【作者】 周志刚；

【导师】 张大斌；

【作者基本信息】 华中师范大学 ， 管理科学与工程， 2015， 硕士

【摘要】 信用风险管理一直是金融风险管理的重要组成部分。伴随着社会信息化的不断发展,信用风险管理问题呈现大规模、复杂化、不确定性的特点,而其管理的方法无法满足大规模跨领域的信用风险度量与评估问题。一方面由于其缓慢的人工评估方法与度量模型无法精准的判断拥有庞大数据业务量对象的信用风险水平,另一方面在应用复杂的信用风险度量模型的时候,都不可避免的依靠不完善的假设及估计。在现阶段的金融领域,传统的金融信用风险模型无法准确地解决这样复杂的风险问题,但是人工智能和数据挖掘技术的发展为这些问题的解决提供了可能。管理的实质是为决策者提供最优决策,信用风险管理实际上也是一个优化问题。差分进化算法在优化问题上具有良好的性能,并且该算法结合了智能演化和群体学习的机制,算法参数少,能够快速寻找到最优区域。为了提高信用风险度量模型的评估能力,本文将评估问题转变为寻优问题,这样有利于差分进化算法的引入,但是在面对复杂多变的信用风险问题时,差分进化算法本身会有这样或那样的不足。针对这样的情况,本文首先对传统差分进化算法提出自己的改进思想,提高其算法的优化性能；其次,利用改进的算法,结合中国上市公司的实情,优化KMV模型,找出符合国情的违约点；最后利用数据挖掘中聚类的思想,改变差分进化算法的形式,来对我国上市公司进行评估。借鉴多种群并行机制和随机搜索策略,提出一种基于随机扩散搜索的协同差分进化算法。该算法引入反向混沌搜索的初始化机制,利用随机扩散搜索策略将种群分为成功和失败两个子群,并对改进的成功和失败子群分别采用不同的差分策略,克服单一差分策略的缺陷,同时,定期将子群的部分最好与最差个体实现一对一的信息交流,从而达到协同进化的目的。通过函数实验仿真,并将它与其它算法进行比较,结果表明该算法的收敛速度和寻优精度得到明显改善,具有较好的收敛速度和寻优能力。为了探寻在KMV信用评级模型中,违约点的短期和长期比率系数以及投资人市场主观态度系数的最优组合,运用差分进化算法,构造了最优违约系数的不确定性DE-KMV测算模型。通过分析近年来中国经济发展环境对上市公司违约风险影响的趋势,厘清人们对当前金融市场环境好坏的基本态势,利用不确定性DE-KVM模型对上市公司的违约情况进行测算,得到了最优的中国上市公司默认违约点系数和投资人态度系数。建立信用风险评价的差分进化自动聚类模型,并将其应用到我国上市公司信用风险评价中。该模型不要求事先知道分类的数据,相反,依靠群体智能去寻找最优的分区。通过数据仿真,并与遗传算法、决策树、BP神经网络模型进行信用风险评价的实证对比研究,结果表明,该模型能够非常准确的找到数据对应的分区,大大提高了信用评估的准确性,降低了风险成本,对信用风险的管理和控制具有很高的利用价值。更多还原

【Abstract】 Credit risk management has always been an important part of financial risk management. With the continuous development of social informatization, the credit risk management problems present large-scale, the characteristics of complexity, uncertainty, and their management methods cannot meet the large-scale interdisciplinary credit risk measurement and assessment of the problem. On the one hand, because of the slow manual assessment methods and measurement model difficult to determine the precise level of credit risk with a large object data traffic, on the other hand in the application of complex credit risk measurement model, are inevitable on faulty assumptions and estimates. At this stage of the financial sector, the traditional financial credit risk models can’t correctly solve the problem of the risk of such complex, but the development of artificial intelligence and data mining techniques for the solution of these problems may be.The essence of management is to provide optimal decision for policymakers, credit risk management is actually an optimization problem. Differential evolution algorithm has good performance in the optimization problem, and the algorithm combines the intelligent evolution and group learning mechanism, less algorithm parameters, able to quickly find the optimal area. In order to improve the assessment of credit risk measurement model, this article will evaluate problem into optimization problem, this is conducive to the introduction of differential evolution algorithm, but in the face of the complex credit risk problem, the differential evolution algorithm itself will have this or that. Against such a situation, this article first to the traditional differential evolution algorithm is put forward their own improvement ideas, to improve the algorithm of optimal performance; Secondly, using the improved algorithm, combined with China’s listed companies of the truth, to optimize the KMV model, find the points of default conforms to national conditions; Finally using the ideas of clustering in data mining, to change, in the form of differential evolution algorithm to evaluation of listed companies in our country.From various species of parallel mechanism and random search strategy, and put forward a kind of cooperative differential evolution algorithm based on stochastic diffusion search. The algorithm is introduced into reverse the initialization of chaotic search mechanism, using stochastic diffusion search strategy population can be divided into two subgroups, success and failure and to improve the success and failure of subgroup, respectively, using different difference strategy, to overcome the defects of the single difference strategy, at the same time, regular parts of the subgroup is best and worst individuals realize one-to-one communication, so as to achieve the goal of cooperative coevolution. Through the function simulation, and compared it with other algorithms, the results show that the algorithm convergence speed and optimization can be obviously improved, has good convergence and optimization ability.In order to explore the KMV credit rating model, the default point of short term and long term ratio coefficient and investor market is the optimal combination of the subjective attitude coefficient, using the differential evolution algorithm, constructs the optimal DE-KMV default coefficient uncertainty calculation model. Through the analysis in recent years, China’s economic development environment for the listed company on the impact of default risk tendency, clarify the basic situation of the current financial market environment is good or bad, with uncertainty DE-KVM model to measure the defaults of listed companies, the optimal point of China’s listed companies default default coefficient and the coefficient of investor attitude.Establish credit risk evaluation of differential evolution automatic clustering model, and apply it to the credit risk evaluation of listed companies in China. The model does not require prior know the classification of data, on the contrary, depending on the swarm intelligence to find the optimal partition. Through data simulation and genetic algorithm, a decision tree, BP neural network model for credit risk evaluation of empirical comparative study, the results show that the model can find very accurate data corresponding partition, greatly improving the credit evaluation of accuracy, reduces the cost of risk, the credit risk management and control of high use value.更多还原