银行信用评估模型效果的量化比较

信用评估模型能有效提高信用评估过程的科学性与结果的准确性.本文围绕主流信用评估模型在性能方面的差异化特征,基于德国信贷数据集、我国个人经营贷数据集与小微企业贷数据集,从六个模型性能评价维度对十二个代表性信用评估模型的拟合能力与泛化能力进行了深入研究.研究发现:(1)逻辑回归模型的总体性能最为优异,其次为判别分析、反向传播神经网络模型,其中逻辑回归模型与反向传播神经网络模型更适用于我国信贷场景;(2)基于无监督学习理论的自组织特征映射神经网络和k均值聚类模型,以及基于惰性学习理论的k最近邻模型的泛化能力较弱,表明各类有监督式主动学习模型更适用于解决信用评估问题;(3)模型理论与结构的复杂性并不必然能够使其在特定应用场景下获得较优的性能评价,结构简单、可解释性更强的模型往往稳健性更好.

A Quantitative Comparison of the Effect of Bank Credit Evaluation Models

Credit evaluation model can effectively improve the scientific nature of credit evaluation process and the accuracy of evaluation results.What are the differences in the performance of the current mainstream credit evaluation models?Based on the German credit data set,domestic personal business loan data set and smallµ enterprise loan data set,this paper fully studies the fitting ability and generalization ability of 12 representative credit evaluation models in six model performance evaluation dimensions.The results show that:Logistic model had the best performance,followed by DA model and BP-ANN model.At the same time,logistic model and BP-ANN model are more suitable for domestic credit scenarios;SOM-ANN model,K-means model based on unsupervised learning theory and k-NN model based on inert learning theory have shown weak generalization ability,which indicates that supervised active learning models are more suitable for solving credit evaluation problems;The complexity of model theory and structure does not necessarily make it possible for them to solve credit evaluation problems,models with simple structure and stronger interpretability are more robust.