消費金融信用評分研究

Abstract

本論文將焦點放在消費金融信用評分上，主要探討：(1)雙維評分模式改善行為評分；與(2)運用存活分析動態預估信用風險及決策。 第一部分是運用雙維評分模式改善行為評分。信用評分模型是統計模型應用在金融業上最成功的運用範例，業界積極發展信用評分模型於申請件之信用給予決策與既有戶之信用風險管理，然而，過去之研究大多以單面向信用評分模型來區隔客戶之風險。本研究針對既有戶的信用風險管理方面，除了使用銀行內部信用卡行為評分模型外，並結合外部聯徵個人信用評分形成雙面向信用評分模型，其正確率及區隔能力皆優於單面向的銀行內部信用卡行為評分模型或外部聯徵個人信用評分模型，得到更為精準之風險判斷與區隔，進而找出其資產組合中，應特別加以管理或監控之部分，設計相對應之因應策略與措施，對銀行實務上可產生極大的效益。 第二部分是運用存活分析動態預估信用風險及決策。多數的信用風險模型採用區隔分析、類神經模型、決策樹及邏輯斯迴歸模型，為在一定期間內預測客戶是否會發生違約的機率，但無法預測一定期間內各個時間的違約機率，本研究採用Cox(1972)提出的觀念“存續時間” (Duration)，以此概念發展的模型即為存活分析，常見的運用領域為臨床醫學與工業工程等，以Cox(1972)的比率危險模型(Proportional Hazard Model, PHM)，找出信用卡戶的人口統計或行為因子中有顯著影響力者，來估計客戶一定期間內各時點的違約機率(或存活機率)與存續時間預測，以利銀行建立預警機制，可更精準掌握信用卡戶的生命週期，即時採取動態風險管理策略，以降低違約之風險成本及增加銀行收益。

This thesis focuses on the topic of consumer finance credit scoring and mainly explores (1)”Improving Behavior Scores Using the Dual Scoring Model” and (2) “Dynamic Credit Risk Management Adopting Survival Model “. The first part of this thesis is to improve behavior scores using the dual scoring model. The credit scoring model is the most successful statistical model used in finance business. However, while the model has been actively developed for credit grant to applicants and for the credit risk management of already existing loan accounts, most past researches have used the one-dimensional credit scoring model to segment customer risk. In focusing on the credit risk management of credit card accounts, we combine the banks’ internal behavior scoring model with the external credit bureau scoring model for constructing the dual scoring model. The predict power of this dual scoring model are found to exceed those of using the one-dimensional internal credit card behavior scoring model or the external credit bureau scoring model separately, which leads to the inference that it is more accurate in terms of judging and segmenting risk. Moreover, in regard to banks’ asset portfolios, the model provides further management and supervisory and, in terms of designing credit strategies that encompass credit limit management, card renewal and collection actions, which can in practice generate even greater benefits. The second part of this thesis is to predict the default probability dynamically by survival analysis. Most credit scoring models adopt linear discriminant analysis, neural networks, classification and regression trees, and logistic regression to construct credit scoring models. Those approaches are to predict default probability in a fixed period, not at individual time point. This study is to adopt survival analysis with duration concept. Survival analysis is common used in medical, biology, and industry engineering. We use Cox’s PHM(Proportional Hazard Model) to find the significant factors for demographic and behavior part and predict hazard rate and duration. Then we could take immediate risk strategy and action accurately.