BACT模型在P2P借款人信用风险评估中的应用

发布时间：2018-05-10 10:00

本文选题：P2P网贷 + 信用风险　；参考：《广州大学》2017年硕士论文

【摘要】：P2P网络借贷,俗称P2P网贷,起步于英国.近年来,伴随着互联网技术的快速普及,P2P网贷在国内外发展特别迅速,呈现出各种各样的形式.P2P网贷属于小额借贷,投资门槛低,备受大众青睐,然而平台跑路、借款人违约等事件层出不穷,因而也得到广泛关注.本文以中国P2P网贷为研究对象,运用两种常用的机器学习方法和一种新的贝叶斯方法来评估网贷借款人的信用风险.这三种方法分别是:支持向量机(SVM)、随机森林(RF)以及贝叶斯可加分类树(BACT).首先将原始数据分成训练数据和测试数据,然后将训练数据等分成三份,称为原始训练集.考虑到数据存在类不平衡问题,故使用SMOTE算法对训练数据进行处理,以下将称这种数据为SMOTE训练集.其次分别以原始训练集、SMOTE训练集分别通过3折交叉验证,同时选取各模型对应参数.然后结合训练集、所选参数及模型对测试集进行建模、预测,进而比较不同训练数据集下训练所得模型之间的分类效果.最终发现,在使用SMOTE训练集来训练模型时,贝叶斯可加分类树和随机森林能够很好地识别违约的借款人,不过他们的AUC值没有显著变化,支持向量机的AUC值显著变大;贝叶斯可加分类树的分类准确率和AUC值都比随机森林、支持向量机大,其两类误分类率均比随机森林、支持向量机小;随机森林和贝叶斯可加分类树对应的ROC曲线明显位于支持向量机所对应ROC曲线的左上方,而随机森林对应的ROC曲线几乎与贝叶斯可加分类树对应的ROC曲线重合;这些都说明在SMOTE训练集下,贝叶斯可加分类树在评估借款人的信用风险方面具有很好的效果。
[Abstract]:P2P network lending, commonly known as P2P network loans, started in the United Kingdom. In recent years, with the rapid popularization of Internet technology, P2P network loans are developing rapidly at home and abroad, showing a variety of forms. P2P network loans are small loans, low investment barriers, popular favor, but the platform runs. Borrowers default and other events emerge in endlessly, and thus get widespread attention. In this paper, two commonly used machine learning methods and a new Bayesian method are used to evaluate the credit risk of Chinese P2P loan borrowers. The three methods are: support vector machine (SVM), random forest forest (RFF), and Bayes additive classification tree (BACTT). First, the raw data is divided into training data and test data, and then the training data is divided into three parts, called the original training set. Considering the class imbalance of the data, the SMOTE algorithm is used to process the training data, which will be called the SMOTE training set. Secondly, the original training set and the SMOTE training set are respectively verified by 3 fold cross-validation, and the corresponding parameters of each model are selected at the same time. Then the test set is modeled and predicted with the training set, the selected parameters and the model, and then the classification effect between the training models under different training data sets is compared. Finally, it is found that when using SMOTE training set to train the model, Bayes plus classification tree and random forest can identify the defaulting borrowers well, but their AUC value does not change significantly, and the AUC value of support vector machine increases significantly. The classification accuracy and AUC value of the Bayes additive classification tree are higher than those of the random forest, and the support vector machine (SVM) is larger than the random forest, and the misclassification rate of the two kinds of tree is smaller than that of the random forest, and the support vector machine (SVM) is smaller. The ROC curve of random forest and Bayes additive classification tree is obviously located at the upper left of the ROC curve corresponding to support vector machine, while the ROC curve of random forest almost coincides with ROC curve corresponding to Bayes additive classification tree. These results show that the Bayes additive classification tree is effective in assessing the credit risk of the borrower under SMOTE training set.
【学位授予单位】：广州大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：F832.4;F724.6

【参考文献】