隨著消費型態的改變，信用卡成為主要支付的工具，消費性貸款因而成為銀行業者主要收入來源之一，但近年因次級房貸導致金融海嘯的發生，使業者必須更重視事前信用評等以及事中的行為評等工作。國內信用卡流通數至97年底達三千多萬張，雖然總卡數有減少趨勢，但過去所引發的問題仍無法徹底解決，且銀行發卡前之信用評等只能做到防範工作，而後續持卡期間之繳款行為更是銀行應注意的事項。因此，本研究使用類神經網路、CHAID以及存活分析，進行二階段行為評等模式建構，針對國內某家發卡銀行之顧客資料進行實證分析，於第一階段使用類神經網路及CHAID分析出顧客還款行為類別，並選擇整體鑑別率較高之分類模式進行第二階段模式建構，第二階段試圖利用存活分析提供之違約機率，針對第一階段較佳模式所分類之同質性顧客予以差異化。研究結果顯示，第一階段以類神經網路所建構之分類模式比CHAID所建立之整體鑑別率較佳，並於第二階段建構中將類神經網路判別之同質性顧客差異化後，發現存活分析較可找出後續還款行為之重要變數及未來期間內違約之機率，因此，本研究所提之二階段行為評等模式，對於顧客未來還款狀態不僅擁有較佳之鑑別能力，且可幫助業者預先掌握經分類模式判別為不同還款狀態下，個別顧客於未來期間還款之違約機率，而使業者於針對個別顧客之信用風險上，如催繳帳款、調整信用額度等動作時做出正確之判決，或未來經營策略上獲得更重要的線索，並提升往後管理決策之成效。 For the evolution of consumption behavior, credit cards have become one of main tools of paying. Banks start to be aware of credit risks associated with delinquency, so they make effort to build up a variety of models selecting good clients before approving their application for credit cards. We have learnt that there are a lot of researches providing many models to solve the problems, whereas the majority of those are related to credit scoring and minority related to behavioral scoring. Therefore, in this study we propose a two-stage behavioral scoring model and use artificial neural networks (ANNs) and chi-square automatic interaction detection (CHAID) as the scoring classifying methods, then we use survival analysis to estimate the different probabilities of each cardholder among different groups predicted by ANNs. The result shows that the ANNs model has a better classification accuracy rate than CHAID model as we expected at the stage one. At the stage two, survival analysis provides not only the important variables of delinquency but the default probability of each cardholder. For banks, they can use this two-stage model to classify clients. Also, they can use the additional information to circumvent the credit risk and improve the efficiency of the decisions on credit-grading strategy.
