REACT: A Lightweight Reliability-Aware Framework for Spatio-Temporal Out-of-Distribution Prediction

In the real world, spatio-temporal forecasting systems inevitably encounter out-of-distribution (OOD) shifts, where temporal shifts evolve due to seasonal patterns or policy changes, and spatial dependencies shift with sensor deployment, removal, or topology reconfiguration. While many approaches attempt to address these challenges via invariant learning or causal adjustment, they implicitly assume that the contextual signals supporting prediction remain reliable. Through empirical studies and theoretical analysis, we find that this assumption often breaks because: (i) node-level context can become biased or incomplete under temporal and spatial OOD, yet existing methods lack mechanisms to assess such reliability; (ii) existing standard graph-based aggregation propagates information with fixed strength, potentially amplifying corrupted context across nodes. To this end, we propose REACT, a lightweight REliability-Aware Context-calibrated spatio-Temporal forecaster that explicitly models and controls context reliability under OOD. REACT decouples prediction from adaptation by first establishing a topology-independent predictive anchor via a graph-free linear encoder, ensuring transferable node-wise representations. It then constructs an availability-aware context that captures both observed signals and prior-based imputations, enabling explicit characterization of context completeness. Building on this, REACT introduces a reliability-guided calibration mechanism that dynamically modulates spatial support through uncertainty-aware gating, selectively leveraging trustworthy neighbors while suppressing unreliable propagation. Extensive experiments on ten real-world datasets with diverse temporal and structural shifts demonstrate that REACT consistently outperforms eleven baselines. These results highlight the importance of explicitly modeling and controlling context reliability as a fundamental principle for robust spatio-temporal forecasting under distribution shift. Our source code is available at https://anonymous.4open.science/r/REACT-A82B.