The paper proposes Relational Stock Ranking (RSR), a deep learning framework for stock prediction that directly optimizes for investment revenue by formulating the problem as a ranking task. Unlike traditional approaches that treat stocks independently and use regression or classification, RSR leverages stock relations (e.g., same sector, supplier-customer) through a novel Temporal Graph Convolution (TGC) component. TGC propagates sequential embeddings from an LSTM network across a relation graph in a time-sensitive manner, allowing the model to capture dynamic influences between stocks. The framework is trained with a combined loss function that includes both pointwise regression and pairwise ranking loss to ensure accurate ranking scores. Experiments on NYSE and NASDAQ data show that RSR achieves average return ratios of 98% and 71%, respectively, significantly outperforming the state-of-the-art SFM method. The results demonstrate the effectiveness of incorporating relational information and ranking optimization for stock prediction.