Research On Financial Risk Early Warning Model Based On Improved SMOTE

With the advancement of the process of economic globalization and the increasingly complex and changeable business environment at home and abroad,the survival of enterprises is also facing more challenges,and accurate and efficient management decisions can bring a strong boost to the sustainable business development of enterprises.Financial risk early warning has always been one of the important research topics of enterprise management and decision-making.Timely discovery of the existence of financial risks and advance resolution or avoidance of risks can reduce the probability of falling into financial crisis for enterprises and avoid certain investment losses for investors.Nowadays,artificial intelligence technology,machine learning and deep learning methods are widely used in the research of various disciplines.The research of financial risk early warning model has also developed from the stage of small sample and univariate analysis to the stage of large sample and application of various machine learning methods today.In view of the characteristics of sample imbalance of financial risk early warning scenarios,this paper starts from the perspective of improving sample imbalance and aims to build a classification model with good generalization performance to carry out research on financial risk early warning models.First of all,according to the set sample selection principles and standards,this paper selects and processes a total of 6,736 data of China's manufacturing listed companies from 2016 to2020 from CSMAR database,involving 1,608 companies,of which 118 are ST companies.The index system of the sample set consists of 51 indicators,including financial indicators and non-financial indicators.Then,according to the characteristics of the Logistic regression model and the shortcomings of the classical SMOTE(Synthetic Minority Oversampling Technique)algorithm,FW?SMOTE(Feature Weighting SMOTE)oversampling algorithm considering Feature weights is proposed to synthesize auxiliary samples.Three public datasets were used as experimental validation data,₁,₂ and₈0)(69))were used as evaluation indicators,and SMOTE,Borderline?SMOTE1 and ADASYN(Adaptive Synthetic Sampling Approach)over-sampling algorithms were compared.It is found that FW?SMOTE algorithm can effectively improve the classification performance of Logistic regression model in unbalanced sample set.Subsequently,FW?SMOTE?LR(Logistic Regression based on Feature Weighting SMOTE)and FW?SMOTE?BBLR(Balanced Bagging Logistic Regression based on Feature Weighting SMOTE)financial risk early warning models were constructed based on FW?SMOTE over-sampling algorithm according to the two conditions,and the results of 5-fold-cross validation were compared with those of LR(Logistic Regression)and BBLR(Balanced Bagging Logistic Regression)models based on SMOTE,Borderline?SMOTE1and ADASYN respectively.In the experiment,AUC,₁ and₂ were used as evaluation indicators,and it was found that LR and BBLR models based on FW?SMOTE has a better comprehensive warning effect.Finally,the advantages and disadvantages of FW?SMOTE?LR model and FW?SMOTE?BBLR model are compared.