论文部分内容阅读
本文运用PCA方法提取出对信用风险具有显著影响的特征指标,同时运用EP-T方法离散样本数据并学习贝叶斯网络的结构与参数,以此建立朴素贝叶斯网络(Naive Bayesian Network,NB)信用风险预警模型;最后通过交叉验证(Cross Validation)对模型进行5次独立建模测试,并利用性能评价指标将NB模型与Logistic模型、MLP神经网络模型、RBF神经网络模型进行对比分析。实证研究结果表明,尽管四种模型均能对上市公司信用风险进行预警,但NB模型表现出了更好的预测精度与稳定性。
In this paper, the PCA method is used to extract the characteristic indexes that have a significant impact on the credit risk. At the same time, we use the EP-T method to discrete the sample data and learn the structure and parameters of the Bayesian network to establish the Naive Bayesian Network (NB ) Credit risk early warning model. Finally, five independent modeling tests were conducted on the model through Cross Validation. The NB model was compared with Logistic model, MLP neural network model and RBF neural network model by performance evaluation index. The empirical results show that, although all four models can predict the credit risk of listed companies, the NB model shows a better prediction accuracy and stability.