放款向來是銀行主要的獲利來源之一，因此放款若是發生逾期，非但銀行原本預期的利息收入不能如期收取，若是透過處分擔保品或向借保人訴追無效，可能徒增代墊訴訟費用並血本無歸。「放款覆審」猶如人體定期健檢，可以提早發現放款逾期徵兆並加以預防。惟過去對於借款人會發生逾期之分析研究多著重於貸放前之落實徵信與審查防範，鮮少研究以貸放後之管理方式作為亡羊補牢的手段，所以本論文針對個案銀行遭遇的放款覆審問題，以該銀行自民國96年起至103年9月底止各月底所列報新發生之逾期授信戶，依本研究擷選逾期放款風險因子，組成資料分析研究。本研究列入的自變數為16項『逾期放款風險因子』，目標變數為『列報逾期距到期天數』，基於決策樹的概念，運用M5'模式樹之分析結果發現關鍵風險因子分別組成企業逾期放款與個人逾期放款分別有7條與8條規則式可供徵審部門作為修訂授信覆審之範圍頻率參考。不論是企業或個人借款列報逾期距離該筆放款到期天數都與放款科目的關聯性最大，所以個案銀行若加入「放款科目」作為覆審頻率之衡量標準，將可避免不必要的覆審人力成本浪費並同時提升覆審效果。

Abstract

SUMMARY
This study focuses on the loan review problems encountered by the case bank. Drawing upon this bank’s new overdue loan accounts reported at the end of each month from Jan 2007 to Sep 2014, this study attempts to capture the risk factors for non-performing loans (NPL). The independent variables are 16 “risk factors for NPL”, and the target variable is “the difference of the date of reporting NPL from the due date”. Based on the concept of decision tree, this study uses the M5' model trees to identify key risk factors and establish 7 and 8 rules for corporate NPLs and individual NPLs, respectively. The results can be a reference for the loan review department in adjusting the scope and frequency of credit reviews. From both corporate and individual loan accounts, it is found that “the difference of the date of reporting NPL from the due date” is most correlated with “loan type”. In other words, by adding “loan type” into the criteria for determining the review frequency, the case bank can save unnecessary labor cost and also improve the effectiveness of their review operation.

Keywords: Non-performing loan (NPL), loan review, decision tree, M5' model trees

INTRODUCTION
This study focuses on NPLs, a common problem encountered by banks. Factors of NPLs are numerous. Previous research of causes of NPLs has paid more attention to credit review before lending, and little research has explored the remedial measures for managing NPLs. Motivated to fill this gap, this study will investigate if the credit review scope and frequency required by the case bank are effective for detecting early signs of NPLs and further reducing occurrence of NPLs or they are so over-stressed that they have caused a burden to credit reviewers? By capturing the associations among the key risk factors for NPL from a dataset of overdue loan accounts, this study attempts to provide the credit review department of the case bank an effective approach to reduce NPLs.
METHOD
In order to analyze the case bank’s NPLs, this study employs the concept of decision tree to capture the key risk factors for corporate NPL and individual NPL and then develop a forecast model. This model will be compared with the current review rules, and overfitting will be avoided. The M5' model trees, an improved extension of CART decision trees, will be adopted for analysis. The M5' model trees involves five procedures, including tree-growing, tree-pruning, tree-smoothing, dealing with enumerated attributes, and dealing with missing values.
EXPERIMENTAL RESULTS
The following experiments are using regression trees, using the K-fold to show the experimental results, and the experiment was repeated for 30 times, Table 4-3 and Table 4-4 in parentheses represents the Standard Deviation after 30 times different experimental results . To calculate the difference between the valuation and the actual value of the estimated value and the actual value and the description of the dispersion of the sample, using the mean absolute error (MAE) and mean square error (RSME) compare two indicators to assess the quality of the model Collocation.
Based on the dataset of corporate NPLs, the following experiment results and model tree can be created:
Table 1. The experiment results of corporate NPLs
2-fold 3-fold 5-fold 10-fold
MAE 217.60(54.81) 192.11(48.36) 181.84(59.39) 174.33(78.57)
RMSE 376.18(122.81) 334.16(123.25) 307.06(137.33) 277.97(164.83)
Figure 1. The model tree for corporate NPLs
Based on the dataset of individual NPLs, the following experiment results and model tree can be created:
Table 2. The experiment results of individual NPLs
2-fold 3-fold 5-fold 10-fold
MAE 975.07(111.62) 895.64(97.41) 840.30(137.73) 801.25(188.13)
RMSE 1251.50(136.31) 1151.17(122.53) 1080.40(183.25) 1019.53(254.24)
Figure 2. The model tree for individual NPLs
CONCLUSION
In this study, the independent variables are 16 “risk factors for NPL”, and the target variable is “the difference of the date of reporting NPL from the due date”. The analysis with the M5' model trees reveals 6 key risk factors for corporate NPLs, including “loan type”, “amount secured by other accounts in the bank-1”, “collateral”, “total amount of loans from the bank”, and “the difference of the date of reporting NPL from the lending day” and 4 key risk factors for individual NPLs, including “loan type”, “the difference of the date of reporting NPL from the lending day”, “total amount of loans from the bank”, and “amount of cash lending”. Based on the above risk factors, this study develops 7 and 8 rules for corporate NPLs and individual NPLs, respectively, which can be an important reference for the bank’s loan review department.
From these factors and rules, it can be observed that particular attention should be paid to three types of corporate loans, including a52 (medium-term unsecured loan), a53 (medium-term secured loan), and a55 (long-term secured loan). As for individual loans, particular attention is required for a55 (long-term secured loan) with a loan duration shorter than 2,799 days (approximately 7.6 years).
The statistical classification method and consideration of more non-financial variables in the review of corporate NPLs can be improved in subsequent research.

中文部份：
1. 中華民國銀行商業同業公會全國聯合會網址：http://www.ba.org.tw。
2. 汪忠平(1997)，「以類神經網路建立財務危機預測模式--考慮產業，總體經濟因素與其穩定性因素」，東吳大學企業管理學系碩士論文。
3. 李婉慈(2000)，「銀行往來關係與公司經營績效及財務危機之關聯性分析」，國立中正大學財務金融研究所碩士論文。
4. 沈丈智(2007)，「中小企業財務預警模型與銀行授信決策之研究」屏東科技大學高階經營管理碩士在職專班碩士論文。
5. 吳明德（2011），「金融機構逾期放款及其影響因素之研究」，國立成功大學高階管理碩士在職專班碩士論文。
6. 施並洲(1999)，「類神經網路、案例推理法、灰色關聯分析於財務危機之應用」國立中央大學工業管理研究所碩士論文。
7. 高琦媛(2005) ，「信用風險資產對銀行獲利率的影響」，東吳大學經濟學系碩士論文。
8. 馬君梅、謝永明(2007)，「事務所內自發性會計師更換對公司盈餘品質之影響」，東吳經濟商學學報，頁79-105。
9. 陳肇榮（1983），「運用財務比率預測企業財務危機之實證研究」，國立政治大學財政研究所博士論文。
10.陳建州（2007），「經理人授信政策對「銀行個人擔保授信」的績效與逾期放款之影響-以SP銀行為例」，國立中正大學企業管理研究所碩士論文。
11.陳自連（2013），「探討商業銀行個人逾期放款之決定因素」，國立中正大學高階主管管理碩士在職專班碩士論文。
12.陳錦富(2014)，「中小企業財務預警模型」，國立高雄第一科技大學財務管理研究所碩士論文。
13.許正春（1990），「銀行授信考量因素之研究」，國立交通大學管理科學研究所碩士論文。
14.郭瓊宜(1994)，「類神經網路在財務危機預警模式之應用」，淡江大學管理科學研究所碩士論文。
15.郭立程（2012），「如何分析財報資訊以提早發現逾期放款之關鍵因素-以個案YHI公司為例」，國立中山大學高階經營碩士班碩士論文。
16.張美玲(2000)，「由交叉持股觀點探討財務危機問題－台灣上市公司之實證研究」，淡江大學管理科學學系碩士論文。
17.黃俊雄(1994)，「企業財務危機預警模型在銀行授信決策之運用」，國立政治大學企業管理研究所碩士論文。
18.黃喜煌(2009)，「以銀行授信角度切入企業財務預警之探討」，國立高雄第一科技大學金融所碩士論文。
19. 曾憲雄、蔡秀滿、蘇東興、曾秋蓉、王慶堯(2011)，「資料探勘 Data Mining」，台北市：旗標有限公司。
20.賴世權(1989)，「企業正常或具失敗危機之關鍵性財務指標比較分析研究」，國立中興大學企業管理研究所碩士論文。
21.劉又華（2007），「借款人償債能力及穩定性因素對小額信用貸款逾期放款之影響」，靜宜大學管理碩士在職專班碩士論文。
22.劉祥熹、黃日鉦（2005），「銀行業借貸風險評估之實證分析－－資料探勘與統計模型之比較」，華人前瞻研究 200505 (1:1期) ，中華民國南洋文化學會，頁45-71。
23.謝依錡﹙2011﹚，「以公司治理因素建構企業財務預警模型-運用模糊判別分析法」，清雲科技大學財務金融所碩士論文。
24.盧瑞山（2006），「漫談資料探勘技術(Data mining)於環境資料庫之應用」，德明技術學院資訊管理學系未出版碩士論文。
25.聶聖之(2004)，「會計資訊與財務危機預測：考量財務危機嚴重程度差異」，中原大學會計研究所碩士論文。
26. SAS資源中心網站 (網址http://www.sasresource.com/artical72.html)

英文部份：
Breiman, L., Friedman, J. H., Olshen, R. A., & Stone, C. J. (1984). Classification and Regression Trees. New Yourk: Chapman & Hall.
Carcello, J. V. and A. L. Nagy. (2004). Audit firm tenure and fraudulent financial reporting Auditing, A Journal of Practice and Theory 23 (2): 55-69.(January), Vol 26, Issue 1-3, 1-34.
DeAngelo, L. E. (1981). Audit size and audit quality, Journal of Accounting and Economics, 3(3), 183-199.
Edward I. Altman,(1968) Financial Ratios, Discriminant Analysis and the Prediction of Corporate Bankruptcy , Jorunal of Finance, Sep, P589~P609.
Gür Ali, Ö., Serpil, S., Tom, v. W., & Jan, F. (2009). SKU Demand Forecasting In The Presence Of Promotions. Expert Systems with Applications , 36(10), 12340-12348.
Katharina, D. C. Stark & Dirk, U. Pfeiffer (1999). The application of non-parametric techniques to solve classification problems in complex data sets in veterinary epidemiology- an example, Intelligent Data Analyusis, 3(1), pp.23-35
Kohavi, Ron (1995). A study of cross-validation and bootstrap for accuracy estimation and model selection. Proceedings of the Fourteenth International Joint Conference on Artificial Intelligence 2 (12): 1137-1143.
Quinlan, J. R. (1992). Learning with Continuous Classes. Paper presented at the 5th Australian Joint Conference on Artificial Intelligence, Hobart, Australia.
Quinlan, J. R. (1993). C4.5: Programs for Machine Learning. Morgan Kaufmann Publishers.
Roiger, R. J., & Geatz, M. W. (2003). Data Mining: A Tutorial-Based Primer. New York: Addison Wesley.
Schwer, R. Keith, & Yucelt, Ugur,(1984）. A Study of Risk-Taking Propensities Among small Business entrepreneurs and Managers, American Journal of small business, Vol.8, Iss.3,pp.31-40.
Wang, Y., & Witten, I. (1997). Inducing Model Trees for Continuous Classes. Paper presented at the Proceedings of the Poster Papers of the European Conference on Machine Learning, Prague, Czech Republic.
William H.Beaver,(1966). How to Separate Financial group.Journal of Finance, pp510-523.
Witten, I. H., & Frank, E. (2005). Data Mining: Practical Machine Learning Tools and Techniques (2nd ed.). San Francisco: Morgan Kaufmann.