Many studies have been carried out on stock price trend prediction, but most of them focused on the public market data and did not utilize the trading behaviors owing to the unavailability of real transaction records data. In fact, trading behaviors can better reflect the market movements, and the fusion of trading information and market information can further improve the prediction accuracy. In this paper, we propose a deep neural network model using the desensitized transaction records and public market information to predict stock price trend. Considering the correlation between stocks, our method utilizes the knowledge graph and graph embeddings techniques to select the relevant stocks of the target for constructing the market and trading information. Given the considerable number of investors and the complexity of transaction records data, the investors are clustered to reduce the dimensions of the trading feature matrices, and then the matrices are fed into the convolutional neural network to unearth the investment patterns. Eventually, the attention-based bidirectional long short-term memory network can predict the stock price trends for financial decision support. The experiments on the price movement direction and trend prediction show that our method achieves the best performance in comparison with other prediction baselines.

已经对股票价格趋势预测进行了许多研究，但是由于缺乏实际交易记录数据，因此大多数研究都集中在公开市场数据上，并且没有利用交易行为。实际上，交易行为可以更好地反映市场动态，交易信息与市场信息的融合可以进一步提高预测准确性。在本文中，我们提出了一个利用脱敏交易记录和公开市场信息来预测股票价格趋势的深度神经网络模型。考虑到股票之间的相关性，我们的方法利用知识图和图嵌入技术选择目标对象的相关股票来构建市场和交易信息。鉴于大量的投资者和交易记录数据的复杂性，然后，将投资者聚类以减少交易特征矩阵的维数，然后将这些矩阵馈入卷积神经网络以挖掘投资模式。最终，基于注意力的双向长期短期记忆网络可以预测股票价格趋势，从而为财务决策提供支持。价格走势和趋势预测的实验表明，与其他预测基准相比，我们的方法可获得最佳性能。