Credit Default Prediction Model Based on Support Vector Mach

发布时间：2023-06-05 01:22

　　信贷审批数据建模是信贷行业的重要研究课题。随着这一领域的快速发展,信用违约预测(CDP)分类器被广泛应用于对客户信用评估和大额贷款组合的审批。违约风险预测数据建模是模式识别理论背景下的二元分类问题,其目的是将新的观察结果传递给预先定义的决策类。信用风险决策的结果受两个主要因素的影响:选择准确的特征选择算法来寻找合理的指标特征集,选择合适的分类模型来构造进行决策。财务风险管理是用于探索许多重要参数的最具研究前景的问题之一。银行业包括影响银行及其利益相关者的众多风险因素。CDP与银行有着紧密的联系,它是一种适用于银行资金借贷的有效而决定性的技术。获取积累有关债权人、监管机构、其他金融和非金融公司、政府等方面的数据,对信用风险进行监管是非常重要的。同时,CDP对于向客户提供贷款的集中评估也是十分重要的。此外,CDP方法有助于将信誉好的客户与信誉差的客户区分开来。这意味着,一些信贷客户拥有良好的信用资质;相应地,银行可以将他们归类为“有偿付能力的债权人”。相反,还有一些没有良好信用资质的客户,因此被归类为“无偿付能力的债权人”。然而,值得注意的是,这种直接的分类程序可能无法提供最佳的信用风险管...

【文章页数】：182 页

【学位级别】：博士

【文章目录】：
Abstract
摘要
1 Introduction
    1.1 Background
    1.2 Research Motivation
    1.3 Research Ideas and Methods
    1.4 Research Questions and Research Objectives
    1.5 Research Content and Structure Arrangement
    1.6 Relation and Differences of Main Contents
        1.6.1 Differences based on Methodology
        1.6.2 Differences based on Characteristics
    1.7 Research Contributions
2 Literature Review
    2.1 Credit Default Risk Prediction:A Theoretical Background
        2.1.1 Credit Default Definitions
        2.1.2 Five Cs Good Lending Concept
        2.1.3 Judgmental Systems Versus Credit Default Prediction Systems
        2.1.4 Risk Management
        2.1.5 Sound Default Risk Management
        2.1.6 Benefits and Criticism of Credit Default Risk Prediction
    2.2 Empirical Literature
        2.2.1 Individual Feature Selection to Default Risk Prediction Based onSupport Vector Machine
        2.2.2 Group Feature Selection to Default Risk Prediction Based on SupportVector Machine
        2.2.3 Hybrid Model for Default Risk Prediction Based on LogitSVM andLogitNeural Algorithms
3 Individual Feature Selection to Default Risk Prediction Based on Support VectorMachine
    3.1 Background
    3.2 Motivation
    3.3 Individual Feature Selection Models
        3.3.1 T-test Approach
        3.3.2 Discriminant Analysis Approach
        3.3.3 Logistic Regression Approach
        3.3.4 CHAID Decision Tree
        3.3.5 QUEST Decision Tree
    3.4 Default Risk Prediction Model
        3.4.1 Support Vector Machine
    3.5 Dataset
        3.5.1 Data Division
    3.6 Performance Measure
    3.7 Empirical Results
        3.7.1 Selecting Significant Features
        3.7.2 Discriminant Analysis
        3.7.3 Logistic Regression
        3.7.4 Decision Trees
        3.7.5 Support Vector Machines
        3.7.6 Type Ⅰ Error, Type Ⅱ Error, and EMCC
        3.7.7 Comparisons of Model's Predictability
        3.7.8 Relative Importance of Selected Features
    3.8 Summary
        3.8.1 Main Results
        3.8.2 Main Conclusion
        3.8.3 Main Characteristics
4 Group Feature Selection to Default Risk Prediction Based on Support VectorMachine
    4.1 Background
    4.2 Motivation
    4.3 'New Age' Group Feature Selectors
        4.3.1 Ridge Regression
        4.3.2 Least Angle Regression
        4.3.3 Lasso (Least Absolute Shrinkage and Selection Operator)
        4.3.4 Gradient Boosted Feature Selection
        4.3.5 Random Forest
    4.4 Default Risk Prediction Models
        4.4.1 Support Vector Machine
        4.4.2 Multilayer Perceptron
        4.4.3 Radial Basis Function
        4.4.4 Classification and Regression Tree
    4.5 Datasets
        4.5.1 Cross-Validation
    4.6 Model's Parameter
    4.7 Performance Measure
    4.8 Statistical Significance Test
    4.9 Empirical Results
        4.9.1 Significant Feature Sets
        4.9.2 Results from Different Datasets
        4.9.3 Credit Default Prediction Average Results
        4.9.4 Cost of Default Prediction Errors
        4.9.5 Verification of Feature Importance
        4.9.6 Robustness Check
    4.10 Summary
        4.10.1 Main Results
        4.10.2 Main Conclusion
        4.10.3 Main Characteristics
5 Hybrid Model for Default Risk Prediction Based on LogitSVM and LogitNeuralModels
    5.1 Background
    5.2 Motivation
    5.3 Default Risk Prediction Hybrid Models
        5.3.1 Logistic Regression
        5.3.2 Neural Network Architecture
    5.4 Datasets
        5.4.1 Training Schemes (TSs)
    5.5 Performance Measure
    5.6 Empirical Results
        5.6.1 Model Prediction
        5.6.2 Type Ⅰ and Type Ⅱ errors with their Corresponding Cost-Benefit Scores
        5.6.3 Selecting the Optimal TS Ratio
        5.6.4 The Most Contributed Feature
        5.6.5 Comparison to the Perfect Models
        5.6.6 LSVM, LNA, SVM, and BPN:A Global Comparison
    5.7 Summary
        5.7.1 Main Results
        5.7.2 Main Conclusion
        5.7.3 Main Characteristics
6 Conclusion
    6.1 Main Conclusion
    6.2 Main Findings
    6.3 Main Contributions
    6.4 Policy Implications
    6.5 Future Roadmaps
References
Appendix A
Appendix B
Publications during PhD Period
Acknowledgement
Curriculum Vitae



本文编号：3831404