Attention-Based Event Relevance Model for Stock Price Movement Prediction

Abstract

Stock prices, in general, can be affected by world events such as wars, natural disasters, government policies, etc. However, the correlations between events and stock prices are often implicit and the influences of events on stock prices can be in indirect ways and act in chain reactions, which brings essential difficulties for precise market prediction. In this paper, we propose an attention-based event relevance model (ATT-ERNN) to explicitly model event relevance for predicting stock price movement. Specifically, in our model, we use long short-term memory neural network (LSTM) and convolution neural network (CNN) to encode event information and stock information to distributional representations. After that, we employ attention mechanism to find related events for each stock to do price movement prediction. Attention weights in our model have a quantitative interpretation as the relevance degree of events affecting the price of a specific stock. We have conduct extensive experiments on a manually collected real-world dataset. Experimental results show the superiority of our model over many baselines, which proves the effectiveness of our model in this prediction problem.