An Ensemble Model Of Credit Scoring Combined With Deep Learning Optimization Algorithm

With the rapid development of the economy,the consumer finance industry is thriving,so the financial market is facing the risk of increased credit defaults.Consequently,it is an burning problem for all financial institutions to establish an complete and effective personal credit evaluation system.As an effective tool for credit risk assessment,the credit scoring model can predict the probability of a customer default or overdue by the basic information and credit data,thereby providing decision support for financial institutions to approve credit business.In view of the shortcomings of existing credit scoring models,this thesis proposes an ensemble model integrating deep learning network,and constructs an ensemble model of credit scoring combining with deep learning optimization algorithm to realize credit risk assessment of credit customers.The research contents of this thesis are as follows:(1)We propose an ensemble model of credit scoring that integrates deep learning networks.Firstly,we introduce the recurrent neural network(RNN)and its extended form,the bidirectional recurrent neural network(BRNN),into the credit scoring field to avoid the limitations of the shallow models.Then,we integrate BRNN,logistic regression(LR)and extreme gradient boosting tree(XGBoost)to build an ensemble model of credit scoring,named as BRNN+LR+XGBoost.Finally,the experimental results have proved the applicability of deep learning algorithm in the field of credit scoring and verified the importance and effectiveness of ensemble learning.(2)We study and analyze a variety of deep learning optimization algorithms.This thesis summarizes the calculation processes,advantages and disadvantages of various deep learning optimization algorithms,and improves adaptive moment estimation(Adam),a commonly used optimization algorithm,and then constructs adaptive moment estimation with dynamic bound(Ada Bound),an adaptive optimization algorithm with dynamic learning rate bound.Experiments have proved that Ada Bound algorithm can not only maintain a fast convergence speed,but also has a good generalization ability,which meets the optimization needs of different stages in the training process.(3)We propose an ensemble model of credit scoring with Ada Bound optimization algorithm.We apply Ada Bound,a deep learning optimization algorithm,to the ensemble model BRNN+LR+XGBoost,and experiments are conducted on three real credit data sets to predict the probability of default or overdue of customers,which can realize the credit risk assessment of customers.The experimental results show that the ensemble model BRNN+LR+XGBoost with Ada Bound optimization algorithm has superior performance and can accurately classify customers' credit.Through the experiments of three credit data sets,compared with several existing credit scoring ensemble models,the AUC of the model proposed in this thesis has increased by an average of 3.1%,the accuracy has increased by an average of 2.3%,and the F-score has increased by an average of 2.2%.In summary,this thesis proposes an ensemble model for credit scoring combined with dynamic learning rate bound optimization algorithm,which can not only accurately classify the credit of customers,but also provide theoretical bases and practical reference for credit evaluation and risk management to financial institutions.