應用支援向量機於汽車貸款違約之預測

Abstract

近年來接連發生的金融危機迫使我們尋求更好的信用評等方式。在經濟發展 的過程中，汽車貸款扮演相當重要的信用借貸角色，對於消費者銀行與汽車車商 而言，判斷汽車貸款違約率是相當的重要。對放款者而言，一方面必須承擔經過 核准的汽車貸款違約所產生重大的損失；另一方面，若錯誤拒絕可以接受的貸款， 放款者則會喪失利益。 過去數十年間，許多研究針對信用評估議題提出新的模型，包括羅吉斯迴歸、 專家系統、資料探勘等。本研究利用某汽車貸款公司所提供的資料，應用「支援 向量機」分析資料，以建立汽車貸款違約之預測模型。 本研究藉由「格子點搜尋演算法」調整核心核數，並以「過濾法」與「包裝 法」進行特徵篩選，結果發現以五個特徵值所組成的分類器具有最高的預測分類 率，可達77.43%，並提出向後逐步選取法為主的羅吉斯迴歸模型，其預測分類率 可達75.70%。研究結果顯示：(1)就預測分類能力，支援向量機較羅吉斯迴歸佳。 (2)考量過多的資訊不一定能夠產生較佳的預測分類力。

The recent financial crises have called for better credit evaluation. Traditionally, car loans constitute an important portion of credit lending for an economy. Determining the default probability of a car loan is a major task for consumer banking and automobile sales companies. On the one hand, the lender will suffer a loss if a granted car loan eventually defaults; on the other hand, the lender will lose a potential gain if a good loan prospect is mistakenly rejected. Previous studies have developed models for credit evaluation. These models include regression analysis, expert systems, and data mining techniques. Using data from an automobile company, this study applies support vector machines to build classifiers for car loan default prediction. By adopting the grid search approach to adjust kernel parameters and using the filter method and wrapper method to select features, this study find that a classifier with just 5 features possess the best classification power. The classification accuracy rate is 77.43%. By comparison, a step-wise logistic regression model has a classification accuracy rate of 75.70%. The results show that (1) support vector machines have better classification power than logistic regression, and (2) considering more factors does not necessary result in better classification.