基于改进Transformer和超图模型的股票趋势预测方法研究

Stock trend prediction method based on improved Transformer and hypergraph model

股票预测是一项令人痴迷又极具挑战的任务。近年来,融合关系信息的股票时序预测方法取得一些进展,但仍存在如下问题:首先,基于图神经网络的方法仅考虑股票之间简单的成对关系,而未考虑股票间的高阶协同关系。其次,现有方法采用预定义图的方式直接给出股票间的静态关系,无法建模股票间潜在的动态变化关系。为了解决上述问题,提出一种端到端的动态超图卷积神经网络股票趋势预测框架。该框架基于改进的Transformer提取股票的时序信息,通过静态超图和动态超图将股票间的协同关系信息引入到时序建模中。在中国A股和美股市场数据集上的实验结果表明,与当前先进模型相比,本文模型的预测性能具有显著优势。

Stock forecasting is an interesting but extremely difficult task.Stock time series prediction methods that incorporate relationship information have progressed in recent years,but the following issues remain.First,graph neural network-based methods consider simple pairwise relationships between stocks but ignore higher-order collaborative relationships.Second,most existing methods utilize the static relationship among stocks with predefined graphs,and modeling the potential changes in dynamic relationships among stocks is challenging.To address the abovementioned issues,a Dynamic HyperGraph Convolutional neural Network(DHGCN)framework for the end-to-end stock trend prediction is proposed.The temporal information of stocks is captured by an improved Transformer model,and the collaborative relationship information is integrated into the time series modeling by the static and dynamic hypergraphs.Experiments on the real-world datasets of the Chinese A-share market and the US stock market show that the prediction performance of the proposed model is significantly superior to that of the contemporary advanced models.