應用LISREL驗證借款企業之財務構面與構建信用評等模式

Abstract

隨著國際金融市場環境的快速發展，企業面臨的風險比以往更為 複雜，因此銀行或金融機構極須建立一套有效的信用評等模式(credit scoring model)，以準確地判斷借款企業是否會違約或預估發生違約的機率，以做為銀行或金融機構對借款企業訂定授信放款策略及違約催討策略之依據。然而，中外文獻在建構銀行或金融機構放款給企業所用之信用評等模式時，所使用之財務構面大多是應用因素分析(factor analysis)從一些財務比率變數中萃取而得，然後再依據這些財務構面，應用一些統計方法或類神經網路來建構一階段信用評等模式。由於財務構面與財務比率變數之間的相關性可能會隨著貸款期間的長短而有強弱之差別，若在構建信用評等模式時，未將此相關性的變化納入考慮，則由因素分析所萃取出之財務構面及依據這些構面所建構之信用評等模式其預測準確性可能不高。此外，由於銀行或金融機構貸款給一個違約者(defaulter)會比拒絕一個未違約者(non-defaulter)所付出之成本高出許多，因此提升信用評等模式預測違約案件之準確率對銀行或金融機構而言，是非常的重要。然而，實務上在建構信用評等模式時，由於違約案件(default cases)通常較未違約案件 (non-default cases)少很多，因此應用此類不平衡資料構建信用評等模式時，預測違約案件之準確性通常較預測不違約案件差很多。為改善此不平衡資料所構建信用評等模式預測違約案件不準確之問題，目前雖已有文獻運用兩階段分類方法來建構信用評等模式，並證實兩階段信用評等模式會較一階段分類方法所建構之信用評等模式之整體預測效果佳，但文獻上所提出之兩階段信用評等模式仍然無法有效提高預測違約案件之準確率。本研究之主要目的是針對上述萃取財務構面未考量貸款期間之不同以及不平衡資料所引起預測違約案件不準確等兩項缺失分別提出建構一階段及兩階段之信用評等模式之方法。本研究建構一階段信用評等模式之方法是應用線性結構方程模式(linear structure relation, LISREL)，根據所投入的財務比率變數，在不同的貸款期間下，分別找出適合的財務比率變數來建構財務構面，然後再利用所萃取出的財務構面及Cox模式來建構一階段信用評等模式。對於建構兩階段信用評等模式方法上，本研究在第一階段利用由LISREL 於不同貸款期間下所建構之適當財務構面及Cox模式將資料分為違約、未違約與不易判別區域等三類；在第二階段則利用支持向量機(Support Vector Machine, SVM)針對不易判別區域之借款企業，再建構第二次階段之信用評等模式。本研究最後利用台灣中小企業借款者之財務資料，驗證了本研究所建構之一階段及兩階段信用評等模式確實能反應借款企業貸款期間越長，銀行或金融機構所面臨的風險越大之現象，因此可同時提升未違約及違約案件之預測準確率。

As the international finance has been developed fast, the risk that enterprises are facing now is more complex than before. Thus, banks and financial institutions must develop a credit scoring model to effectively predict default probability and assess borrower default risk. Banks or financial institutions can utilize the results of credit scoring model to devise appropriate loan strategies for borrowers to reduce risk or losses from improper loans. Previous studies on constructing a credit scoring model, the financial dimensions were generally extracted from financial ratio variables using Factor Analysis (FA) without considering various loan periods. These dimensions were used to construct a one-stage credit scoring model using statistical models or neural networks. However, the relationship between financial dimensions and financial ratio variables may vary according to various loan periods. If various loan periods are not taken into account when constructing the one-stage credit scoring model using financial dimensions extracted by FA, the predictive power of the model may not be high. In addition, because the cost of granting a loan to a defaulter is much larger than that of rejecting a non-defaulter, enhancing the accuracy rate of defaults cases is urgently important for banks or financial institutions. However, in practice, default cases are usually much less than non-default cases when constructing a credit scoring model. Thus, the effect on predicting default cases is usually much smaller than that of non-default cases when constructing the credit scoring model using the imbalanced data. Now, although some studies constructed the credit scoring model using a two-stage classification method to improve the problem of predicting default cases caused by imbalanced data. These studies claimed that the accuracy rate of the credit scoring model constructed using a two-stage classification method is higher than that using a one-stage classification method. However, using two-stage classification methods to construct the credit scoring model still can not effectively increase the accuracy rate of defaults cases. Thus, this study proposes methods of constructing the one-stage credit scoring model and the two-stage credit scoring model, respectively to overcome the problems mentioned above. In constructing a one-stage credit scoring model, linear structure relation (LISREL) is utilized to find proper financial ratio variables to construct financial dimensions for various loan periods. Then, the constructed financial dimensions and Cox model are utilized to construct a one-stage credit scoring model. In constructing a two-stage credit scoring model, this study constructing the two-stage credit scoring model is composed of two stages. The fist stage is using the constructed financial dimensions and Cox model to divide the borrowers into three classes: default, non-default and undetermined borrowers. In the second stage, the data in the undetermined class are utilized to construct a classification model using Support Vector Machine (SVM). Finally, this study using the financial data of borrowers from the small -and-median sized enterprises in Taiwan to demonstrate that the proposed one- stage credit scoring model and two- stage credit scoring model can reflect that when an enterprise borrower’s loan period is longer, the risk that a bank or a financial institution must face is higher. Also the proposed methods effectively enhance the accurate rates of both default and non-default cases.