Time Series Enhanced Graph Convolutional Networks for Stock Excess Returns Forecasting

Stock excess return prediction has always been a challenging task due to the high volatility and complexity of the data. It is still an open issue to leverage the valuable information embedded in the data effectively. This paper proposes a new graph neural network, named Time Series Enhanced Graph Convolution Network (TSE-GCN), which integrates graph-structured spatial information into the time series data for predicting stock excess returns. The model used a number of risk factors in the market, size, book-to-market ratio, profitability, and investment patterns. Experimental results showed that TSE-GCN outperforms several state-of-the-art baselines with up to 13.6% improvement on average for different prediction metrics. Besides, the model tends to exhibit high prediction accuracy even with a smaller volume of data. Generally, this research contributes a new dimension to the investigation in predicting stock excess return through special attention to how Graph Neural Networks can be applied and utilized for stock market data analysis.