【作者】 成洪静；

【导师】 陈立潮；

【作者基本信息】 太原科技大学 ， 计算机软件与理论， 2008， 硕士

【摘要】 银行风险管理一直是国际国内金融界关注的焦点,入世后,国内银行面临着国内外银行的激烈竞争;国内银行由于信贷决策手段跟不上企业经营与市场信息的快速变化,导致不良资产率高、信贷资产安全性差、不良贷款比重大、贷款损失严重,隐含着相当大的危机。如何建立有效的企业信用评估体系和银行信贷风险评估模型,为银行决策提供科学的量化决策依据,全面降低不良贷款率,提高信贷资产质量,是国内银行共同面临的研究课题。因此,研究银行信贷风险评估具有重大现实意义。目前,国内银行大多采用定性与定量相结合的风险量化模型对信贷风险进行评估,由于银行信贷风险等级的划分是一个多分类问题,这种建模技术直接影响信贷风险预测精确度和模型评估能力。本文在深入研究了国内外银行信贷风险评估采用的建模技术基础上,针对BP神经网络在建立信贷风险评估模型时存在的推广能力差和预测时间长等缺点,同时考虑到支持向量机可以处理一些多分类问题,采用支持向量机方法建立银行信贷风险评估模型。然而,支持向量机在处理多分类问题时多采用基于二值分类器的多分类器构造算法,这种算法在提取样本子集时仅考虑到二值分类而忽视与其他类之间联系,造成了类间信息丢失降低了分类精度,针对这一问题,本文提出了基于支持向量回归机的多分类器构造算法。该算法从决策函数入手把二值分类器扩展成三值分类器,增强了类间联系,解决了信息丢失问题;通过引入支持向量回归机代替三值分类器,解决了三值分类器存在的由于参数调整过多造成的计算代价大的问题。并采用中国银行滨河支行提供的数据,通过多组数据实验,将实际情况与该算法结果进行了对比验证。实验结果表明,该算法可以有效地提高预测精度,同时能解决计算代价大的问题,减少了占用的内存空间,提高了运行效率,缩短了训练时间,为银行信贷风险快速、有效的评估提供了更可靠的依据。更多还原

【Abstract】 The bank risk management has been the focus which the international and domestic financial circles have paid attention to all the time. After joining the WTO, the domestic banks are faced with the severe competition from the domestic and international banks. Since the means of credit decision support in the domestic banks can’t keep up with the rapid changes in the business operation and market information, the proportion of the non-performing assets and loan is quite high, and the assets security of credit is poor, as well as the loan losses are serious. So a tremendous crisis is hided in the above-noted problems. A common research subject by domestic financial circles faced is how to establish an efficiency enterprise credit evaluation system and a good bank credit risk evaluation model, so as to provide the scientific quantized reference for making decisions in the banks, reduce non-performing loan rate in an all-round way and improve credit assets quality. Hence, it is great realistic significance to study the bank credit risk evaluation.The risk quantization model combining qualitative analysis with quantitative analysis is adopted by most banks to evaluate the credit risk currently. The risk rank division of the bank credit is a problem of multi-classed. The accuracy of credit predicted and the ability of model evaluating are affected directly by the modeling techniques. This paper makes deeply research on the modeling techniques of bank credit risk evaluation adopted by international and domestic. Aiming at the poor Generalization Ability and the long prediction time in using Back-Propagation NN to establish the model of bank credit risk evaluation, in the meantime, considering that Support Vector Machine can solve some problems of multi-classed; the paper adopts SVM to establish this model. However, the algorithm about structuring multi-classifier based on bi-classifier is adopted widely in using SVM to solve these problems. But the relation between two classes is considered only by the above-noted algorithm rather than the relation among other classes, so losing information among classes is inevitable. The information loss results in the decreases of accuracy classed necessarily. To improve the performance, the new algorithm about structuring multi-classifier based on Support Vector Regression is put forward in this paper. Bi-classifier is expanded to tri-classifier bringing about strengthen relation among classes and reduce information loss in the view of decision function, then tri-classifier is replaced by SVR so as to solve the problem of high computational cost caused by adjusting overabundant parameters. Finally, the facts and proposed algorithm results are compared by much test data provided by Bank of China binhe sub-branch. The experimental results show that the algorithm can improve efficiently the forecast accuracy, solve the problem of high computational cost, decrease the memory occupied, and increase efficiency. Therefore, it can offer more dependable reference for efficient and effective evaluation in bank credit risk.更多还原