Stock prediction aims to assess future price trends and assist investment decisions. With the recent success of graph convolutional networks (GCNs) in modeling relational data, they have also shown promise for stock prediction. However, vanilla GCNs lack the ability to capture long-range dependencies in graphs and have not fully utilized the structured knowledge with data available. In this paper, we propose a novel framework of Deep Augmented Relational Stock Prediction (AR-Stock). We first detect the long-range links using pre-trained knowledge graph embeddings, leading to a new geometrically augmented edge type into the provided stock market graph. We then construct the GCN model on this augmented graph, that predicts each company’s stock prices by leveraging its related corporations; specifically, to train the GCNs better over this complex graph, we introduce two novel self-supervised regularizers (graph partition and graph completion) to inform the model with the global and local topology features. Unifying the above ingredients, AR-Stock has the unique strength in capturing long-term and hidden graph node dependencies. Experiments on two popular stock market datasets, NASDAQ and NYSE, demonstrate the prediction superiority of AR-Stock. Particularly, in terms of the investment return ratio, AR-Stock improves 65.77% in NASDAQ, and 30.48% in NYSE, over state-of-the-art models, respectively.