| The development of society and economy has changed people’s consumption concept.Overspending behaviour is beginning to be noticed and more widely accepted,and people are using credit loans.Personal credit loans account for an increasing proportion in the bank loan business and have become an important business of commercial banks.However,with the rapid development of personal credit business,it also brings many risks.Among them,credit default risk is the biggest threat to credit business and it is also a risk that banks focus on.According to CBRC,the events and amounts of non-performing loans have increased in 2021,and the banking industry is facing great pressure of bad debts,so it is necessary to take measures to avoid the possible default risk.Many small and medium-sized banks still use the manual audit method when conducting loan qualification review,but the personal credit business has the characteristics of small loan amount,large number of strokes and complex loan process,relying on manual audit alone will cost huge labor costs.In addition,the online business and the explosive growth of online users have increased the difficulty of risk control.Traditional manual review and outdated model tools cannot meet the needs of risk control.Digital transformation is urgently needed.At present,machine learning algorithms have matured,and various industries have started to apply machine learning to solve practical problems,and even many Internet financial companies pay high salaries to hire big data technicians to solve financial risk problems and achieve intelligent risk control.This paper mainly studies from the following aspects.First of all,it expounds the research background and significance,discusses the research status at home and abroad,introduces the relevant theories of personal credit and the algorithm basis used.Secondly,based on the data set of personal credit default records provided by Zhongyuan Bank,the relationship between features and labels is preliminarily explored through visual description,and the data set is preprocessed,including data cleaning such as data format conversion,coding,and extracting numerical values from strings.Handling of missing and outliers;building derived features from existing features and using filtering and embedding methods for feature selection;balancing datasets through random oversampling,SMOTE sampling,and comprehensive sampling and selecting optimal sampling methods.Then,try to use the random forest in the ensemble learning bagging algorithm and XGBoost and Light GBM in the boosting algorithm to build personal credit default prediction models to help banks realize credit risk identification,adjust the parameters by grid search method,select the optimal model with accuracy,precision,recall rate,F1 score,KS value and AUC value as model evaluation indexes,and output the feature importance.The results show that most of the derived features are helpful to model building.Compared with the original features,the features screened by filtering method and embedding method can improve the model performance better.The effect of random oversampling on test set is better than SMOTE sampling and comprehensive sampling.For the empirical part,through comparison,it is found that the three models have little difference in each index,but the XGBoost algorithm and Light GBM algorithm are superior to random forest,and they can be used as risk prediction models,but considering that the KS value and the AUC value are not affected by the change of the sample proportion,and Light GBM algorithm runs faster than XGBoost algorithm,Light GBM with adjustedparameters is selected as the final credit default risk prediction model.The results of Light GBM show that debt-to-income ratio,the number of outstanding credit lines on a borrower’s file,interest,total credit revolving balance and loan amount play a key role in predicting whether the borrower defaults,and commercial banks can focus on the features that contribute greatly to the default prediction model.Therefore,the model constructed in this paper provides a good solution for realizing credit risk identification.According to the above model,the judgment of the borrower’s default possibility can be effectively improved,thereby reducing the default risk and bank losses. |
