Comparison of Classification Methods for Issuer Rating2

• Main Authors: Shih, Jen-Ying, 施人英
• Other Authors: Wun-Hwa Chen
• Format: Others
• Language: zh-TW
• Published: 1997
• Online Access: http://ndltd.ncl.edu.tw/handle/82004136785347902475

碩士 === 國立臺灣大學 === 商學研究所 === 85 === A credit rating is an opinion of the general creditworthiness of an obligor based on relevant risk factors. In the U.S., the credit rating has been practiced for more than a century and has played an important role in corporate capital raising, investment information providing for both individual and institutional investors, and bank''''s credit granting. However, until May 1997, Taiwan did not have its own credit rating agent. In this early development stage, it is urgent to have both academia and professionals participate in the design of a credit rating system for our financial market. For the past two decades, the thrust of the research in credit rating has been toward applying statistical methods such as linear regression, multivariate discriminant analysis (MDA), ordered probit and multinominal logit to the classification problem; however, they all require that the input data follow some distribution assumptions. In recent years, artificial neural network (ANN) has been proposed to overcome this drawback. Majority of these researches have tried to explore the applicability of neural network for the classification problem of binary nature, i.e., classifying bonds into either one specific rating or not. The main objective of this research is to design the neural network based rating system able to handle multi-class classification. To build the rating models, we made an extensive literature survey and consulted with senior analysts from Taiwan Ratings Corporation regarding S&P''''s rating criteria. Eleven financial variables are selected as the inputs for all the methods tested (with the exception for the airline industry model where five more operational performance variables are added); and the data were gathered from Standard & Poor''''s Global Sector Review Creditstats. Our empirical results showed that the multi-class model is significantly harder to solve than the binary-class model. Also, most of the past research applied their analyses on the data set across all industries. Here,we have tested our models on both the industry-specific data and the cross-industry data.Based on our empirical results, we can conclude some findings as follows: 1. When dealing with unstructured problems like credit rating, the neural network models seem to outperform both the MDA and ordered logit methods. 2. For MDA and ANN, the industry- specific rating models seem to outperform the cross-industry models. However, for the ordered logit, the results are not conclusive. 3. For all methods tested, the correctness ratios for binary classification are much higher than that for multi- class classification. 4. For the airline industry models tested, the addition of operational performance ratios significantly improves the performance of the ANN models 5. The number of hidden layers in the ANN is strongly affected by data complexity. When the data are of homogeneous nature, the addition of the hidden layers does not improve the performance of the model. To further improve the performance of the ANN models, it is essentialto consider other business risk factors.