Local-constraint transformer network for stock movement prediction

Abstract

Stock movement prediction is to predict the future movements of stocks for investment, which is challenging both for research and industry. Typically, stock movement is predicted based on financial news. However, existing prediction methods based on financial news directly utilise models for natural language processing such as recurrent neural networks and transformer, which are still incapable of effectively processing the key local information in financial news. To address this issue, local-constraint transformer network (LTN) is proposed in this paper for stock movement prediction. LTN leverages transformer network with local-constraint to encode the financial news, which can increase the attention weights of key local information. Moreover, since there are more difficult samples in financial news which are hard to be learnt, this paper further proposes a difficult-sample-balance loss function to train the network. This paper also researches the combination of financial news and stock price data for prediction. Experiments demonstrate that the proposed model outperforms several powerful existing methods on the datasets collected, and the stock price data can assist to improve the prediction.