다분류 Support Vector Machine을 이용한 한국 기업의 지능형 기업채권평가모형

Investors, debt issuers, and others are quite interested in corporate credit rating (i.e. bond rating) because it is considered to be an important measure for managing financial risk of their portfolios. But, in spite of its importance, corporate credit rating is typically very costly to obtain, since it requires professional agencies to invest large amount of time and human resources to perform deep analysis of the company’s risk status based on various aspects ranging from strategic competitiveness to operational level details. As a result, it has been a popular research topic for researchers to predict companies’ credit ratings by applying statistical and artificial intelligence techniques The researchers in the early days mainly focused on applicability of statistical techniques such as multiple discriminant analysis (MDA) and logistic regression (LOGIT) analysis. However, more recent studies have shown that artificial neural networks (ANNs) achieved better performance than traditional statistical methods and other artificial intelligence methods in bond rating. Consequently, ANN has been the most widely-used technique for corporate credit rating for a long time. However, despite ANN’s superior performance, it has some critical limitations. First of all, it suffers from difficulty in selecting a large number of controlling parameters which include relevant input variables, hidden layer size, learning rate, and momentum term. And, it generally requires a large data set for the effective training. Furthermore, there is a danger of overfitting, and ANN usually requires huge amount of data samples for the effective training. In addition, it is usually difficult to explain why it produces a specific result, i.e. poor explanation ability. To mitigate these limitations, this paper suggests a novel machine learning technique, multi-class support vector machine (SVM), as a tool for corporate credit rating. General SVM is simple enough to be analyzed mathematically, and leads to high performance in practical applications. Although many traditional neural network models have implemented the empirical risk minimization principle, SVM implements the structural risk minimization principle. The former seeks to minimize the misclassification error or deviation from the correct solution of the training data but the latter searches to minimize an upper bound of the generalization error. In addition, the solution of SVM may be a global optimum, while other neural network models may tend to fall into a local optimal solution. Thus, overfitting is unlikely to occur with SVM. In addition, SVM does not require too many data samples for training since it builds prediction models by only using some representative samples near the boundaries (so-called support vectors). However, original SVMs were originally devised for binary classification. Thus, in order to apply them to multi-class classification problems such as corporate credit rating, the original SVMs should be extended to the multi-class SVM models. So far, researchers have proposed various approaches for this extension. In this study, we have tried three different methods for multi-class SVMs including (1) bound-constrained multi-class support vector classification (SVC), (2) multi-class SVC from solving a bound-constrained problem, and (3) multi-class SVC from Crammer and Singer. And, we try to find the most appropriate method and parameters for our data set. To examine the feasibility of multi-class SVMs in corporate credit rating, we applied these methods to the real-world bond rating case for Korean companies. We applied multi-class SVMs as well as ANN and MDA to the same dataset, so we tried to validate the superiority of multi-class SVM to other comparative algorithms. The experimental results showed that multi-class SVM outperformed ANN and MDA with a statistical significance level of 0.05 to 0.10. As a result, we found that multi-class SVM is a promising alternative to corporate credit rating prediction.