Improving stock price forecasting using a large volume of news headline text

Abstract

Previous research in the area of using deep learning algorithms to forecast stock prices was focused on news headlines, company reports, and a mix of daily stock fundamentals, but few studies achieved excellent results. This study uses a convolutional neural network (CNN) to predict stock prices by considering a great amount of data, consisting of financial news headlines. We call our model N-CNN to distinguish it from a CNN. The main concept is to narrow the diversity of specific stock prices as they are impacted by news headlines, then horizontally expand the news headline data to a higher level for increased reliability. This model solves the problem that the number of news stories produced by a single stock does not meet the standard of previous research. In addition, we then use the number of news headlines for every stock on the China stock exchange as input to predict the probability of the highest next day stock price fluctuations. In the second half of this paper, we compare a traditional Long Short-Term Memory (LSTM) model for daily technical indicators with an LSTM model compensated by the N-CNN model. Experiments show that the final result obtained by the compensation formula can further reduce the root-mean-square error of LSTM.