Model and forecast stock market behavior integrating investor sentiment analysis and transaction data

Social network media analytics is showing promise for prediction of financial markets. However, the true value of such data is unclear due to a lack of consensus on which instruments can be predicted. In this paper, we investigate whether measurements of collective emotional states derived from large scale network feeds are correlated to the stock transaction data over time. The information space corresponding to stocks is divided into the network public opinion space $$Opinion\_Space$$Opinion_Space and the realistic transaction space $$Behavior\_Space$$Behavior_Space. We then handle the information and generate the multidimensional time series from them respectively. Furthermore, Granger causality analysis and information theory measures are used to find and demonstrate that social media sentiments contain statistically significant ex-ante information on the future prices. At last, we propose our separate-LSTM model and the experimental results of six stocks which are randomly selected indicate that financial data predictions can be significantly improved through our model by the fusion of network public opinion emotions and realistic transaction data.