風險評估模型之最適重新取樣策略

Abstract

近年來經濟環境變遷快速，銀行或金融機構之放款風險隨之增高，致使國內銀行或金融機構對於企業之放款態度漸趨保守，造成許多的企業面臨資金嚴重短缺的窘境。中、外文獻通常使用統計方法或類神經網路來建構企業之風險評估模型，將借款客戶分成正常或違約(normal / default)兩類型。然而在利用上述方法進行風險評估時，雖然有不錯的整體預測正確率，但常會出現某類借款客戶（如：正常客戶）之預測正確率高，而對另一類客戶(如：違約客戶)之預測正確率卻偏低的情況。然而對於銀行或金融機構而言，違約客戶之預測正確率須相當高，風險評估模型才具實用價值。因此，本研究之主要目的是針對不平衡類別資料發展出一套最適重新取樣策略，然後再用由各類別重新抽取之資料量建構風險評估模型，以改善現有風險評估模型正確率偏向某一類客戶的問題，進而提升現有風險評估模型之預測正確率。本研究首先利用實驗設計與雙反應曲面法找出最適之重新取樣策略，再應用邏輯斯迴歸使用依照策略取出之資料量來建構風險評估模型，以供銀行或金融機構在放款給企業時，能制定出最佳之放款策略。本研究最後利用國內某金融機構所提供之企業借款歷史資料，驗證了本研究方法確實有效可行。

In recent years, the global economic environment is changing rapidly. Consequently, the financial risks of banks or financial institutions are also increased. Therefore, financial institutions such as domestic banks are generally conservative in their lending policies of enterprises loans and relied mainly on the availability of collateral. The most available risk assessment models use classification methods to construct the models and classify the loan customers into default and normal groups, respectively. It is often found that the accuracies of both groups are imbalanced, i.e. the accuracy of a particular group is significantly higher than another group. Those risk assessment models are not practically used by financial institutions, although they are having high average accuracy in terms of both groups classified. Therefore, this study aims to firstly determine the optimal re-sampling strategy using design of experiments (DOE) and dual response surface methodologies, and then use the amount of data in accordance with this strategy to construct the risk assessment model utilizing logistic regression to improve the imbalanced results. The effectiveness of the proposed procedure has been exampled with a real case drawn from a Taiwanese financial institution. The empirical results demonstrate that the average accuracy of the proposed study is better than that of the existing risk assessment models.