應用商業智慧技術於信用卡違約風險之預測

過去發卡銀行核卡作業僅根據授信人員的經驗及傳統的計分卡來審核客戶風險的可能性，缺乏公正及一致性的評估標準，不夠有效率，也不夠E化。銀行業者若能結合商業智慧，發展出一套本身的核卡系統，將更進一步深化銀行競爭的優勢。分類問題是資料採礦作業中最普遍的一種，其目的在於事先預測「尚未發生」的分類事實，信用卡違約風險預測模型正是分類問題的一項應用。本研究將以商業智慧的觀點，配合Microsoft SQL Sever 2005軟體所提供的資料採礦工具，利用發卡銀行龐大的客戶歷史資料，透過先進的資料採礦技術（決策樹、類神經網路）和統計方法（羅吉斯迴歸），建構出一套完全符合自身銀行客戶特性的信用卡評分模型之流程。以本研究的結果所示，在模型的預測能力上，羅吉斯迴歸優於類神經網路，類神經網路又優於決策樹，且根據BASEL對信用評分卡的規定，羅吉斯迴歸為其標準的演算法，因此最終模型即選擇以羅吉斯迴歸所建立的模型，並發現在基本資料上，會影響申請者違約機率的因素為「教育別」、「性別」、「行業別」、「職稱別」、「工作年資」、「緊急聯絡人與正卡關係」；在與銀行往來紀錄上，影響申請者違約機率的因素為「是否申請結餘代償」以及「是否使用扣款帳號」；在金融聯合徵信中心資料上，會影響申請者違約機率的因素為「延遲月數比率」、「循還月數比率」、「近一年M2次數」、「近期是否逾期」、「負債所得比」、「有效卡張數」、「被查詢總家數」。

關鍵字

商業智慧；資料採礦；信用評分；分類模型

並列摘要

In the past, the issuer's verificative card operations only reviewed the risk of customer risk based on the experience of the credit officer and the traditional scoring card. The lack of fair and consistent evaluation criteria was not efficient enough. If the bankers combined with business intelligence, develop a set of their own verificative card system, it will further deepen the advantages of bank competition.The classification problem is the most common of the data mining operations, and its purpose is to predict the fact that ＂haven’t been occurred＂ events, one of the application of the classification problem is that the credit card default risk forecasting model. This study will use the data mining tools provided by the Microsoft SQL Sever 2005 software, leveraging the huge customer data of the issuing bank, through advanced data mining technology (decision tree, neural network) and statistical methods (logistic regression) to construct a set of credit card scoring models that fully comply with the characteristics of their own banking clients. Based on the results of this study, the logistic regression is superior to the neural network in the predictive ability of the model. The neural network is superior to the decision tree. According to BASEL's credit scorecard, logistic regression is the standard algorithm, so the final model is chosen by the logistic regression model established.And found that in the basic information, the factors that affect the probability of the applicant's breach of contract are ＂education＂, ＂sex＂, ＂industry＂, ＂title＂, ＂working year＂, ＂emergency contact and positive card＂.In terms of bank records, the factors that affect the applicant's probability of default is ＂whether to apply for the balance of compensation＂ and ＂whether to use the charge account＂. In the financial information center credit information, the factors that affect the applicant's probability of default is ＂delay monthly ratio ＂,＂ monthly ratio ＂,＂ recent year M2 times ＂,＂ recent overdue ＂,＂ debt ratio ＂,＂ effective card numbers ＂,＂ total number of households ＂.