A dual encoder-based approach to predicting stock price by leveraging online social network

We propose a dual-encoder which encodes the investor sentiment and technical indicators separately to improve the accuracy of the encoder-decoder model in predicting stock price by using two types of information. For the dual-encoder and decoder, we revise the gated recurrent unit (GRU) by removing the reset gate, using the update gate instead of the reset gate function and replacing tanh activation function with ReLU activation function to improve the speed of network training and the accuracy of the model. We regard market sentiment as a discrete-time stochastic process. When fixed time, market sentiment is a variable subject to a certain probability distribution. Sentiment score formulas are built for investor sentiment by a pseudo-label based sentiment classifier, and the market sentiment is estimated through ensemble Bagging learning. The orthogonal table experiment design is used to select parameters in our dual-encoder based model, which greatly reduces the time of parameter adjustment. Finally, experiments are conducted to show that our dual-encoder based model is more accurate than encoder-decoder model, and investor sentiment helps improve the stock forecasting in our model.