Evac-Cast: An Interpretable Machine-Learning Framework for Evacuation Forecasts Across Hurricanes and Wildfires

Bo Li, Ali Mostafavi

Abstract: Evacuation is critical for disaster safety, yet agencies lack timely, accurate, and transparent tools for evacuation prediction. This study introduces Evac-Cast, an interpretable machine learning framework that predicts tract-level evacuation rates using over 20 features derived from four dimensions: hazard intensity, community vulnerability, evacuation readiness, and built environment. Using an XGBoost model trained on multi-source, large-scale datasets for two hurricanes (Ian 2022, Milton 2024) and two wildfires (Kincade 2019, Palisades--Eaton 2025), Evac-Cast achieves mean absolute errors of 4.5% and 3.5% for hurricane and wildfire events, respectively. SHAP analysis reveals a consistent feature importance hierarchy across hazards, led by hazard intensity. Notably, the models perform well without explicit psychosocial variables, suggesting that macro-level proxies effectively encode behavioral signals traditionally captured through time-consuming surveys. This work offers a survey-free, high-resolution approach for predicting and understanding evacuation in hazard events, which could serve as a data-driven tool to support decision-making in emergency management.