Support Vector Machine

With the development of Chinese capital market, the number of listed corporates increased and scale of listed coporates became bigger. Meanwhile because the system of entry and exit became more reasonable,these operating failure corporates become mine for investors .Especially after the special treatment and the delisting regulation,the bankruptcy corporates perhaps make the investors lose a lot. Therefore the financial state of corporates which are the cornerstone attract concerns from various aspects. Whether theoretical fields or practical fields have been concerned the problem how to forecast the financial distress .The accurate prediction for financial distress can supply important references for investors. The managers can take actions to avoid the deterioration if they get the good prediction. The proper prediction can serve the debtors ,such as commercial bank .with warning indexes they manage the loans more efficiently. The regulatory section can benefit from the prediction to stabilize the capital markets.The prediction of financial distress has been being improved since the Altman's pathbreaking study. The development supplied more appropriate methods than lonely financial ratios analysis for investors, debtors, and banks .With the development of information technology,The multiple descriminant analysis(MDA) became easy as rule of thumb .However the limitation of MDA urged the researchers to seek better forecasting methods. Although Olhson applied Logit regression to financial distress prediction and improved the forecasting performance on the Z score method, the recent advances on intelligent skill made it possible to get more accurate forecasting of financial distress .The artificial neural networks(ANN) has been successfully used in financial distress prediction . The support vector machine(SVM) which based on statistical learning theory was proper for small sample learning. The new method could balance between the complexity of model and the generalization. It overcame the limitation of local optimization and overfitting for ANN. On the basis of a full review of theory of financial distress prediction, the present dissertation constructed the financial distress prediction models five years prior to bankruptcy using support vector machine .The main contents and contributions were outlined in the following paragraphs.Firstly,the dissertation presented the reasons of listed companies as financial distress corporates .Such companies approached closely bankruptcy companies which were used as samples in the financial distress prediction researches in some representative studies by authorative specialists. Additionally, the number of such corporates amounted to nearly 80,with the healthy companies which were chosen according to the matching rule there were enough samples to construct financial distress prediction model.Then the prediction variables were chosen. The prediction variables included not only the representative variable in some famous study but also the proxy variables embodied the corporates'occupied capital by some controlling shareholders. Four ratios were used to reflect the situation. The four ratios were accounts receivable-others to equity, receivable-others to total asset, (accounts receivable-others + accounts receivable ) to equity and (accounts receivable-others + accounts receivable ) to asset. Then the test of mean equality was carried out. The variables with significant difference between two groups included such variables that reflected the occupied capital. Next the principal component analysis also showed that such variables that reflected the occupied capital contributed to variance a lot .In succession the least squares support vector machine was used as tool and the variables with significant difference were used as input variables. Finally five financial distress models were constructed.the support vector machine recognized the holdin samples and holdout samples by find out the support vectors on the border. During the process the genetic algorithm was used to choose parameters including error penalty item and kernel parameters. At last ,being compared the prediction difference of accuracy and robust between the two methods .Because the support vector machine only depended on the support vectors on the border ,it efficiently overcame the overfitting problem. The support vector machine was good at the nonlinear classification.Although the multiple discriminant analysis was good at the linear classification , the samples must meet the strict limitations.The support vector machine outperformed multiple discriminant analysis in the accuracy by comparing the type I error and the type II error for estimated samples and test samples. The support vector machine also outperformed multiple discriminant analysis in the robust of financial distress prediction.by using the Levene test. From the misclassification cost's view, the support vector machine also outperformed multiple discriminant analysis.