建構複合式信用評等模型

Abstract

風險評估及信用評等是金融機構用以評量借款企業償債能力的重要依據，然而身處目前經濟不景氣的環境下，逐漸身高的逾放比率使得越來越多的金融機構必須檢討其現有信用評等模式的缺失，以對貸款企業做出更正確有效的放款決策。現有之中、外文獻雖發展出許多信用評等模式來探討此問題，一般說來以類神經網路架構出的信用評等模型分類正確率表現較傳統統計方法建構出之模型為佳。但基於類神經網路(Artificial Neural Networks: ANN)理論上的不足，使得類神經網路架構出之信用評等模型解釋能力不佳，在實務層面上難以使用。因而本研究乃針對台灣金融機構之中小企業借款者，發展出一套複合式信用評等模型，此模型流程首先建立分類迴歸樹(Classification and Regression Tree: CART)，然後再將分類迴歸樹的預測結果及事後機率作為後續的類神經網路的輸入變數，藉此來增加整體複合式信用評等模型的分類正確率；此外，藉由使用分類迴歸樹來鑑別具有顯著影響的變數，增加整體複合式信用評等模型的模型解釋能力。 同時，本研究也廣泛比較現存的信用評等模型預測能力的差異，分別利用了線性判別分析(Linear Discriminant Analysis: LDA)、曲線判別分析(Quadratic Discriminant Analysis: QDA)、羅吉斯迴歸(Logistic Regression: LR)、機率類神經網路(Probabilistic Neural Network: PNN)、倒傳遞網路(Back Propagation Neural Network: BPN)、一般迴歸神經網路(General Regression Neural Network: GRNN)、自組性演算法(Group Method of Data Handling: GMDH)、K最近鄰居法(K-Nearest Neighbor: KNN)及學習向量量化網路(Learning Vector Quantization Neural Network: LVQ)等不同的信用評等模型，透過台灣某金融機構所提供中小企業借款者的實際歷史資料，驗證了本研究所提出之複合式信用評等模型確實有效可行。

Credit scoring is an essential task for banks and loan companies in the last few decades. The demand of developing a credit scoring model with reliable accuracy has become an urgent issue. Among many studies of credit scoring, artificial neural network (ANN) is a promising technique to achieve high accuracy of classification compared to existing conventional techniques. However, the poor explanation power makes ANN difficult to produce interpretable result. This drawback also decreases the power of ANN applied in practical problems. The objective of this study is to propose a hybrid credit scoring model which is combined with CART and other algorithms to enhance the accuracy of credit scoring model, and increase the interpretable capability as well. Financial loan companies can employ this study when establishing their credit scoring models.