Progressive Dependency Representation Learning for Stock Ranking in Uncertain Risk Contrasting

The practice of ranking a list of stocks to facilitate investment decisions has garnered a lot of attention in the fintech field, aiming at minimizing investment risk while maximizing profitable returns. With recent developments in deep representation learning such as temporal/relational dependency, prior efforts either strive to explore the temporal dynamics behind distinct stocks or expect to expose the collaborative signals from predefined relations, resulting in promising achievements in stock ranking. However, owing to the profound or intricate fluctuations of stock markets, existing insights rarely consider the uncertain risks underlying the learning of dependency representation, which could bring a narrow perspective on how to perceive market laws and ultimately yield an unprofitable decision-making procedure. In this study, we introduce a novel Progressive Dependency representation learning solution with Uncertain risk contrasting (PDU), primarily seeking to progressively uncover multiple dependency dynamics from historical trading signals for stock ranking in addition to addressing the uncertain risks. Specifically, we devise a Progressive Dependency learning block (or PD) in PDU that can progressively capture the temporal and relational dependencies besides multi-term dependencies in the latent space, allowing a coupled exposure of diffusion impacts over historical trading. Furthermore, we introduce an uncertain risk contrasting mechanism in PDU by placing the PD block in a contrastive environment (i.e., certainty vs. uncertainty), aiming to stably enhance dependency learning in the latent space. The experimental results conducted on four real-world stock market datasets demonstrate the superiority of PDU over several baselines.