| Research on risk prediction of unbalanced financial data based on integrated learning The current economic situation is changing rapidly.Whether it is the change of the capital market or the change of the company’s operating conditions,it may bring certain financial risks to the company,thus damaging the interests of investors.Therefore,in order to reduce the losses caused by potential financial risks,it is necessary to establish an accurate financial risk early warning system to monitor the company’s financial situation,so as to take appropriate measures to avoid losses as soon as possible.Most of the current research is to directly use financial statement data to predict financial risks,without considering that most of the financial data are unbalanced data,and the results obtained by direct use are not as expected.However,the research on data imbalance also has shortcomings,such as not considering the spatial location of different categories of samples,the cost of misclassification of different categories of samples is difficult to determine,and the algorithm is too complex.To solve these problems,the following work has been done in this paper:1.The noise reduction algorithm is studied.First,the original data is preprocessed with missing value filling,feature filtering,normalization,etc.Then a cluster-based noise reduction algorithm is used for undersampling,and the training set and test set are divided.The experiment shows that the recall rate of the noise reduction algorithm is increased by 20 percentage points and the F1-score is increased by 34 percentage points compared with the direct use of raw data.Compared with the traditional random undersampling method,the recall rate is increased by 5 percentage points,so this method has certain advantages.2.The combination model and data enhancement are studied,and four generation models,GAN,WGAN-GP,CycleGAN and VAE,are proposed to enhance the data of a few samples,so as to achieve the purpose of oversampling.The four classification models of Adaboost,Lightgbm,Xgboost and Catboost are combined with the four generation models in pairs,and the best model combination is obtained by comparison.The experiment shows that the prediction result of the combination of CycleGAN-Catboost is the best.The recall rate for a few categories reaches 80%,the accuracy rate reaches 95%,and the F1-score reaches 88%.Comparing this result with the traditional oversampling methods SMOTE,BSMOTE and ADASYN,it is found that the CycleGAN-Catboost combination model is superior to the traditional method in the four indicators of recall,accuracy,F1-score and AUC.3.A DS information fusion prediction model is proposed.Support vector machine is introduced to integrate with the optimal combination CycleGAN-Catboost obtained from oversampling.The prediction probabilities of the two models for different categories are output respectively.The joint evidence trust is calculated according to the DS information fusion calculation formula,and the prediction category is determined by comparing the threshold and the joint evidence trust.At the same time,the influence of the threshold on the prediction results is discussed.According to the actual needs,0.1 is set as the threshold.Experiments show that the DS integration method is better than the single classification model,which improves the recall rate by 8 percentage points,and improves the accuracy rate,F1-score and AUC to a certain extent. |
