Using Kernel Method to Include Firm Correlation for Stock Price Prediction

Abstract

In this work, we propose AGKN (attention-based graph learning kernel network), a novel framework to incorporate information of correlated firms of a target stock for its price prediction in an end-to-end way. We first construct a stock-axis attention module to extract dynamic and asymmetric spatial correlations through the kernel method and a graph learning module into which more accurate information can be integrated. An ensemble time-axis attention module is then applied to learn temporal correlations within each stock and market index. Finally, we utilize a transformer encoder to jointly attend to obtain information from different levels for correlations' aggregation and prediction. Experiments with data collected from the Chinese stock market show that AGKN outperforms state-of-the-art baseline methods, making up to 4.3% lower error than the best competitors. The ablation study shows that AGKN pays more attention to hidden correlation between stocks, which improves model's performance greatly.