Recent traffic flow prediction methods are lacking abilities to determine predictive features. Thus, they will propagate the error in the next timestamps. In this paper, first, we assess the role of spatial and temporal features on the traffic speed prediction task. Secondly, we propose an attention-based architecture to effectively leverage both cues. Our model mainly consists of two major building blocks to capture the spatial-temporal features in the data and dynamically calculate the attentive features. More specifically, the first block sequentially applies temporal convolution to produce time-based features and then employs graph convolution to capture spatial features. The second component determines the attention between spatial and temporal features. The combination of the component’s output will be calculated to generate the final prediction results. Experiments on two real-world large-scale road network traffic datasets (i.e., METR-LA and PEMS-BAY) demonstrate that the proposed STATNet (spatial-temporal attention traffic network) model outperforms the state-of-the-art baselines such as graph-wavenet and STGCN (spatial-temporal graph convolutional networks).