Research On Credit Scoring Based On Adaptive Ensemble Model

In recent years,with the deepening of economic globalization and financial liberalization,consumption financial services in China have developed rapidly.Credit quota of Internet Finance consumers has exploded with the overall popularity of the Internet.The continuously issued regulatory policies require a more in-depth reform and innovation in consumption financial industries.Currently,the profit mode of financial institutions has changed from the extensive credit mode based on user scale to the intensive credit mode based on user quality.Credit risk management is a key factor for financial institutions to stand out in the increasingly fierce competition.Therefore,constructing a more scientific,reasonable and effective credit scoring model has become an important part of current credit scoring research.Credit scoring is essentially a classification problem which aims at distinguishing good credit applicants from all applicants.Many domestic and abroad scholars have carried on deep research and discussion of credit scoring,and put forward various credit scoring methods.However,some disadvantages,such as lack of using hybrid feature selection model for feature selection need to be further addressed.Based on the extant research,this study first elaborated the background and significance of credit scoring,and then systematically summarized the research status of credit scoring from domestic and abroad perspective.In addition,feature selection,classifier algorithm and classifier ensemble are introduced in detail.Two novel self-adaption credit scoring models are proposed as follows.(1)This study proposed a credit scoring ensemble model adapting of different imbalance ratios.Firstly,an extended BalanceCascade approach was designed to deal with the imbalanced data beyond the threshold and effectively solve the interference of imbalanced data to the proposed model;then,two DT-based ensemble models(Random Forest and XGBoost)were used as base classifiers for ensemble model to effectively guarantee the prediction performance of the ensemble model.;finally,stacking approach was applied for classifier ensemble to construct credit scoring ensemble model with higher prediction performance.(2)This study proposed a multi-stage self-adaption credit scoring ensemble model.First,this study proposed an enhanced multi-population niche genetic algorithm(EMPNGA)for feature selection;second,several filter methods are combined with EMPNGA to construct a hybrid feature selection method for adaption of feature selection;third,forming a candidate classifier repository containing several independent base classifiers and employing the enhanced EMPNGA and combining the priori knowledge of base classifiers to select the optimal classifier subset;finally,using stacking method to stack the obtained optimal classifier subset and construct hybrid credit scoring ensemble model with higher prediction performance.Four datasets and four evaluation metrics are applied to validate the comprehensive performance of these two models for credit scoring.Experimental results indicate that:(1)extended BalanceCascade can improve prediction performance in imbalanced ratios dataset;(2)average performance of models with feature selection is superior to that without feature selection;(3)the optimization performance of EMPNGA is superior to that of binary-encoded genetic algorithm,and binary particle swarm optimization algorithm;(4)compared to five comparative features selection methods,credit scoring model combining with HEMPNGA algorithm can obtain optimal prediction results;(5)the prediction results of multi-stage self-adaption credit scoring model is superior to that of comparative models.In all,both the models and algorithms proposed in this study have obtained superior performance.This study is one of the innovations and development in credit scoring,which provides a new research perspective for theoretical research in credit scoring.It has good theoretical significance to this extant.In addition,this study is conducive to preventing systematic financial risks in financial consumption industries,improving financial institutions competitiveness,and providing technical support for financial institutions to realize rapid,rational and efficient loan approval,which is of great practical significance.