摘 要
本研究從決策與認知的觀點，提出銀行信用評等模式由
銀行信用風險概念、有限理性決策及評等分類構成。以學習
專家智慧的策略，分成三個步驟建立這個模式。
銀行信用風險概念來自S&P、Moody’s 及 BankWatch
的銀行信用評等方法論文件，以紮根理論質的文件分析法，
得到101個概念，並系統化分類後得到CAMEL+架構。
有限理性決策代表因應認知能力有限及決策資訊不足時
，簡化決策的策略，僅考慮少數重要的銀行信用風險概念來
評等。BankWatch 除使用十五個財務變數外，經由對臺灣區
銀行評等報告進行內容分析，發現重要的非財務變數是主要
股東、業務差異及放款集中度三個概念。
評等分類係依據銀行在風險概念上的表現，分到不同的
信用等級。以BankWatch的台灣地區商業銀行及中小企銀為
對象，比較以下八種分類技術的效果：複變數區別分析的線
性或二次區別函數、Logit分析、Ordered Logit 分析，無
母數統計法的k 個最鄰近法、核心法、使用倒傳遞演算法
的類神經網路及有序類神經網路。類神經網路在預測能力、
模式風險及信用等級可靠度上，均較其他分類技術為佳。在
分類投資級及投機級的預測能力上，類神經網路能夠達到百
分之九十六的預測正確率。
由本研究的類神經網路敏感度分析，得到以下結果：
BankWatch評等最佳的銀行特徵重要的為問題資產比率低、
公營行庫、商業銀行、風險資產比率高。而評等最差的銀行
特徵重要的有資產報酬率低、中小企銀、呆帳提列率高、風
險資產比率低。

Abstract
From the aspects of decision-making and cognition science, this study
proposes bank credit rating model consisting of bank credit risk concepts,
bounded rationality and rating process. This model is constructed by
learning from credit rating expert’s knowledge.
Based on literatures of S&P, Moody’s and BankWatch and the
qualitative data analysis method of Grounded Theory, we extracted
101 key concepts for bank’s credit rating. We then analyzed these
concepts into categories, and constructed the CAMEL+ model.
Due to the limitations of cognition and information, experts overcome
bounded rationality by reducing the variables used in credit rating decision
making process. In addition to 15 financial variables used by BankWatch,
this study revealed three essential factors which are major shareholder,
business franchise and loan concentration.
Banks’ credit are rated according to the 18 variables. The sample
of these banks used in the study are rated by BankWatch. The performance
of eight classification techniques are compared. These eight
techniques are Multivariate Discriminant Analysis using linear and
quadratic discriminating functions, Logit Analysis, Ordered Logit Analysis,
K-nearest Neighborhood method, Kernel Estimation, Back-propagation
Neural Network and Ordered Back-propagation Neural Network. Neural
Network outperforms in predictive correct rate, cost risk and reliability.
Especially in discriminating invest level and non-invest level banks, Neural
Network can reach 96% correct rate.
Finally, based on BankWatch’s credit reports, a Sensitive Analysis
of Neural Network finds the important characteristics in these
high-rated banks are low problem asset ratio, government-owned,
commercial bank and high risk-weighted capital ratio. On the other hand,
the important characteristics in these low-rated banks are low return on
assets ratio, medium-and-small size business bank, high bad loan
provision and low risk-weighted capital ratio.