基于预训练模型的A股停牌预测研究

The Research on Prediction of Trading Halts in A-Share Based on Pre-trained Model

下载PDF

导出

摘要股票停牌是基于提高信息披露程度、抑制股票剧烈波动、消除信息不对称等目的所采取的机制,但个股的停牌行为存在随意性以及其复牌时间不确定性的问题,对投资者的合法权益造成了较大的损害。本文研究从股票波动中挖掘停牌的内在规律,提出了基于预训练模型的股票停牌预测机制。首先预训练模型学习A股上市公司股票停牌的共性特征,然后通过预训练模型参数的迁移学习获得个股停牌的特征,进而构建特定个股的停牌预测模型。实证分析选取2 539家上市公司作为研究对象,对比分析了多个深度学习组合模型,选择以TADM(TCN-Attention-Dense Model)网络结构作为预训练模型具有较好的总体预测效果,在此基础上个股的迁移学习模型表现出更强的样本识别能力且误报率更低。研究发现,虽然诱发个股停牌的因素很多,但仍能从股票波动中预测停牌事件。研究成果为投资者规避停牌风险、减少投资损失或利用个股停牌实现收益最大化等提供借鉴,并可为监管机构对停牌趋势做出前瞻性判断,对宏观调控和政策调整提供参考。 Trading halts in stock market is a mechanism designed to improve the degree of information disclosure, restrain stock volatility and eliminate information asymmetry.However, uncertainty associated with the suspension and resumption of stocks can cause great trouble and loss to investors.The paper studies mechanism of trading halts from perspective of stock fluctuations, and proposes a prediction pre-trained model of trading halts.First, the pre-trained model is used to learn the common features of A-share listed companies’ trading halts.These common features are further employed to get features of individual stock with the technique of transfer learning, then the prediction model of specific stocks is constructed.In the experiment, 2 539 listed companies are selected as the research objects.Comparative analysis of multiple deep learning combination models, the pre-trained model adopts the TADM(TCN-Attention-Dense Model),which has a better overall prediction effect.Based on TADM,the transfer learning model of individual stocks holds stronger sample recognition ability and lower false positive rate.It is found that although there are many factors inducing trading halts of individual stock, trading halts can still be predicted from the stock fluctuations, which provides reference for investors to avoid the risk of trading halts and reduce investment losses, or to maximize returns using trading halts, and for regulators to make forward-looking judgments on the trend of trading halts, as well as macro-control and policy adjustment.

作者孙夫雄谢翔熊平梁嘉欣彭畅 SUN Fu-Xiong;XIE Xiang;XIONG Ping;LIANG Jia-Xin;PENG Chang

机构地区中南财经政法大学信息与安全工程学院

出处《中央财经大学学报》 CSSCI 北大核心 2022年第11期39-51,共13页 Journal of Central University of Finance & Economics

基金高等学校学科创新引智基地项目(项目编号:B21038) 中南财经政法大学研究生拔尖人才培养项目(项目编号:XKRH202101,CXJH202105)。

关键词股票停牌预训练模型迁移学习深度学习非平衡样本 Trading halt Pre-trained model Transfer learning Deep learning Unbalanced sample