STAGformer: AI model slashes bike demand forecasting cost by 99%

STAGformer, developed by Ye Zihao and detailed in a July 2026 arXiv paper, tackles the challenge of station-level bike-sharing demand forecasting. The key innovation is a two-step agent attention mechanism that uses a small set of learnable spatial and temporal 'agent tokens' to first aggregate global information, then broadcast it back to individual stations and time steps. This reduces the computational complexity of standard self-attention from quadratic O(N²) to linear O(NT), making it feasible for large urban networks.

The model combines four modules: a spatio-temporal encoder that fuses dynamic node features with external factors (weather, time, points of interest), a graph propagation module for spatial neighbor aggregation, a temporal convolution module for local pattern extraction, and the agent attention module for global dependencies. On real-world datasets (NYC Citi-Bike and Chicago Divvy-Bike), STAGformer consistently outperformed state-of-the-art baselines across multiple prediction horizons, achieving significant improvements in RMSE and MAE. Ablation studies confirmed that the agent attention mechanism is critical for modeling global spatio-temporal dependencies.

This work has practical implications for city planners and bike-sharing operators, enabling more accurate demand forecasts at lower computational cost. The paper is available on arXiv (2607.06614) and provides a promising direction for efficient spatio-temporal modeling in large-scale urban systems.