本研究嘗試提出藉由應用貝氏統計的潛在變項模型 (Latent variable model)與分類迴歸樹(Classification and Regression Tree)建構出一套整合模式以解決目前金融機構面對授信業務時需面對的三項難題:第一,銀行無法明確地得知目前的授信決策對未來的風險影響程度為何;當金融機構欲針對申請者的歷史交易和還款情況授予適當的產品(如:信用額度或循環利率)時,常苦無適當的資訊說明授信過程與結果;最後,在授信作業受限於人工審核機制下,易使作業成本增加且效率不彰。因此,為驗證本研究所提整合模式的可行性,針對台灣某大銀行的顧客歷史交易與還款紀錄,以及人口統計變數、歷史信用行為及銀行的授信程度(信用額度及循環利率)資訊於模式中。經由研究證實,藉由持卡人的交易行為中萃取的顧客價值對評分模式而言,其重要程度不僅遠高於人口統計變數,且具有顯著的價值性。此外,整合模式在辨識顧客的能力上除平均鑑別正確率可達91.425%外,亦具有誤判成本較低的優勢。
A Bayesian latent variable model with classification and regression tree approach is built to overcome three challenges encountered by a bank in credit-granting process. These three challenges include (1) the bank wants to predict the future performance of an applicant accurately; (2) given current information about cardholders’ credit usage and repayment behavior, financial institutions would like to determine the optimal credit limit and APR for an applicant; and (3) the bank would like to improve its efficiency by automating the process of credit-granting decisions. The data set consists of each credit card holder’s credit usage and repayment data, demographic information, and credit report. Empirical study shows that the demographic variables used in most credit scoring models have little explanatory ability with regard to a cardholder’s credit usage and repayment behavior. A cardholder’s credit history provides the most important information in credit scoring. Compared to the performance of discriminate analysis and logistic regression, the proposed model has a 91.425% average accurate rate in predicting customer types, and has the lowest misclassification cost.