Stocks in the Chinese stock market can be divided into ST stocks and normal stocks, so to prevent investors from buying potential ST stocks, this paper first performs SMOTEENN oversampling data preprocessing for the S...Stocks in the Chinese stock market can be divided into ST stocks and normal stocks, so to prevent investors from buying potential ST stocks, this paper first performs SMOTEENN oversampling data preprocessing for the ST stock category, and selects 139 financial indicators and technical factor as predictive features. Then, it combines the Boruta algorithm and Copula entropy method for feature selection, effectively improving the machine learning model’s performance in ST stock classification, with the AUC values of the two models reaching 98% on the test set. In the model selection and optimization, this paper uses six major models, including logistic regression, XGBoost, AdaBoost, LightGBM, Catboost, and MLP, for modeling and optimizes them using the Optuna framework. Ultimately, XGBoost model is selected as the best model because its AUC value exceeds 95% and its running time is less. Finally, the XGBoost model is explained using the SHAP theory and the interaction between features is discovered, further improving the model’s accuracy and AUC value by about 0.6%, verifying the effectiveness of the model.展开更多
文摘Stocks in the Chinese stock market can be divided into ST stocks and normal stocks, so to prevent investors from buying potential ST stocks, this paper first performs SMOTEENN oversampling data preprocessing for the ST stock category, and selects 139 financial indicators and technical factor as predictive features. Then, it combines the Boruta algorithm and Copula entropy method for feature selection, effectively improving the machine learning model’s performance in ST stock classification, with the AUC values of the two models reaching 98% on the test set. In the model selection and optimization, this paper uses six major models, including logistic regression, XGBoost, AdaBoost, LightGBM, Catboost, and MLP, for modeling and optimizes them using the Optuna framework. Ultimately, XGBoost model is selected as the best model because its AUC value exceeds 95% and its running time is less. Finally, the XGBoost model is explained using the SHAP theory and the interaction between features is discovered, further improving the model’s accuracy and AUC value by about 0.6%, verifying the effectiveness of the model.
