商业银行信用风险的KPCA-PSO-SVM智能预警研究

发布时间：2018-05-01 06:02

本文选题：信用风险 + 商业银行　；参考：《成都理工大学》2017年硕士论文

【摘要】：信用风险作为商业银行面临的主要风险之一,一旦发生,不仅会造成商业银行经营损失,甚至引发商业银行破产危机。因此,如何对商业银行信用风险进行预警分析,进而采取行之有效的手段提前防范与控制信用风险,成为了当前理论与实务界探讨的热点问题之一。就中国商业银行而言,起步较晚,发展时间较短,在信用风险管理方面还缺少经验。同时,随着中国资本市场的逐步开放,国外资本不断涌入中国,在加快中国资本市场发展的同时,也很可能对中国脆弱的商业银行信用风险体系形成潜在威胁。因此,优化信用风险预警方法,提升信用风险管理水平,完善信用风险管理体系,对中国商业银行发展具有重要意义。基于上述分析,本论文以中国商业银行的贷款企业,即沪深两市的部分上市公司为研究对象,基于中国金融的现实环境,选择出16项诱发商业银行爆发信用风险的指标变量并进行预处理,从而获得14项能够显著区分中国商业银行信用风险与非信用风险样本的指标变量,进而运用核主成分分析(KPCA)方法对这14项指标变量进行提取,以消除指标变量间的高相关性特征;引入支持向量机(SVM)人工智能技术,构建商业银行信用风险的SVM智能预警模型,并运用粒子群优化(PSO)方法优化SVM模型的参数,以此来开展信用风险预警的研究工作,并通过实验证明了本论文提出的KPCA-PSO-SVM模型在商业银行信用风险预警中的优异的预测性能。本论文的主要研究内容如下:1.对风险预预警样本的预处理研究。由于直接基于样本的原始指标变量来构建预警模型存在诸多问题,因此,本论文运用了归一化处理方法和统计分析方法对原始的指标变量进行了筛选。通过实证结果表明,运用归一化方法能够将各指标变量转换为正态分布,从而能够消除指标变量的量纲问题;运用统计分析方法发现,营业收入增长率和税后利润增长率两项指标变量无法显著区分信用风险与非信用风险样本,因而需要将其从指标变量中删除。通过实验,本论文就获得了具有无量纲特征、能够显著区分信用风险与非信用风险样本的指标变量。2.对指标变量提取方法进行研究。诱发商业银行爆发信用风险的指标变量众多,且这些变量之间往往呈现高相关性特征。如果不消除这些变量的高相关特征而直接运用其进行建模,则很容易引发数据冗余问题,最终降低SVM智能预警模型的预测效果。因此,本论文引入了常用的主成分分析方法(PCA)以及其改进方法——核主成分分析方法(KPCA)进行了实证对比研究。实证结果表明,KPCA方法在指标提取上较PCA方法更为高效,同时,与SVM相结合,KPCA能够显著地提升SVM的预测效果,然而PCA方法却会降低SVM的预测效果。从而表明,银行信用风险预警指标变量存在非线性特征,而KPCA方法正好能够提取指标变量的非线性特征,能够有效地提升SVM的预警能力。3.对SVM参数优化方法研究。SVM智能预警模型的预测能力在很大程度上取决于惩罚参数和核函数参数,如果不恰当地选择这两类参数,就很可能导致SVM模型出现过拟合或欠拟合。为此,本论文对比研究了以网格寻优法(GS)为代表的传统参数优化方法和以遗传算法(GA)、粒子群算法(PSO)为代表的启发式算法在SVM参数寻优中的效果。实验结果表明,启发式算法在参数寻优上优于传统的GS参数寻优方法,其中,以PSO为代表的启发式算法又比以GA为代表的启发式算法具有更为优异的预测性能,能够更为有效地提升SVM预警模型的预测性能。通过上述一系列实验,本论文认为,基于KPCA-PSO-SVM的商业银行信用风险预警模型是商业银行信用风险监管部门应对与防范信用风险的最优的应用工具。监管部门能够运用本论文构建的KPCA-PSO-SVM智能预警模型,对未来一段时间内商业银行的信用风险进行全面而准确的预测,即时制定并实施应对信用风险的相关政策措施,从而加强市场监管,有效地防范信用风险。
[Abstract]:As one of the main risks faced by commercial banks, credit risk will not only cause business losses, but also lead to the bankruptcy crisis of commercial banks. Therefore, how to carry out early warning and Analysis on commercial banks' credit risk and then take effective measures to prevent and control credit risks has become the current theory and As far as China's commercial banks are concerned, China's commercial banks have a late start, a short development time and lack of experience in the management of credit risk. At the same time, with the gradual opening up of China's capital market, foreign capital is constantly pouring into China, while the development of China's capital market is accelerated, and it is likely to be vulnerable to China's business. The bank credit risk system has a potential threat. Therefore, it is of great significance to optimize the early warning method of credit risk, improve the management level of credit risk and improve the credit risk management system. Based on the above analysis, this paper takes the loan enterprises of Chinese commercial banks, that is, some listed companies in Shanghai and Shenzhen two cities as the research. Based on the real environment of China's finance, we selected 16 indexes to induce the credit risk of commercial banks and pretreated them, so as to obtain 14 index variables that can distinguish between credit risk and non credit risk samples of Chinese commercial banks, and then use the nuclear principal component analysis (KPCA) method to enter the 14 index variables. In order to remove the high correlation characteristic between the index variables, we introduce the support vector machine (SVM) artificial intelligence technology to construct the SVM intelligent early warning model of the credit risk of commercial banks, and optimize the parameters of the SVM model by using the particle swarm optimization (PSO) method to carry out the research work of the credit risk early warning, and prove the thesis through the experiment. The outstanding performance of the KPCA-PSO-SVM model in the early warning of credit risk in commercial banks is proposed. The main contents of this paper are as follows: 1. research on Prewarning samples for risk prewarning. There are many questions in the construction of early warning model based on the original index variables based on the sample. Therefore, this paper uses the normalized processing party. The original index variables are screened by method and statistical analysis method. The empirical results show that the normalization method can convert the index variables into normal distribution and can eliminate the dimensionless problem of the index variables. The statistical analysis method shows that the growth rate of revenue and the rate of profit growth after tax are two variables. The method clearly distinguishes the sample of credit risk and non credit risk, so it needs to be deleted from the index variable. Through the experiment, this paper obtains the dimensionless characteristics, which can distinguish between the index variable.2. of the credit risk and the non credit risk sample, and induces the credit risk of the commercial bank to break out. There are many index variables, and these variables often show high correlation characteristics. If they do not eliminate the high correlation characteristics of these variables and directly use them for modeling, it is easy to cause data redundancy and ultimately reduce the prediction effect of SVM intelligent early warning model. Therefore, this paper introduces the common principal component analysis method (PCA). The positive results show that the KPCA method is more efficient than the PCA method in the index extraction, and the KPCA can significantly improve the prediction effect of SVM, while PCA method can reduce the predictive effect of SVM, which indicates that the credit risk of the bank can be reduced by the PCA method. Thus, the bank credit risk is indicated by the PCA method. Therefore, the credit risk of the bank is shown to be the risk of the bank's credit risk. The early warning index variable has nonlinear characteristics, and the KPCA method just can extract the nonlinear characteristics of the index variables. It can effectively improve the early warning capability of the SVM.3.. The prediction ability of the SVM parameter optimization method for the.SVM intelligent early warning model depends largely on the penalty parameters and the kernel function parameters, if the two is not chosen properly. The class parameter is likely to lead to the over fitting or less fitting of the SVM model. Therefore, this paper compares the traditional parameter optimization method represented by the grid optimization (GS) and the effect of the heuristic algorithm represented by genetic algorithm (GA) and particle swarm optimization (PSO) in the optimization of SVM parameter optimization. The experimental results show that the heuristic algorithm is in the parameter. The optimization method is superior to the traditional GS parameter optimization method. Among them, the heuristic algorithm represented by PSO has more excellent predictive performance than the heuristic algorithm represented by GA, and it can improve the prediction performance of the SVM early warning model more effectively. Through a series of experiments above, this paper considers that the commercial bank credit based on KPCA-PSO-SVM is based on this series of experiments. The risk early warning model is the best application tool for the credit risk supervision department of commercial banks to deal with and prevent the credit risk. The supervisory department can use the KPCA-PSO-SVM intelligent early warning model constructed in this paper to make a comprehensive and accurate prediction for the credit risk of commercial banks in the future period, and make and implement the coping credit immediately. Risk related policies and measures to strengthen market supervision and effectively prevent credit risks.

【学位授予单位】：成都理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：F832.33

【参考文献】