Watch video to learn more about AI Powered Refugee Forecasting
Refugee crises are often seen as unpredictable emergencies. In a context of acute suffering, assistance is often rushed to those who have just fled conflict and violence and delivered within the communities that receive them. But what if refugee movements could be forecasted? What if hosting countries and their partners had the time to prepare for large inflows of people? With more than 122 million forcibly displaced people worldwide — double that of ten years earlier — these questions are pressing.
As both humanitarian and development financing are under pressure, the use of forecasting tools can provide major gains in the effectiveness of development responses. This is why the World Bank has been looking at artificial intelligence (AI) models to help countries be better prepared when a crisis hits. This is part of a broader effort to work together with UNHCR on promoting responses that integrate refugees into national systems. This approach provides improved social services and fosters better economic outcomes for both refugees and host communities.
Uganda’s progressive policies enables refugees access to employment, land, and public services alongside host communities. The World Bank’s Social Development and Prosperity teams have been piloting AI-based approaches to further strengthen preparedness under the Development Response to Displacement Impacts Project (DRDIP). DRDIP is a community driven development project investing in social service infrastructure and livelihoods in refugee-hosting districts. The project includes a contingency fund called the Displacement Crisis Response Mechanism (DCRM). Based on defined thresholds, the DCRM disbursed rapid funding to refugee hosting districts facing public service pressure following large refugee inflows.
The DCRM provides an agile mechanism to support infrastructure investments in refugee and host communities. But funds were disbursed after large inflows occurred and after host communities felt the impacts. So, at the request of the Government of Uganda, the World Bank has explored the potential of AI to predict refugee inflows. This would allow contingency financing to be triggered before large refugee inflows take place, enabling host districts to invest in public service capacity and minimize the impact on host communities.
AI and Machine Learning to predict refugee inflows
The World Bank has developed an AI-driven model to forecast refugee flows into Uganda from the Democratic Republic of the Congo (DRC) and South Sudan. The model draws on hard data, specific to the country context spanning the economic, social, natural and built worlds. Data is collected on conflict, climate, vegetation, built structures, economics, as well as online language about them. This is based on the logic that human behavior is not only caused by concrete measurable change, but also by perceptions of change (Figure 1).
Figure 1: Machine Learning Model ingesting dimensions of reality and perception of it
The model predicted change in daily refugee arrival data by ingesting over 90 independent variables into a machine learning model. The model was tested on unseen data and forecast change in future volumes of refugee inflows from DRC and South Sudan with over 80 percent accuracy (see Figures 2 and 3).
Key Drivers of Refugee Movements
The model goes beyond predictions alone to identify which factors have the most significant relationships with changes in refugee flows. Key factors include armed conflict, economic activity, climate, food prices, and the volume and sentiment of language about these and other things (see Figure 4). Understanding the factors associated with change can help governments, development and humanitarian actors anticipate refugee movements and better understand their drivers. In the areas from which refugees migrate, this evidence can support more effective approaches to addressing the drivers of forced displacement. The contextual knowledge this model provides can also inform World Bank country strategies, project design, project implementation, and policy dialogue.
Figure 4: DRC Model-identified Factors
Transforming DCRM Implementation
By integrating the predictive model, the DCRM can initiate public service scale-up approximately 4-to-6 months prior to refugees’ arrival and public service impact, enabling a visible response (e.g., construction of schools, health facilities, water points) to get underway early. This proactive approach focuses on the World Bank’s comparative advantage to provide necessary social infrastructure in a timely manner, while also allowing for integration of refugees into national services and local planning processes.
The model strengthens Uganda’s commitment to refugees by enabling more efficient allocation of limited resources, reduced tension between communities, and improved operational planning.
Figure 5: DCRM with prediction-based capability
Related Content
- Forcibly Displaced – Toward a Development Approach Supporting Refugees, the Internally Displaced and their Hosts
- Development Response to Displacement Impacts Project
- Tackling forced displacement as a development challenge
- Data-Driven Development Response to Displacement Crisis in Uganda: The Displacement Crisis Response Mechanism
- According to plan: Second activation of Uganda’s Displacement Crisis Response Mechanism (DCRM)
- Watch: AI Predictive Model For Refugee Modeling Presentation in Kampala
Join the Conversation