Enhancing Stock Trend Prediction Models by Mining Relational Graphs of Stock Prices

• Published in: 2020 International Conference on Pervasive Artificial Intelligence (ICPAI) pp. 110 - 117
• Main Authors: Li, Hung-Yang, Tseng, Vincent S., Yu, Philip S.
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.12.2020
• Subjects: arbitrary time series prediction backbone model; artificial intelligence; Companies; company embedding; correlated stock; data mining; Data models; Dynamic time warping; external resources; finance; finance domains; generated relational graph; graph embedding model; graph theory; incorporate relations; knowledge graph; Knowledge graph embedding model; Lead; learning (artificial intelligence); mining Relational graphs; Predictive models; pricing; Relational Graph Stock Enhancing Network; relative performance gain; stock market; stock markets; stock price movement; Stock prices; Stock trend prediction; Stock trend prediction models; time series; Time series analysis
• DOI: 10.1109/ICPAI51961.2020.00028

Stock trend prediction has been a subject that attracts lots of attentions from a diverse range of fields recently. Despite the advance by the cooperation with artificial intelligence and finance domains, a large number of works are still limited to the use of technical indicators to capture the principles of a stock price movement while few consider both historical patterns and the relations of its correlated stock. In this work, we propose a novel framework named RGStocknet (Relational Graph Stock Enhancing Network) that can boost performance on an arbitrary time series prediction backbone model. Our approach automatically extracts the relational graph into which the graph embedding model can be easily integrated. Treated as an additional input feature, company embedding from the graph embedding model aims to improve performance without the need for external resources of the knowledge graph. The experiment results show that the three benchmark baseline can benefit from our proposed RGStocknet module in relative performance gain on the S&P500 dataset with 2.97%, 2.48%, and 7.03% on profit-score and with 25.50%, 17.53%, and 12.75% on accuracy respectively. Applied to a real-world trading simulation environment, our approach also outperformed the backbone model and doubled the average return on ResNet over the buy and hold (BH) strategy from 4.42% to 7.38%. Visualization of the generated relational graph and company embedding also shows that the proposed method can capture the hidden dynamics of other correlated stocks and learn representation across the whole stock market. Moreover, the proposed method was shown to carry the potential to incorporate relations with external resources to achieve higher performance further.