A News-Driven Recurrent Neural Network for Market Volatility Prediction

Extracting hidden information embedded in the financial news is an effective approach to market volatility prediction. In this paper, we propose a recurrent neural network (RNN) based method that dynamically extracts latent structures from the sequence of news events for market prediction. Specifically, we first train a skip-thought model on the financial news datasets to represent the semantic meaning of sentences. Then we aggregate the representations from a single day to form a daily feature to align with the market index. Finally, to make use of the news released some days ago for a better prediction, we exploit the long short-term memory RNN (LSTM-RNN) to integrate the information by exploring the dynamic patterns embedded in the sequence of news events. Extensive experiments on the public available Reuters and Bloomberg financial news datasets verified the effectiveness of our method and demonstrated that our method achieves the state-of-the-art performance.