Attention enhanced long short-term memory network with multi-source heterogeneous information fusion: An application to BGI Genomics

Abstract

The recent availability of enormous amounts of both data and computing power has created new opportunities for predictive modeling. This paper compiles an analytical framework based on multiple sources of data including daily trading data, online news, derivative technical indicators, and time–frequency features decomposed from closing prices. We also provide a real-life demonstration of how to combine and capitalize on all available information to predict the stock price of BGI Genomics. Moreover, we apply a long short-term memory (LSTM) network equipped with an attention mechanism to identify long-term temporal dependencies and adaptively highlight key features. We further examine the learning capabilities of the network for specific tasks, including forecasting the next day’s price direction and closing price and developing trading strategies, comparing its statistical accuracy and trading performance with those of methods based on logistic regression, support vector machine, gradient boosting decision trees, and the original LSTM model. The experimental results for BGI Genomics demonstrate that the attention enhanced LSTM model remarkably improves prediction performance through multi-source heterogeneous information fusion, highlighting the significance of online news and time–frequency features, as well as exemplifying and validating our proposed framework.