中国股票市场操纵识别研究——基于机器学习分类算法

Manipulation Recognition of China’s Stock Market:Based on Machine Learning Classification Algorithm

本文整理了2006—2021年证监会行政处罚涉及的股票市场操纵案例,并通过Wilcoxon秩和检验来筛选构造解释变量,之后综合运用各种过采样算法和机器学习模型对其进行实证分析。研究发现:第一,经过过采样算法扩充样本的模型预测精度明显大于样本不平衡的模型;第二,综合比较各种过采样算法,Borderline-SMOTE过采样算法的预测精度大于SMOTE和ADASYN过采样算法;第三,综合比较各类机器学习分类模型,SVM模型的预测精度明显大于其他机器学习分类模型。本文结论对股票市场操纵的及早识别预测,促进资本市场良性发展具有一定的理论意义和实践价值。

We sort out the cases of stock market manipulation involved in the administrative punishment of the CSRC from 2006 to 2021,selecting and constructing explanatory variables through Wilcoxon rank sum test.Then we make an empirical analysis by comprehensively using various over-sampling algorithms and machine learning models.We find the prediction accuracy of the model expanded by over-sampling algorithm is significantly higher than that of the model with unbalanced samples.Besides,we also find that the prediction accuracy of Borderline SMOTE over-sampling algorithm is greater than that of SMOTE and ADASYN.What’s more,by comprehensively comparing various machine learning classification models,we conclude the prediction accuracy of SVM model is significantly better than that of other machine learning classification models.Our conclusion has certain theoretical significance and practical value for the early identification and prediction of stock market manipulation and the promotion of the healthy development of the capital market.