A Study On Personal Credit Risk Assessment Methods And Applications For Consumer Credit

Severe impacts have taken palce with the rapid development of the COVID-19,with economic growth being decreased and peoples’ lifestyle being reshaped.In this context,plenty of polices have been conducted to invigorate the potential domestic consumption market.In recent years,consumer credit service has developed rapidly in China,which is of great importance for improving the inclusiveness of consumer credit and activating internal consumption demand.However,in the context of uncertain policy and unstable economic situation,credit risk for the consumption loan has risen sharply,and severe default events have emerged ceaselessly.Therefore,it has become a hot research topic in terms of assessing the credit risk for personal comsumer loans,and mining the credit risk information contained in credit data.Accurate credit assessment models are indeed of intensive demand for commercial banks and other financial institutions.This paper focuses on the research topic on personal credit risk assessment in the consumer credit business,and in the view of core credit paraments in consumer credit application and default stages,we proposed three credit assessment models concerning the probability of default and loss given default.Hence,three main research parts are included in this thesis:1.Research on credit risk assessment of consumer credit based on multi-task learning.Concerning the shortcomings of the individual credit risk assessment literature,such as single research perspective,insufficient use of related information,and one-sided model evaluation indicators,this paper,based on the characteristics of consumer credit,from the dual perspectives of risk and income,mines the risk and income related information contained in consumer credit data.,using the multitask random forest algorithm to build a quantitative assessment model of consumer credit personal credit risk.Firstly,based on the default situation of consumer credit samples and the information of project internal rate of return,a multi-task target variable is constructed;secondly,considering the information correlation among multiple tasks and the fitting optimization problem,the loss function of the multitask learning algorithm is designed;finally,Based on the random forest algorithm after reconstructing the loss function,a multi-task assessment model of personal credit risk is constructed.In addition,in order to comprehensively evaluate the performance of the credit risk assessment model,this paper constructs two indicators,the default reduction rate and the profit improvement rate,based on the cost loss matrix.The empirical research on domestic real consumer credit data found that the method proposed in this paper is significantly better than other comparative credit risk assessment models in multiple evaluation indicators,and the conclusion has good robustness.2.Research on reject inference of consumer credit based on feature transfer learning.This part of the content extends and expands the research on personal credit risk assessment,and further considers the problem of sample selection bias in traditional credit assessment models.In view of the shortcomings of the previous rejection inference model,such as the premise of sample distribution assumptions that are difficult to verify,the low quality of rejection sample label inference,and the strong noise interference of rejected samples,this paper takes a new approach,starting from the perspective of feature learning,and researches rejection inference methods based on transfer learning technology.This part of the work mainly includes three points: first,in order to avoid the negative transfer phenomenon,the rejected samples are screened based on the three decision theories,the noise samples are filtered out,and the transfer learning efficiency is improved;secondly,based on the filtered rejected samples,the accepted samples are introduced The sample is used as a data bias,and the sparse self-encoding network is used for risk feature transfer learning;finally,the credit data is reconstructed based on the learned highlevel credit risk representation,combined with the multi-task learning algorithm proposed in the previous section,to construct personal credit risk Evaluate the model.The empirical results show that the rejection inference method proposed in this paper based on feature transfer learning can effectively improve the performance of credit risk assessment,and is significantly better than the commonly used rejection inference models.3.Research on loss given default prediction from perspective of multigranularity default information and hybrid ensemble model.Focusing on the subject of personal credit risk assessment in consumer credit business,this paper further evaluates and predicts the credit risk of post-loan default samples on the basis of the previous two researches.In view of the shortcomings in the LGD prediction literature,such as single perspective of default sample risk distribution,insufficient utilization of relevant information of LGD,and low prediction performance of evaluation models,this paper based on the perspective of multi-granularity default information,using integrated learning algorithm and intelligent optimization technology,Conduct in-depth research on the prediction of consumer credit LGD.First of all,this paper draws on the idea of granular computing,uses the default risk exposure and default loss rate information of default samples jointly,and granulates the risk information of default samples based on genetic algorithm technology,so as to realize multi-granularity and multi-level disassembly of the risk information of default samples.Secondly,based on the granulated default samples,the multi-class integration algorithm is used to calculate the membership probability of the default samples belonging to different information granules;finally,for the risk information mining problem of the granulated subspace samples,the random forest is used.The regression algorithm quantifies them separately,and performs a hybrid integration of the subspace prediction models based on the default sample membership probability.In the empirical study of consumer credit default samples,the superiority of the method proposed in this paper is verified.In addition,this paper also uses SHAP algorithm to conduct a fine-grained analysis of the influencing factors of LGD.The study finds that borrowing information and borrower’s work information are important factors affecting the LGD rate of consumer credit.