銀行建構「信用卡信用風險即時預警系統」之研究

A Research on Constructing a Real-time Alert System of Credit Card Issuing Bank

張文生(wen-sheng zhang)

指導教授 ： 嚴奇峰

中原大學/商學院/企業管理研究所/碩士(2001年)

https://doi.org/10.6840/cycu200100436

摘要

信用卡業務與銀行傳統放款業務相較，其本質上是屬於高利差、高風險且為全球化的產業，亦有頗具發展潛力的市場。面對這全球化競爭環境及當前景氣不佳前提下，現階段發卡機構的核心競爭力，應是如何做好信用風險管理（Credit Risk Management）以降低營運風險、減少呆帳損失，並追求利潤之極大化。

信用卡核卡後之信用風險管理是防範呆帳的重要防線。若能建立一信用風險預警系統，當持卡人拿卡從事高風險交易，在持卡人刷卡交易之同時，即能透過即時預警系統給予授權額度控管人員立即之警訊，進而採取凍結額度或拒絕交易等措施，即可減少呆帳並防止不當損失。

本研究以國內某家發卡機構為研究對象，從民國81年至89年所核發通過之持卡人資料總計162,173筆，以簡單隨機抽樣方式，抽出1,000筆為資料樣本，分為實驗組與對照組兩群，先將實驗組900筆，分別用SPSS以區別分析(Discriminant Analysis)及Logistic Regression建構風險預警模式；再將對照組100筆樣本代入已建立之預警模式，以驗證其正確率。研究發現Logistic Regression模式只需4個變數、預測能力達95﹪，與區別分析需5個變數、預測能力94﹪相較，模式較為簡單且正確率較高。兩種模式預測的正確性相差雖僅有1﹪，惟信用卡每年累積之簽帳金額龐大，以個案發卡機構為例，一年累積簽帳金額約為68億元，1﹪即可產生6,800萬元之誤差(約為個案發卡機構全年之獲利)，因之選擇Logistic Regression之結果為本研究之最適預警模式。

在實務的作法上，在建構即時預警系統時需與信用卡授權交易系統相整合，藉由發卡機構前端主機授權交易系統將持卡人基本資料、繳款狀、及消費項目等變數導入預警模式，預測此持卡人消費此商品產生違約倒帳之機率。若產生之預測值在設定之預警值(Alert)之內，則前端授權系統逕自核准交易。若在預警值之外，則系統將該筆資料送至發卡機構後端主機，由系統或人工判斷方式來處理，以產生即時預警效果給予授權額度控管人員，採取適當因應措施減少呆帳及不當損失。持卡人消費行為不斷在改變，發卡機構惟有隨時餵養違約倒帳之資料，讓即時預警系統藉由多樣變數之選擇不斷的學習、不斷的修正預警模式，方可達到更即時、更精確之效果。在目前先進的資訊及通訊科技下，建構即時預警系統是非常合乎成本效益考量的。

關鍵字

信用風險預警模式 ； 信用卡 ； Logistic Regression ； 區別分析 ； 即時預警系統

並列摘要

Comparing to the traditional banking loan, credit card business is characterized as high spread, high risk and globalized business with enormous market potential. While facing the global market competition and economic recession, the core competence for the issuing banks is to conduct credit risk management so as to minimize operating risk and to maximize profit.

Credit risk management is a very critical and helpful safeguard against the possible bad debts. Under a well-developed credit risk management alert system, the credit officers will, in time, be able to freeze the credit limit or to reject the related on-going transaction to prevent the potential bad debts and losses when the card holders try to conduct high-risky transaction, to over-utilize the credit limit, and/or to do the funding illegally.

This research defines the 162,173 cards issued by one of the issuing banks in Taiwan during the period of 1992~2000 as the population. And the sample size of 1,000 cards are collected by simple random sampling and were assessed by the basic data of card holders, the payment and historical records to analyze the reasons of incurred bad debts. Furthermore, according to the statistical results of Discriminant and Logistic Regression Analyses, the author tries to construct credit risk management alert models for the credit card and then choose one optimal model based on the accuracy of forecasting.

The captioned samplings of 1,000 card holders, 900 for experimental purpose and 100 for compared purpose. The experiment group was analyzed and to build the alert models by Discriminant Analysis and Logistic Regression using the SPSS software. And the comparison group was used to test these two models. The accuracy of Logistic Regression was scored at 95%, higher than that of Discriminant Analysis of 94%. Because the fewer variables was required and the higher accuracy for using the Logistic Regression, this analysis is chosen as the optimal alert model in this research.

The implementation of a real-time alert system is required to coordinate the authorization system of credit card business. The variables, such as basic data payment record and consumption item of card holders are input into the alert model built in the front-end authorization system to predict the possibility of bad debit associated with the related transaction conducted by the card holders. If the predicted value is located within the set-up alert value, the authorization system will automatically approve the transaction to speed up the authorization procedure for most normal transactions and to improve the transaction efficiency. If not, the front-end system will transmit the related data to the back-end system of issuing bank, the host system or expert judgment will take over for the exceptional or abnormal transactions by the credit control officers adopting appropriate measure to minimize bad debits and losses.

並列關鍵字

Credit Risk Management Alert Model ； Credit Card ； Logistic Regression ； Real-time Alert System ； Discriminant Analysis

參考文獻

23.黃小玉（民77），銀行放款信用評估模式之研究－最佳模式之選

5.李桐豪、呂美慧（民89），金融機構房貸客戶授信評量模式分析

11.徐健進（民74），銀行放款信用評估等模式之研究，國立政治大學企業管理研究所碩士論文，民國七十四年六月。

2. Espahibodi , P.（1991）,”Identification of Problem Bank and Binary choice Models” ,Journal of Banking and Finance，15，PP53-71。

3. Ingram , J. F. and Frazier ,E. L.（1982）,”Alternative Multivariate Testss in Limited Dependent Variable Models：An Empirical Assessment”, Journal of Financial and Quantitative Analysis ,Vol. 17,PP.227-240。