| In the interdisciplinary research of finance and deep learning,predicting the movement of financial asset price with intelligent computational models is a longstanding topic.Recently,finance researches have acknowledged the Momentum Spillover Effects,which indicates that the fluctuation of a firm’s stock price can also influence those of its related ones,and thus stocks prices among firms can have leadlag relation.With the progress made in Graph Neural Networks(GNN),few works have adopted the GNN to incorporate the influence from related firms into the predictions of stock price fluctuation,rather than modeling the prediction just as a function of the historical input information of one single firm.Cleary,these studies provide new perspectives and practical openings for modeling financial problems using emerging deep learning methods.However,by reviewing previous studies,this paper argues that previous studies still suffer from a few problems.Specifically,past studies failed to treat the modeling of momentum spillover effect as a spatial-temporal learning problem.Actually,the stock prices of an individual firm are influenced by their own historical risk information,and thus exhibit temporal dependency.At the same time,stock price is also influenced by risk spillovers of their related firms,and thus exhibit spatial dependency.More importantly,the spillover of risk from related firm to shape the price of target firm is a continuous process.On the one hand,the stock price would respond sluggishly to the historical risk of its related firms.On the other hand,changing market states would bring out new risk from related firms to spill.The stock price is spatially and temporally correlated with the historical risk information of both itself and its related firms.In line with the above idea,this paper proposes a spatial-temporal graph neural network,which is HST-GNN,to model the momentum spillover among listed firms as a spatial-temporal learning problem,and predict price trend of multiple stocks.HST-GNN is mainly composed of three distinctive sub-modules,namely the spatial learning module,hysteresis learning module,and the temporal learning module.The spatial learning module is responsible for calculating the potential risks brought out by the changing risk status of related companies;Next,the hysteresis learning module is responsible for modeling the continuous process of momentum spillover to influence the return of target firm;At last,the temporal learning module is responsible for memorizing the evolving pattern of historical risk.Three modules work alternatively and recurrently to learn the spatial and temporal dependency between the price fluctuation of one firm and historical information of both itself and its related firms.This paper conducted experiments on five-year real financial data of over 2700 stocks in NASDAQ and NYSE market to demonstrate the effectiveness of our proposed model,HST-GNN.Experimental results showed that the proposed HSTGNN have achieved the best predictive performance among both the baseline and state-of-the-art models in terms of three classification evaluation metrics,with enhancement of 17.54%,20.13%,19.17% in DA,AUC,F1 score compared to baseline model,and 5.29%,6.56%,3.49% compared to state-of-the-art models.In addition,the paper also compared the effectiveness of each sub-modules in HSTGNN,and the results shows that removal of any sub-modules can lead to a degradation of the overall predictive power.In addition,the visual analysis of interfirm correlation weights reveals that in some industries,individual firms only have strong interactions with a combination of related firms,while in others,individual firms tend to interact evenly with all related firms.At last,visualization of hysteresis influence of momentum spillover indicates that firms would respond to the risk spilled firm relevant firms to varying degree,and the hysteresis influence caused by it generally diminished over time. |
