Spatio-temporal sequence forecasting (STSF) aims to predict the future sequence of spatio-temporal inputs based on previous observations. Most of the existing methods focus on training graph neural networks and Transformers to extract spatial and temporal features of the input data. However, real-world spatio-temporal sequences often contain complex dependencies and lots of irrelevant information for forecasting, which might degrade the model’s performance. To tackle these issues, we propose a novel Spatio-Temporal Sequence Conditional Information Bottleneck (STSCIB) approach for STSF based on structural information theory. First, we establish an information bottleneck principle to extract the minimal and sufficient information for STSF, which could maximize the relevant information while minimizing the negative effects by eliminating irrelevant information based on the community dependencies among data. Second, we provide an efficient model to approximately implement the proposed STSCIB upon an encoding tree with the minimal structural entropy. Finally, we design an effective contrastive loss to train the model. Experiments on five datasets show the superiority of our method compared with other state-of-the-art methods.