A machine learning and agent-based modeling framework for tourism crisis management

Abstract This study proposes and tests a new Destination Resilience Assessment Framework (DRAF) to overcome the weaknesses of existing models which have a limited focus on multi-hazard prediction and on multi-layered interactions of stakeholders/infrastructure/markets during recovery in the context of cascading crises (pandemics, natural disasters and geopolitical shocks). The research employs a mixed-methods design that combines decision-tree based machine learning (random forests), an LSTM-based neural network, and ensemble approaches with agent-based modeling in NetLogo. Simulations are informed and calibrated using real-time streams from 847 sources globally (2019–2024), such as economic indicators, tourist flows, infrastructure capacity and stakeholder sentiment, combined with case data for 12 crises. DRAF achieves 94.7% accuracy in predicting recovery trajectories, identifies key resilience variables with 89.3% accuracy, and forecasts recovery time within ± 2.3 months. Simulation-based counterfactual analyses indicate that destinations implementing DRAF would recover an estimated 67% faster and regain visitor confidence an estimated 43% more quickly than the simulated baseline scenario.The five principal resilience factors are adaptive governance (β = 0.847), stakeholder collaboration networks (β = 0.723), infrastructure flexibility (β = 0.692), market diversification (β = 0.634), and community engagement (β = 0.587). Conceptually, the framework advances tourism crisis management models and, practically, provides transparent decision support that empowers stakeholders to assess impacts and coordinate evidence-based recovery actions.