AI for social risk forecasting and explanation: The power of machine learning–based social risk models

This discussion paper presents three proof-of-concept models that use large, multimodal datasets including satellite imagery, economic indicators, and natural language processing of news and social media to predict changes in conflict, migration, and crime. The report demonstrates how these models capture dynamic interactions between climate stressors, economic conditions, and social perceptions, strengthening risk identification and early warning capacities.

The paper concludes that the optimal predictive and explanatory utility of machine learning and AI methods requires certain conditions, including the availability or producibility of data representing change in frequently occurring phenomena. In eastern Democratic Republic of Congo, models achieved validation accuracies of up to 76 per cent in forecasting violence, while in the Horn of Africa, population movements were predicted with over 70 per cent accuracy. By enabling anticipatory action and more precise targeting of interventions, these approaches contribute to disaster risk reduction through improved risk monitoring, early warning, and risk-informed decision-making in fragile and climate-affected contexts.