Stock Forecasting using Neural Network with Graphs

Due to the complex characteristic in the stock market, it is always a challenge and interesting topic to predict stock price. With the development of neural network models, deep learning has become a popular way to solve the stock prediction problem. Many of the current studies focus on how the stock own historical information which will affect the stock price in the future. Although the individual historical features are essential, the stock price is also affected by the other stocks.

To capture such internal relations and influence, we propose to join stock graphs with the neural network model. The reason we choose to use graphs is that the connected graph structure can compress such relation between stocks. We investigate different graph construction methods so that we can describe the stock relation in a comprehensive way. Although graph convolutional network(GCN) has already been proved effective in the prediction of stock movement, it only considers one single graph. Here, we build a combination model based on the GCN that the model can deal with multiple graph features. Apart from GCN, we also applied the transformer-based model to learn the correlation between the stocks. Transformer is a popular model for natural language processing and the implementation in stock prediction is focus on dealing with the public mood. In our research, we applied the stock graph as a mask to attention layer so that the transformer can have prior knowledge.

Our experiment applies the stock data from the New York stock exchange. We propose our model using graphs outperforms the recurrent neural network or other methods which do not take the graph structure into account. In the experiment, we investigate how various type of graphs influence the prediction result. The results show that the combination of multiple graphs effectively improves accuracy. But it does not outperform the general GCN model due to the quality of our constructed graphs. Furthermore, we introduced three graph construction methods and examined their impacts on stock prediction problem. The result indicates that the correlation graph is the optimal choice among them. Both multi-graph GCN and transformer with graph mask outperform the LSTM model. Besides, pure transformer+LSTM also produces a better result than the LSTM model. The result reveals our assumption that the internal relation provides sufficient improvements for the stock prediction problem.