信用卡違約態樣及循環信用者特性研究-以某銀行為例

The Study of The Patterns of Credit Card Defaulters And Revolving Credit–A Case Study of Taiwan’s Bank

張家源(Chia-Yuan Chang)

指導教授 ： 梁德昭

淡江大學/商管學院/資訊管理學系碩士在職專班/碩士(2020年)

https://doi.org/10.6846/TKU.2020.00587

摘要

近年來因銀行間彼此競爭激烈，信用卡業務獲利正逐漸減少，為避免虧損，及早地辨識信用風險將更顯得重要。
本研究採國內某銀行信用卡客戶作為分析對象，第一階段經觀察違約者之樣態後發現，多數違約者違約前會先有使用循環信用的歷程，爰第二階段以決策樹、隨機森林、支持向量機、羅吉斯迴歸、類神經網路及深度學習等演算法分析循環信用者之特性。
研究結果發現，多數違約者違約前會有遲延繳款情形，且違約時使用額度大於核給額度8成的現象；循環信用使用者之特性為帳單金額/額度之比率、差別利率級數、帳單金額、近3、6期循環信用次數，以及使用循環信用後將會有持續使用的情形；就模型表現而言，有近6個月繳款紀錄之各分類模型，模型AUC值均大於0.9%，均有好的鑑別能力；深度學習在準確度、靈敏度、特異度、精度等四個指標中囊括兩項第一，而決策樹有運算速度快，容易理解等優點，兩者於建置相關模型時可共同列為優先適用之演算法。

關鍵字

信用卡 ； 循環信用 ； 信用風險 ； Rapidminer

並列摘要

In recent years, the credit card business has been losing profits due to fierce competition among banks. In order to avoid losses, it is even more important to be able to identify credit risks early.
This study adopts the credit card customers of a domestic bank as the objects of analysis. After observing the patterns of credit card defaulters in the first stage, we have found that most credit card defaulters have a history of using revolving credit before defaulting on their credit contracts. In the second stage, the characteristics of revolving credit users were analyzed using decision trees, random forests, support vector machines, logistic regression, artificial neural networks and deep learning algorithms.
The results show that most credit card defaulters will postpone payment before defaulting, and use more than 80% of the authorized amount at the time of default. The characteristics of revolving credit users are shown in the bill amount-to-limit ratio, the differential interest rate, the bill amount, the number of times that revolving credit was used in the past 3 or 6 billing cycles, and the fact that they will continue using revolving credit on an ongoing basis. In terms of model performance, all models with payment records of the past 6 months have AUC values greater than 0.9%, which suggests good discrimination ability. The deep learning model achieved the best performance in two out of four metrics, including accuracy, sensitivity, specificity and precision, while the decision tree model has the advantage of being fast and easy to understand. Therefore, both are preferred algorithms to use when building relevant models.

並列關鍵字

Credit Card ； Revolving ； Credit Risk ； Rapidminer

參考文獻

[1]Breiman, L. (2001). Random Forests Machine Learning, vol. 45.

[2]Cortes, C., & Vapnik, V. (1995). Support-vector networks. Machine Learning, 20(3), 273-297. doi:10.1007/BF00994018

[3]Green, D. M., & Swets, J. A. (1966). Signal detection theory and psychophysics (Vol. 1): Wiley New York.

[4]Hosmer, D. W., & Lemeshow, S. (2000). Applied Logistic Regression. New York: Wiley.

[5]JunLiang. 特徵篩選的原理與實現（上）. Retrieved Apr 12, 2020, from https://www.twblogs.net/a/5d3fa1c6bd9eee517422b54e