标题: 金融机构对中小企业放款风险评估与信用评等整合模式及放款决策之研究
Integrating Credit Scoring and Risk Assessment Models and Constructing Mortgage Loan Policy for Small and Medium Enterprise
作者: 唐丽英
TONG LEE-ING
国立交通大学工业工程与管理学系(所)
关键字: 信用评等;偿还率;风险评估;抵押放款;放款决策;类神经网路;罗吉斯回归;理想度类似顺序偏好法;自主性演算法;复回归分析;主成份分析;分类回归树;Loan Policy;Credit Rating;Recovery Rate;Risk Assessment;Logistic RegressionAnalysis;Neural Network;TOPSIS;GMDH;Multivariate Regression Analysis;Principal Component Analysis;Classification and Regression Tree
公开日期: 2008
摘要: 信用评等是金融机构用以评量借款企业偿债能力的重要依据。由于国内近年来倒帐
事件频传,导致银行对中小企业逾放比居高不下,且“新巴塞尔资本协定”(New Basel
Capital Accord)首度明文规定允许银行使用内部信用评等方法来衡量借款企业之信用风
险,因此,金融机构建构一套有效的信用评等模型已迫在眉睫。目前中外文献利用统计
方法(Statistical methods)、无母数方法(Nonparametric methods)以及类神经网路(Artificial
Neural Networks)来探讨此问题,这些方法所构建的模式中以类神经网路架构出的信用评
等模型分类正确率表现较其他方法建构出之模型为佳,但由于类神经网路的运算过程为
一黑盒子,且无法找出重要之解释变数,使得类神经网路架构出之信用评等模型在实务
应用效果不佳。由于过去中外文献大多是提供一套判别信用等级之模式,很少文献提及
如何针对违约样本之存活期大小(即借款者于发生违约前依约还款之期间长短)构建一套
存活期预估模式,以瞭解其可能发生违约之严重程度,这些文献中有些假设不符合实务
状况,亦或是需进行适配度检定(Goodness of Fit Test)来验证存活期资料是否服从某种分
配,使得文献中所提之存活期预估模式不恰当或应用不易,因而导致放款机构决策者较
无法瞭解关于放款企业可能还款期长短及担保品多寡的资讯。放款机构在建构出信用评
等模式之后,可依其放款对像做出多等级之判别,以增加其放款之弹性。然而如何决定
各等级之放款策略,使放款机构能在放款前决定借款者需提供多少担保品或是应设定多
高之利率,则是放款机构在放款前须考虑之课题。此外,近年来由于企业倒闭违约事件
频传,银行及放款机构之逾期放款比率更趋恶化,造成逾期后偿还之比率逐年降低。因
此银行或放款机构针对偿还率逐年下降的问题,急需建立一套完善的违约后偿还率评估
模式,以做为银行制订客户违约后催讨决策之依据,并降低违约后之损失。因此本计画
欲建立一个信用评等与放款策略及催收策略之整合流程,本计画共分三年完成,第一年
将利用罗吉斯回归以及类神经网路方法发展出一套复合式违约预测模型,此信用评等模
型首先建立罗吉斯回归模式(Logistic Regression),然后再将罗吉斯回归模式的预测结果及事后机率作为后续的类神经网路的输入变数,藉此来增加整体复合式模型的分类正确
率;此外,藉由使用罗吉斯回归来鉴别具有显着影响的变数,可增加整体复合式模型的
模型解释能力。接下来,利用自组性演算法构建一套存活期预估模式,然后再根据存活
期的长短作违约企业的等级划分,建立出多等级之信用评等模型;本计画第二年的工作
则是将每一个等级中的借款者的预估存活期以及财务及非财务变数,利用TOPSIS 法进
行决策分析以决定可使放款机构利润最大或是借款者违约时损失最小之担保品价值、借
款利率或契约期限的长度。本计画第三年的工作则是利用本计画综合借款者特质、信用
评估项目、产业指标及总体经济指标,纳入利用分类决策树及类神经方法之复合式模型
中,所建立之呆帐及结清组判别式,再利用复回归分析构建较现有文献更合理的偿还率
评估模式。最后,本计画利用偿还率预测值、风险暴险额、信用评估项目及总体经济等
指标构建风险等级模式,作为放款机构催收决策之参考指标。本计画将利用国内某金
融机构所提供近几年借款企业之历史资料验证本计画所发展之整合流程确实能
有效地判别借款企业之信用等级。
Due to the increasing global competition, gathering sufficient capital has been a very
important issue to many banks or loan companies. For this reason, credit rating become a
crucial task for banks or loan companies in the recent decades. Consequently, developing a
reliable credit rating model has become an urgent issue for loan companies. Among many
studies on credit rating, many methods such as statistical methods, nonparametric methods,
and artificial neural networks (ANN) are used to construct credit scoring models. Among
these methods, ANN has been proven to have greater predictive power as compared to other
credit rating techniques. However, ANN still has some drawbacks such as “black box
procedure”, “lack of explanation”, “complex network design”, “lack of feature selection”etc.
These drawbacks make ANN difficult to be used in practice. Some studies employ survival
analysis to construct a credit rating model, but these methods require that data must possess
certain distributions. After classifying the loan applicants by credit scoring models, banks or
loan companies need to make the loan decisions. Due to economic recession and unstable
financial market, some companies may not be able to make on-time payment after receiving
mortgage loan. In this case, forecasting the recovery rate of these companies to collect the
defaulted loan becomes an very important issue.
This three-year study is divided into three phases. In the first year, logistic regression and
ANN are used to construct hybrid models. First, logistic regression is utilized to select
significant input variables. The chosen variables are then used as inputs for ANN to enhance
the total predictive accuracy. Group method of data handling (GMDH) is used to construct the
credit scoring model. The model can predict the survival rate of loan applicants which are
classified as bad loaners. In the second year, TOPSIS method is used to make the loan
decisions. The financial variables and non-financial variables of loan applicants are utilized as
the attributes to determine the value of collaterals, the rate of loans and time length of the
contract. In the third year, the variables such as the features of the borrowers, the terms of credit, the index of industry and the index of macroeconomic, etc. are utilized as the input
variables to construct the classification and regression tree (CART). CART is then used to
select significant input variables. The chosen variables are then used as inputs for ANN to
construct a model to classify whether the loans are uncollectible. A multiple linear regression
is also constructed to predict the recovery rates of the loaners. Finally, this study employs the
predicted recovery rate, the exposure at default, the terms of credit, and the index of
macroeconomic to construct a risk assessment model. A real case from a Taiwan’s loan
company is utilized to demonstrate the effectiveness of the proposed method.
官方说明文件#: NSC95-2221-E009-187-MY3
URI: http://hdl.handle.net/11536/102078
https://www.grb.gov.tw/search/planDetail?id=1600156&docId=274918
显示于类别:Research Plans