利用軟性計算法建構金融機構債權資產分類模型

Construction of Classified Model for Financial Institution Loan Potfolio by Soft Computing Method

陳文淵(Wen-Yuan Chen)

指導教授 ： 蔡碩倉博士

亞洲大學/管理學院/經營管理學系碩士在職專班/碩士(2009年)

摘要

金融機構在經營環境日趨艱困之情況下，想要藉由傳統授信業務來獲利的空間已日益壓縮，而龐大的不良債權不僅會大幅降低金融機構的獲利能力，對於整個金融體系的穩定性也會造成相當的影響。本研究藉由彙整金融機構債權資產分類因素，研析其分類評估指標。並利用軟性計算法建構金融機構債權資產分類模型，建構一套更精細可靠的分類模型，提供金融機構不良資產差異化標售或議價之參考。最後測試分類模型指標敏感度，釐清分類之關鍵因素。
本研究以中部地區某公營大型行庫2002年到2007年的授信資產評估表為本研究個案資料來源。依財政部「逾催辦法」規定銀行對資產負債表表內及表外之授信資產評估，除正常還本繳息及授信戶積欠本金或利息在清償期ㄧ個月以內(含)之授信資產列為第一類外，餘不良之授信資產，應按債權擔保情形及逾期時間長短予以評估，分別列為第二類應予注意者，第三類可望收回者，第四類收回困難者，第五類收回無望者。本研究由實際案件中篩選出已逾期放款案件資料360筆及正常放款案件資料90筆，作為金融機構債權資產分類模型的樣本依據。在進行模型之建構與驗證前，為便於後續鑑別金融機構債權資產分類正確率，必須將實證樣本區分為訓練樣本與測試樣本兩組樣本。訓練樣本係用於推估金融機構債權資產分類模型，測試樣本則用於鑑別金融機構債權資產分類之正確預測率。實證結果顯示，訓練樣本之正確預測率於四種網路架構上只有網路Ⅰ及網路Ⅲ為100%，惟就其RMSE數值表現，以網路Ⅲ(以Norm-Cum-Delta-Rule為學習法則，TanH為轉換函數)為最佳，RMSE為0.1476，而網路Ⅳ(以Norm-Cum-Delta-Rule為學習法則，Sigmoid為轉換函數)之RMSE值最差，為0.1955。另在測試樣本方面，模型正確預測率於四種網路架構上只有網路Ⅰ及網路Ⅲ為94.67%，惟就其RMSE數值表現，以網路Ⅰ(以Delta-Rule為學習法則，TanH為轉換函數)為最佳，RMSE為0.2275，而網路Ⅳ(以Norm-Cum-Delta-Rule為學習法則，Sigmoid為轉換函數)之RMSE值最差，為0.1995。
換言之，本研究以Delta-Rule為學習法則，TanH為轉換函數之類神經網路模式數值最佳。因此以該模式進行敏感度分析，發現影響資產分類最顯著的變數為逾期期間，其次依序為擔保品價值、擔保品、從債務人、擔保品種類及擔保品座落，此可提供金融機構人員實務上處理案件參考，以爭取時效、減少審核期間。另可提供金融機構處分不良債權(NPL)標售予資產管理公司(AMC)時，考慮案件之逾期期間、擔保品價值、擔保品、從債務人、擔保品種類及擔保品座落等變數，始得作出最適宜金融機構不良債權及承受擔保品出售之資產組合，作為差異化標售或議價之處理，以助於行銷過程中吸引更多投標者，提高金融機構債權回收率，同時本分類模型亦可強化金融機構貸後管理之債權預警系統，做好事前預防管理工作。

關鍵字

不良債權 ； 資產管理公司 ； 軟性計算 ； 倒傳遞類神經網路

並列摘要

Financial institution wants to make money by traditional loaning service is becoming more difficult in such difficult management environment. The huge non-performing loans are not reduce the earning capacity of the financial institution, but also influence the stability to the whole financial system. This study collect the financial institution loans portfolio classifies factors and analyzes its classified indicator. Utilize the soft computing method to build and construct the financial institution loans portfolio classified model and rebuild new one which is more exactly and more reliable classified model. To provide a reference resource of sale by sealed tender or negotiated price when financial institution treat their non-performing loans. Test the indicator susceptibility of the classified model finally, distinguish the key factor of classification.
We select one large public bank in the middle in Taiwan, use its loaning assets appraisal table from 2002 to 2007 as our case study source. According to the 「non-performing loans measures」 rules made by Ministry of Finance , financial institutions evaluate the balance sheet and out of the balance sheet, except the performing loans and overdue loans within a month ( including a month) are classified type 1, the other non-performing loans should be classified into type 2—pay attention、 type3—maybe will regain、 type 4—difficult to regain、 type5—regain hopeless, by evaluating the creditor's rights guarantee situation and the exceeding the time limit time length. In this study, we sieve 360 non-performing loans cases and 90 normal loans cases from the really cases as the samples of the loans portfolio classified model. Before constructing and proving the building of the model, it must be divide the samples into training sample and testing sample to facilitate follow-up distinguishing loans portfolio classes correctly. Training samples are used for estimating the loans portfolio classes model of financial institution, testing samples are used for distinguishing the loans portfolio classes of financial institution. The real example result shows that the correct prediction rate of training sample only the networkⅠand networkⅢ are 100% on four kinds of network structure. As its RMSE number value, the network Ⅲ (regards Norm-Cum-Delta-Rule as the rule of studying, TanH as a transfer function) is the best one, RMSE is 0.1476, and the networkⅣ (regards Norm-Cum-Delta-Rule as the rule of studying, Sigmoid is the transfer function) its RMSE number value 0.1955 is the worst one. On the test sample side, the correct prediction rate only the networkⅠand networkⅢ are 94.67% on four kinds of network structure. As its RMSE number value, the network Ⅰ (regards Delta-Rule as the rule of studying, TanH as a transfer function) is the best one, RMSE is 0.2275, and the networkⅣ (regards Norm-Cum-Delta-Rule as the rule of studying, Sigmoid is the transfer function) its RMSE number value is the worst on 0.1955.
That is to say, our research regarding Delta-Rule as the rule of studying and TanH as a transfer function get the best value of back-propagation neural network model. So we use the model to analyze the sensitivity, find that noticeable parameter of influence loans portfolio classes is the the exceeding the time limit time length, and the other parameter in sequence is the collateraling value, collaterals, the debtor, the type of collateral, and collaterals located. This result can be a reference to the member of financial institute to deal with practical cases for saving time and reducing exam time. And supply financial institute some references (the case exceeding the time limit period, the collateraling value, collaterals, the debtor, the type of collateral and collaterals located) to deal with non-performing loans (NPL) sale by sealed tender gives to asset management company (AMC), make sure they can combine the optimal asset portfolio composition of NPL and undertaking the collateral, to be the treatment of sale by sealed tender or negotiated price. This can help to attract more biders on the sale process, and enhance creditor's rights returns-ratio. At same time, the classified model also can strengthen the prediction system of loan assets management, to do preventive management work well.

並列關鍵字

non-performing loans ； asset management corporation ； soft computing ； back-propagation neural network

參考文獻

鄭銌鋪，2004年，「壽險業健康醫療險理賠預警系統之研究」，臺中健康暨管理學院經營管理研究所碩士論文。

劉鼎康，2005年，「使用類神經網路進行垃圾郵件過濾之研究」，中原大學資訊管理學系碩士學位論文。

林信成，「軟性計算理論與知識處理技術之研究」，教育資料與圖書館學，第三十八卷，第二期，頁148-173，2000年12月。

Blinks,D.L.and N.M.Allinson.,1991,＂Financial Data Recognition Prediction Using Neural Networks.＂

Coats, P. K. & L. F. Fant, 1993 , “Recognizing financial distress patterns using a neural network tool＂, Financial ManagementVol.12, p142-155.