基于分类器组合的企业信用评级研究
控制理论在经济领域中的应用控制理论在经济领域中的应用
摘 要 对客户的信用评级是银行贷款业务经营的核心，信用评级是否科学可靠、健全，直接关系到银行贷款的成败。因此，借助信用评级模型对企业的信用等级做出准确、可靠的评判，具有重要的现实意义和使用价值。
国内银行对客户的信用等级的评定，还处在对企业的某些单一指标进行评价，然后加权平均的阶段。因此，迫切需要建立更为准确的定量模型来解决信用评估问题。国内外学者已利用神经网络技术来建立企业信用评级模型。然而，基于单个神经网络建立的评级模型不稳定，即学习样本较小的变化会引起神经网络结构和预测性能的较大改变，无法保证模型的可靠性；而且，单个神经网络评级模型的准确率，仅仅依靠参数和算法的调整也难有较大提高。针对这些问题，本文使用了分类器组合技术，对企业信用等级评估建模方法进行了更为深入的研究。
本文首先阐述了分类器组合的原理和方法，介绍了分量分类器输出结果的组合规则，以及产生分量分类器的两种最为流行的算法——Bagging和Boosting。在前人研究的基础上，本文考察了BP神经网络对信用评级问题的适用性，并选用BP神经网络作为多分类器系统的元学习算法。
本文的主要工作是：针对选定的元学习算法对数据进行相应处理，使用Wrapper方法对评级指标进行选择；通过实验，研究了Bagging和AdaBoost算法在不同算法参数和组合规则下所产生的多分类器模型的性能差异，以此为基础确定了适合建立评级模型的组合方法。实验结果表明，使用Bagging和平均法得到的评级模型性能良好，Bagging方法更适合于建立信用评级模型。
[英文摘要]： ABSTRACTIt is important for the commercial banks to assess the credit rating of their customers since the assessments can have significant impact on bank lending decisions and profitability. The banks must carry out a scientific, reliable assessment method for the credit rating. So developing accurate credit rating models is very valuable.
At present, the approach to assess the credit rating used in China is to produce an internal rating, which takes into account various quantitative as well as subjective factors through a scoring system. The problem with this approach is the subjective aspect of the assessment, which makes it difficult to make consistent estimates. There is a need, therefore, to develop fairly accurate quantitative assessment models that can serve as very early warning signals for counterparty defaults.
Some researchers have developed the assessment models using neural networks. But the neural networks are instable; perturbing the learning set can cause significant changes in their structure and accuracy. The instability affects the reliability of neural networks model. Further more, it is difficult to improve accuracy of model based on single neural network by adjusting the parameters and learning algorithm. Focusing on these problems, we study on developing assessment models using classifiers combination method.
In this paper, we review principles and methods of classifiers combination, introduce to combining rules of component classifiers and approaches for producing component classifiers——Bagging and Boosting. Inspired by previous researches, we study on the applicability of backpropagation algorithm to credit rating assessment, and then we select backpropagation neural networks as the meta-learning algorithm of multiple classifiers system.
The major work of this paper includes: data processing for the meta-learning algorithm; selecting financial ratio index system using wrapper approach; studying on assessment models based on Bagging and AdaBoost with different parameters by experiments; and confirming the combination method applied to credit rating assessment. The result of experiments shows that the performance of models based on Bagging is superior, and the Bagging technique can well be applied to credit rating assessment.
Key Words: Credit Rating Assessment; Classifiers Combination; Bagging; Boosting; AdaBoost
