由於金融危機導致金融機構對於信貸風險較過去更加重視,並根據巴塞爾協議III,信用評等不僅影響到資本充足率,也影響金融風險管理。因此,在近年信用評等預測模型重於2類型分類,良好信用與較差信用,是受到限制。此外,多數研究信用評等預測模型多集中在資料探勘技術與演算法上,較忽略資料預處理的重要性與帶來的效益。 本研究採用兩階段方式建立信用評等預測模式,第一階段是資料預處理方法,包括變數選擇、集群、以及重採樣之方法;第二個是資料探勘技術,包括決策樹、貝氏網路、類神經網路、支持向量機、袋法、投票法。再來,選擇TEJ企業信用風險指標(TCRI)共9等級作為分類依據,並且收集41篇相關文獻,在從主要10篇文獻中選擇30個輸入變數。 由實驗結果可以得知,混合袋法與決策樹以及使用重採樣方法為最佳的信用評等預測模型,可達到82.96%預測率,也說明了兩階段的預測模型是優於一階段模式。在研究限制中,本研究由於時間限制只使用3種資料預處理的方法,建議未來研究可選擇更多的資料前預處理方法。
As the financial crisis resulted in unprecedented attention of financial institutions on credit risk, and under the New Basel Capital Accord, Basel III, the real and precise estimation of the unexpected loss not only affects the capital adequacy, but also has influence on the evaluation of risk management. As a result, the classification categories for credit rating concentrated on classifying 2 classes, good credit and bad credit from recent research, are limited. In addition, most research so focused on employing data mining techniques to construct models that they lost sight of the importance in data preparation i.e. data pre-process. This paper uses two-stage way, 1st stage is data pre-process methods including feature selection, cluster, as well as resample, and 2nd is data mining techniques comprising DT, BN, ANN, SVM, Bagging, and Vote, to construct prediction model for Taiwan credit rating. Taiwan Corporate Credit Risk Index (TCRI) from TEJ is used for experimental analysis, and the research collects 41 related studies and selects 10 noticeable papers to acquire 30 input variables; moreover, output variable is TCRI and has 9 classification categories. In experimental results, the contribution is not only research for the optimal model, Bagging (DT) with Resample method to achieve excellent accuracy 82.96%, as well as demonstrates that two-stage prediction model is better than one-stage model. The limitation is simply employs three kinds of data pre-process methods on account of needing more time to prepare, the future study could involve more features and research for more data pre-process methods such as backwards and another cluster algorithm.