Research On Financial Credit Forecasting Based On Ensemble Decision Tree

Issuing loans to potential users is the core business of lending institutions around the world.The loan business brings considerable benefits to the organization,but also makes the company bear the huge risk of financial losses due to the default of the borrower.Therefore,the lending institution needs to comprehensively analyze the basic information and credit history submitted by the applicant to assess the possibility of repayment,and decide whether to approve the application.Although there are relatively mature prediction methods in this field,in the face of the inconsistency of credit records of various institutions,credit data has complex characteristics,a large number of missing values,and sparse data.Research on the prediction of defaults applied by borrowers Still very challenging.The ensemble decision tree model is a powerful regression and classification method,which has made outstanding achievements in data competitions for many times.It has attracted wide attention of researchers for its good explanatory ability in financial credit prediction.Appropriate hyper-parameter setting is the necessary basis for good performance of the model.In order to ensure the accuracy of the credit prediction model,we use the improved adaptive optimization algorithm to optimize the limit gradient lifting tree and establish the credit scoring model.In addition,the traditional prediction models mainly based on hard information data modeling usually ignore the importance of soft information in credit application to predict default.In order to supplement the limitations of hard information features in credit model,we propose to use multi-granularity text feature extractor to extract text information features from word level granularity and sentence level granularity,and combine soft and hard information to enrich the data support and prediction basis of credit model.In view of the above problems,this paper constructs an ensemble decision tree model based on improved hyper-parameter optimization to predict default.With more reasonable super parameter settings,the model obtains more accurate prediction results.In addition,the text constructs an integrated decision tree model for credit prediction based on multi-granularity text feature extraction combined with soft and hard information.Compared with the traditional prediction model,this model can use the characteristics of text soft information to improve the prediction performance of the model.The main contributions of this paper are:Firstly,based on the optimization algorithm of adaptive learning strategy,the hyper-parameters of the limit gradient boosting tree are optimized,and the credit scoring model is established to accurately determine the optimal setting of the super-parameters of the model.The model structure more matches the characteristics of financial credit data,the model makes more accurate prediction and judgment.Secondly,the proposed improved particle swarm optimization algorithm is based on adaptive subgroup division,and uses two different learning strategies to update different types of particles to enhance the diversity of particle populations.The improved optimization algorithm can average the information of local optimal particles to prevent premature convergence caused by local aggregation of particles.Thirdly,a new credit risk assessment model based on text feature extraction of multi-granularity analysis is proposed.The model can extract multi-granularity feature of credit text,extract text context semantics and word level semantics,and improve the richness of text features.The model uses the feature combination characteristics of decision tree to integrate soft and hard information to achieve the purpose of accurate prediction of credit data.