基于非平衡数据分类的贷款违约预测研究

发布时间：2018-03-01 22:16

本文关键词： 贷款违约预测非平衡数据随机森林并行计算　出处：《中南大学》2013年硕士论文　论文类型：学位论文

【摘要】：如何在发放贷款前有效的评价和识别借款人潜在违约风险,计算借款人的违约概率,是现代金融机构信用风险管理的基础和重要环节,也是数量经济学、金融学等领域中的研究热点问题。现有的贷款违约数据大部分都是非平衡的,以往的研究并未足够注意这一特征也就重视这一问题的深入研究,本文主要研究如何借助非平衡数据分类的思想对银行等金融机构的历史贷款数据进行分析,预测贷款违约的可能性。对于非平衡数据问题,采用基于数据平衡的方法构造随机森林；针对数据较大的问题,采用具有并行特性的随机森林算法。基于上述研究,本文提出了一种改进的带权重的并行平衡随机森林算法(WPBRF)。WPBRF算法在构造随机森林的每个决策树的同时利用OOB数据估计该决策树的预测性能,并据此赋予每个决策树不同的权重；此外,WPBRF算法利用了随机森林算法的可并行计算的特点,减少了单个决策树的训练时间。实验结果表明,WPBRF在准确率和平衡准确率等方面超过了SVM、KNN、C4.5等常见分类算法和随机森林算法。此外,利用随机森林的并行性的WPBRF算法大幅降低了算法的学习时间,提高了算法的执行效率。
[Abstract]:How to effectively evaluate and identify the borrower's potential default risk and calculate the borrower's default probability before making loans is the foundation and important link of credit risk management of modern financial institutions, and it is also the quantitative economics. Hot issues in finance and other fields. Most of the existing loan default data are unbalanced. Previous studies have not paid enough attention to this feature and paid attention to the in-depth study of the problem. This paper mainly studies how to use the idea of disequilibrium data classification to analyze the historical loan data of banks and other financial institutions to predict the possibility of loan default. A method based on data balance is used to construct a random forest, and a stochastic forest algorithm with parallel characteristics is used to solve the problem of large data. In this paper, an improved parallel balanced stochastic forest algorithm with weight is proposed, which not only constructs each decision tree of random forest, but also uses OOB data to estimate the prediction performance of the decision tree, and gives different weights to each decision tree. In addition, the WPBRF algorithm takes advantage of the parallelism of the stochastic forest algorithm to reduce the training time of a single decision tree. The experimental results show that WPBRFs outperform common classification algorithms such as SVMKNNC4.5 and stochastic forest algorithms in terms of accuracy and balance accuracy. In addition, the WPBRF algorithm using the parallelism of stochastic forests greatly reduces the learning time of the algorithm. The efficiency of the algorithm is improved.
【学位授予单位】：中南大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：F830.59;F224

【参考文献】