Transformer-based attention network for stock movement prediction

Stock movement prediction is an important field of study that can help market traders make better trading decisions and earn more profit. The fusion of text from social media platforms such as Twitter and actual stock prices is an effective but difficult approach for stock movement prediction. Although some previous methods have explored this approach, there are still difficulties with the temporal dependence of financial data and insufficient effectiveness of fusing text and stock prices. To solve these problems, we propose the novel Transformer Encoder-based Attention Network (TEANet) framework, which is based on precise description through small-sample feature engineering and uses a small sample of 5 calendar days to capture the temporal dependence of financial data. In addition, this deep learning framework uses the Transformer model and multiple attention mechanisms to achieve feature extraction and effective analysis of financial data to achieve accurate prediction. Extensive experiments on four datasets demonstrate the effectiveness of our framework. Further simulations show that an actual trading strategy based on our proposed model can significantly increase profit and has practical application value. • The stock market is a time series problem, leading to temporal dependence. • A small-sample feature engineering network framework is proposed. • The fusion of the Transformer and various attention mechanisms is introduced. • Transformer model is used to extract deep features.