植基於基因演算法與支援向量機之銀行不良債權影響因子研究

Research on Factors Influencing Bad Loans Using Genetic Algorithms and Support Vector Machine

翁民義(Ming-Yi Weng)

指導教授 ： 陳同孝

共同指導教授 ： 陳民枝(Jeanne Chen)

國立臺中科技大學/資訊學群/資訊科技與應用研究所/碩士(2010年)

摘要

現今微利及經濟低迷時代，銀行間彼此競爭更加激烈，而利差也有逐漸縮小的趨勢，因此如何降低銀行不良放款戶及增加呆帳回收，已是銀行業首要研究課題，且金管會已明定各銀行在2013年時，會計財務報表需符合IFRS財報透明揭露及金融資產價值精準評估之原則，因此極需金融資產減損參考的科學評估模組。基此本論文利用銀行所擁有之不良放款及呆帳客戶資料作篩選分析後，發現研究資料集的類別樣本分佈呈現偏斜狀況，故使用增加正例類別的樣本資料方法以提高預測準確率，再結合基因演算法及支援向量機作為研究方法，建構特徵值選取模組，從中選取出影響客戶不良放款及呆帳回收發生的重要風險因子，再依所選取出關鍵因素欄位資料組合成樣本集，透過SVM演算法來建構不良放款預警及呆帳回收預測模組。經實驗結果證實，本研究不良放款預警模組平均準確率99.91%，呆帳回收預測模組平均準確率95.36%，顯示所建構之模組具有穩定性及高度精準性。透過本研究不良放款預警模組，可提供給銀行業務人員判定採取因應措施防範，以降低不良債權之發生；而呆帳回收預測模組，則可提供給銀行催理人員進行積極追討程序，以增加非利差收入。最後在IFRS運用上，模組所統計出之逾期金額，可提供給銀行作為金融資產減損參考預估值，來因應配合金管會要求2013年與全球金融體系接軌之目的。

關鍵字

資料探勘、基因演算法、支援向量機、不良放款、呆帳回收、IFRS

並列摘要

Competition among banks has become more intense causing. Interest rates are to deflating. Profits have become marginal low. Therefore, how to reduce the amount of non-performing loan (NPL) accounts and to increase the recovery rates of bad debts are the primary issues of banks. Furthermore, the Financial Supervisory Commission have announced that after year 2013, credit card companies and insurance intermediaries will be required to prepare IFRS comprehensive financial statements. Consequence, there is desperate need for a model to evaluate financial lose.
This paper filtered and analyzed a number of customer data of NPL accounts and bad debts of a bank. Results showed that the category sample distribution of the research data was partially skewed. Hence, the sample data of positive category were added to increase the accurate rates of prediction. The Genetic Algorithm and Support Vector Machine were used in the methodology to construct an Eigen value selection model. The significant risk factors, which might influence the NPL accounts and bad debt recovery, were also chosen. The data were queried according to these factors and combined as the sample data. To construct the module for the precaution of NPL accounts and prediction of bad debt recovery, the SVM algorithm was used. Experimental results showed that the average accurate rate of the precaution module was 99.91% and the average rate of prediction module was 95.36%; these results indicated the stability and effectiveness of the experimented models. The precaution module may bank personnel’s to adopt strategies to eliminate NPL accounts. The prediction module may provide debt-collection managers with reference data to actively pursue the recovery of those bad debts that can increase of non-interest revenues. Finally, the overdue payment which these two modules statistically compiled could be used on IFRS and used by banks to estimate the value of financial assets in order to meet the goal set by the Financial Supervisory Committee in order to integrate with the global financial system.

並列關鍵字

Data Mining, Genetic Algorithm, Bad Loan, Debt Recovery, IFRS

參考文獻

[1] B. Baesens, T. V. Gestel, S. Viaene, M. Stepanova, J. Suykens, and J. Vanthienen, “Benchmarking State-of-the-art Classification Algorithms for Credit Scoring”, Operational Research Society, pp.627-635, 2003.

[3]S. Floyd, and M. Warmouth, “Sample Compression, Learn Ability, and the Vapnik-chervonenkis Dimension Machine Learning”, Kluwer Academic Publishers, Boston, Manufactured in the Netherland, pp. 269-304, 1995.

[4]D. E. Goldberg, “Genetic Algorithms in Search, Optimization, and Machine Learning”, Addison Wesley, pp. 146-168, 1989.

[5] F. H. Grupe, and M. M. Owrang, “Database Mining Discovering New Knowledge and Cooperative Advantage”, Information System Management, pp.26-30, 1995.

[6] J. J. Grefenstette, and J. E. Baker, “How Genetic Algorithms Work: A Critical Look at Implicit Parallelism”, George Mason University, pp.20-27, 1989.