建構整合式信用風險評估模型

Abstract

現今金融機構評估借款者之信用非常重視，目前財務風險資料大多屬於二元資料，分為違約案件(default)與非違約案件(non-default)。目前信用風險評估模型有非常多的種類，因不同資料特性對於相同風險評估模型不一定會有相同的效果，且目前的文獻當中，還未對資料特性與建模方法作相關的分析，所以金融機構大多以試誤法，選擇判別準確率較高的模型為最佳的信用風險評估模型。然而，當多種風險模型判別準確相近時，其判別結果不盡相同。因此，本研究針對當多種信用風險評估模型判別結果相近時，提出一套整合式信用風險評估模型，可根據不同信用風險評估模型之判別結果，建構模糊隸屬函數，產生新的判別結果。本研究以某金融機構之資料為例，使用自組性演算法、K-最鄰近法(KNN)、邏輯斯迴歸(logistic regression)結合模糊多數法則建構信用風險模型，結果顯示本研究提出之信用風險模型，在違約判別準確率、非違約判別準確率、整體判別準確率上，皆優於自組性演算法、K-最鄰近法(KNN)、邏輯斯迴歸(logistic regression)。

The importance of investigating the credit of borrowers is highly considered among current financial organizations. Most of the financial risks data in present distinguished into 2 different cases, which are: default and non-default. The credit risk assessment model is widely accepted in most of financial industries to evaluate the degree of risk impact of borrowers. There are a number of credit risk assessment models for any specific applied areas, most financial organizations construct numerous assessment models by history data. The effects of risk with different data character do not have the same behavior in the same model. Since the lack of analysis in the relations between data characters and model mythologies, most organizations use try-error method to classify the most accurate model as adapted model. While the accuracy in models do not have significant difference, the suggested decisions may different. This paper provides an integrated credit risk assessment model when the differences of multiple models are trivial. Based on the consequences of various models, formulating fuzzy membership function generates new decision result. An anonymous financial organization is as example in this research. Applied Group Method of Data Handing (GMDH), k-Nearest Neighbor algorithm (KNN) and logistic regression with fuzzy theory to construct the credit risk assessment model to increase the predicted accuracy of decision rate.