建立以決策樹為基礎之短期違約放款案件信用風險評估模型

Abstract

傳統的信用風險評估模型未考慮放款是一個隨時間進行的過程，僅關心借款者未來是否會違約，而非借款者何時會違約，這樣的判別結果不能提供管理者做出獲利最大化的決策。因為即使借款者違約，在某些情境之下金融機構依然可以從中獲利。目前的研究主流是使用 Cox 模式存活期模型建立時間信用風險評估模型，預測借款者違約的時點，以解決前述決策無法獲利最大化之問題。然而影響違約的因子眾多且複雜，目前的文獻對於哪些因子會影響違約發生的機率、以及如何影響，並沒有共識，Cox模式的預測結果因此並不夠準確。本研究提出一個以決策樹為基礎的時間信用風險評估模型，有別於以往預估每一筆案件違約時點的方式，本研究使用決策樹直接篩選出可能會在短期內違約的案件 (Short-term default)，目標是以高準確率找出會在短期內違約的高風險案件，這些高風險案件會給金融機構帶來巨大損失。為了改善決策樹不穩定，和放款資料呈現高度不對稱的情形對決策樹判別準確率的影響，本論文之信用風險模型還結合拔靴集成法 (Bootstrap aggregating, Bagging) 和增生少數類別技術 (Synthetic minority over-sampling technique, SMOTE)，用以提昇決策樹的判別能力。套用國內某金融機構的中小企業放款資料顯示，本研究所提出之風險評估模型，在判別高風險案件的準確率和精確率之上，都明顯優於現在廣泛使用的羅吉斯迴歸和Cox模式。

Traditional credit scoring models do not put time factor into consideration, only assess whether a customer will default, but not when. However, when making profit maximum decisions, managers of financial institutions can hardly rely on this kind of model. That is because even if a customer defaulted in the future, there is still possibility that the financial institution could gain profit from issuing the loan. Most of recent researches applied Cox proportional hazard model into their credit scoring models, predicting the time when a customer is most likely to default, to solve this problem. Nevertheless, the prediction provided by Cox proportional hazard model is not accurate enough. It is because there are vast amount of factors contribute to the timing of default, in a perplexing way. Today, there is no consensus among researchers on which indicators, and by which means, influences the probability of default. This thesis develops a decision tree based credit risk model. Unlike conventional approach, which gives individual duration, we filter out borrowers who tend to default in short-term, i.e., short-term default loans. These loans post the highest risk of significant loss to a financial institution. Our goal is to produce a highly accurate model which could distinguish high risk lending from others. This research integrate Bootstrap aggregating, or Bagging, and Synthetic minority over-sampling technique, otherwise known as SMOTE into the credit risk model, to improve the decision tree stability and its performance on unbalanced data respectively. Empirical result on real world small-and-median-enterprise loan data, provided by a local financial institution, shows that our credit risk model has good recall and precision on classifying high risk customers, out performs popular methods namely logistic regression and Cox proportional hazard models.