Hodos: A Traffic Forecasting Foundation Model with In-Context Spatial Modeling

Accurate traffic forecasting is essential for intelligent transportation. Since traffic conditions at one sensor are shaped by those nearby, spatial modeling lies at the core of the task. Prior methods usually bind spatial dependencies to predefined graphs, adaptive adjacency matrices, or node embeddings, which limits their transferability to new regions. Recent spatiotemporal foundation models (STFMs) pursue zero-shot transfer through pretraining on large-scale spatiotemporal data, yet their spatial modeling still hinges on target-domain spatial priors such as sensor coordinates or connectivity graphs and fails when these are missing or inaccessible. Cross-sensor self-attention needs none of these, but it compares sensors with shared projections and has no adaptive control over how much spatial information each sensor should receive. To fill this gap, we propose In-Context Spatial Modeling (ICSM), which models spatial interactions from the observed traffic context alone. It augments spatial self-attention with context-conditioned low-rank query–key modulation and a context gate that controls the strength of spatial update. Building upon this, we develop **Hodos**, a traffic forecasting foundation model, and construct **JaRTI**, a pretraining corpus of about 31 billion real traffic observations from 51 Japanese regions over eight years. Pretrained on JaRTI and evaluated zero-shot on six widely used US traffic benchmarks, Hodos reduces MAE and RMSE by 26.9% and 26.2% without spatial priors on average over the strongest baselines. Visualizations reveal transferable cross-region spatial patterns and show how Hodos responds as congestion emerges and propagates. Code and a subset of the dataset are available at https://anonymous.4open.science/r/Hodos-C85E.