Stock trend prediction remains a significant challenge in quantitative investment due to nonlinear market dynamics, low signal-to-noise ratio, and evolving inter-stock dependencies. Traditional sequential methods treat stocks as independent time series, overlooking temporal and relational patterns across stocks. Recent graph-based methods attempt to simplify inter-stock relations as pairs but face three key limitations: 1) inability to distinguish and utilize positive/negative correlations, 2) neglect of lead-lag effects common in financial time series, and 3) failure to model temporal evolution of stock dependencies. To address these issues, we propose a Dual-Relational Dynamic Graph Representation Learning (DRDGRL) approach. Specifically, we introduce MT-DDTW, a novel multivariate time series similarity measure that captures both positive and negative relations under potential lead-lag conditions. These dual-relational information are then processed via a heterogeneous dynamic graph neural network to jointly encode neighbor information at different time points. The aggregated representations are finally used to forecast future stock trends. Without relying on static relations or hard-to-obtain textual data, our method achieves state-of-the-art portfolio optimization performance on multiple real-world stock datasets including CSI300, NASDAQ100, and S&P500.