Temporal similarity-constraint graph networks for stock prediction with stock relations

Stock prediction plays a crucial role in assisting investors with decision-making by leveraging predictive models to forecast future trends. Traditional stock prediction models primarily focus on price regression, treating stocks as independent entities and neglecting inter-stock relationships, which limits their applicability in investment decision-making. To address this limitation, we introduce temporal similarity-constraint graph networks (TSCGNs), a novel approach that integrates stock relations and ranks stocks based on expected returns. Experimental results on NASDAQ and NYSE datasets demonstrate that TSCGN significantly improves stock selection performance over baseline models. Specifically, TSCGN achieves a 34.1% and 29.4% improvement in cumulative investment return ratio (IRR) compared to rank-LSTM and TGC, respectively. Additionally, TSCGN yields a 4.8% increase in mean reciprocal rank (MRR) while maintaining a comparable mean squared error (MSE), demonstrating its effectiveness in stock selection. These findings highlight the importance of incorporating inter-stock relations and temporal similarity constraints in financial modeling to enhance investment decision-making.