標題: 信用卡資產組合風險之研究
A Study of Predicting the Credit Risk of Credit Card Portfolio
作者: 羅文綺
Wen-Chi Lo
林進財
張保隆
Chin-Tsai Lin
Pao-Long Chang
經營管理研究所
關鍵字: 多元邏輯特迴歸;馬可夫鏈;信用卡;風險矩陣;信用風險;multinomial logistic regression;Markov chain;credit card;risk metrix;credit risk
公開日期: 2001
摘要: 隨著信用卡市場的競爭日益激烈及信用卡貸款的資產證券化、購併及風險值(VaR)研究等議題的興起,使得信用卡資產組合的信用風險預測的研究成為一項重要的課題。本文試圖導入摩根銀行所發展之信用風險矩陣(1994)觀念並整合Smith and Lawrence (1996)的研究,發展一套動態且在實務上可行的信用風險預測模型。本研究以國內某金融機構,民國89年10月至90年10月的持卡戶為研究對象,以電腦隨機抽樣方式,抽出35,000筆客戶資料,並以K-S檢定樣本之代表性後,觀察其連續十三個月的信用狀態變化,用以推估母體資料。研究方法以多元邏輯特迴歸分析(Multi-Logistic Regression)及非定態之馬可夫鏈機率模型(Markov Chain)等方法,發展信用卡資產組合的信用風險矩陣及風險預測模型,同時分析模型之適合度及預測能力。實證結果顯示,模型適合度及預測能力良好,命中率均在86%以上,資產組合預測的誤差水準也在4.7%以下。本文並對可能影響信用狀態進行卡方分析及變異數分析,檢定結果顯示,持卡戶的信用狀態,在年齡、性別、學歷、有無不動產、行業別與居住地等變數的影響下有顯著差異。在總體經濟變數上失業率、MIB等變數的影響下均有顯著差異。而信用狀態正常的持卡戶,其額度較大、信用額度使用率較低、預借現金累計額較低、循環息較低。上述因素在邏輯特的轉移機率迴歸式中的分析結果顯示對於信用狀態轉移的影響亦為顯著。
With the hot competition credit card market, and the issues of ABS (asset-backed securities) ,M&A (merge and acquisition) and the researches of VaR is raised, the risk management of credit card portfolio has become an critical problem to survive and be profited. This thesis attempt to integrate the concept of risk metrics developed by J.P Morgan bank (1994) and Smith & Lawrence (1996) researches and develop an dynamic model of the portfolio of credit card loan which can be used practically .The purpose of this thesis is to evaluate credit card portfolio prediction performance under customer’s profiles. We sampled randomly the monthly data sets of thirty five thousand credit card accounts form a credit card issuer dating from October 2000 to October 2001. The data sets include information form cardholders’ credit applications, monthly transactions, credit limits, balances, and detailed delinquent information. The dominated research methods included multinomial logistic regression and Markov chain. The evidence result shows that the fitting of model and the validation of the model are both well. The predicted accuracy of observed individuals reaches 86% and the prediction error of portfolio is below 4.7%.
URI: http://140.113.39.130/cdrfb3/record/nctu/#NT900457039
http://hdl.handle.net/11536/69043
Appears in Collections:Thesis