為了符合國際清算銀行在 2006 年通過的新巴賽爾資本協定，且有鑑於近年 來整體經濟環境欠佳，銀行業者面對外部的規定以及內部的需求，積極地尋求 信用風險模型的建置方法，希望將整個融資的評等過程系統化以提高對信用風 險的控管。 本研究希望利用 92 至94 年未上市上櫃中小企業之營造業的資料，依循新 巴賽爾資本協定之規定並配合資料採礦的技術，擬出一套信用風險模型建置與 評估的標準流程，其中包含企業違約機率模型以及信用評等系統的建置，前者 能預測出授信戶的違約情形以及違約機率；後者則是能利用前者的分析結果將 授信戶分成數個不同的等級，藉此區別授信戶是否屬於具有高度風險的違約授 信戶，期待能提供銀行業者作為因應新巴賽爾協定中內部評等法的建置，以及 中小企業的融資業務上內部風險管理的需求一個參考的依據。 研究結果共選出 5 個變數作為企業違約機率模型建立之依據，訓練資料以 及原始資料的AUC 分別為0.799 以及0.773，表示模型能有效的預測違約機率 並判別出違約授信戶以及非違約授信戶。接著，經過回顧測試與係數拔靴測試， 證實本研究的模型具有一定的穩定性。另外，透過信用評等系統將所有授信戶 分為8 個評等等級，並藉由等級同質性檢定以及敏感度分析的測試，可以驗證 出本研究之評等系統具有將不同違約程度的授信戶正確歸類之能力。最後，經 由轉移矩陣可以發現，整體而言，營造業在2003 年到2005 年間的表現有逐漸 好轉的趨勢，與營造業實際發展情形相互比較之下，也確實得到相互吻合的結 論。 / In order to conform to the New Basel Capital Accord passing in 2006 by the Bank for International Settlements and due to the slump faced by economies globally and the rise in the number of defaulters in the recent years, the banking industry has aggressively looked for ways to establish the reliable credit risk model that can accommodate required regulations set forth by the Accord as well as the internal banking procedure demands. The banking industry attempts to standardize the process of evaluating credit rating in regards to capital risk in the loan business to enhance the control of credit risks. The attempt of this research is to perform the process of the establishment and evaluation of the credit risk model which includes the default risk model of companies and the credit rating system within the framework of the New Basel Capital Accord using the statistical tool known as data mining. The data adopted in this study is taken from the construction industry of the SMEs from 2003 to 2005. The default risk model assesses the probability whether a company is at risk of being defaulted. In addition the credit rating system assigns credit scores to a company in question based on the application result from the default risk model to differentiate those who have high risk of being defaulted. More importantly this research provides banking industry of varying degrees of complexity to monitor its risk assessment as well as becoming a reference basis of the loan business in the SMEs. Based on the result of this study, five variables are selected as the default probability model basis. The AUC for the training data is 0.799 and for the raw data is 0.773 which represents the accuracy and reliability of the model in predicting the probability of default risk and determining the likelihood of the companies to default. After series of testing, our model stability plays a key role in determining whether the algorithm produces an optimal model in this study. The credit rating system formulates credit scores of the companies into 8 credit ratings. Applying homogeneity test and sensitive analysis, this study is able to verify the validity and accuracy of the rating system to correctly classify different levels of credit risk that could have jeopardized the companies to default. Finally, through the transformation matrix, there has been an improvement trend of performance in the construction industry from 2003 to 2005 which coincides with the result of this study.