AI for Social Risk Forecasting and Explanation

What Is AI-Powered Social Risk Modeling?

Social risks - conflict, crime, population displacement, and poverty - compound in ways that impede drivers of poverty reduction and shared prosperity, such as access to better jobs and social protection.

The World Bank Group's Social Policy practice has developed AI and machine learning models that forecast and explain change in social risks so development more effectively target drivers. These models analyze thousands of variables - from climate, crime, conflict, demographic shifts, micro and macro-economic indicators, change in the bult environment, and online language about these factors.

Why Does Social Risk Prediction & Explanation Matter?

Social protection and social development are critical tools for moving people inclusively out of poverty and into jobs. Sustainable job creation, including reaching 500 million people - including 250 million women and girls - with social protection and employment support by 2030, can be better targeted by measuring, anticipating, analysing and getting ahead of social risks that trap people in poverty and cut them off from opportunity.

Crime raises the cost of doing business. Conflict disrupts infrastructure and markets. Displacement strains public services and separates vulnerable populations from schools, health facilities, and jobs. By the time traditional surveys capture these dynamics, the window for early action has often closed.

These models shift social protection delivery from reactive to anticipatory - enabling governments and development partners to target investments in the people, places, and risk mitigants of greatest impact to accelerate the movement of people out of poverty and into jobs.

How is the World Bank Using Machine Learning to Forecast Social Risk?

How the Models Work

The World Bank is exploring how machine learning can complement and improve forecasting and explanation of social phenomena. A multidisciplinary team - combining data scientists, economists, climate scientists, political scientists, linguists, and development specialists - identifies the factors most relevant to each risk, sources data representing them, and builds models that predict change and identifies which inputs are most associated with that change. To mitigate the risk that bias undermines findings, diverse context-specific expertise informs every stage of data selection, curation, and modeling.

Three Cases: How the Models Have Been Applied

The Bank has developed three proof-of-concept models, each addressing a first-order development issue:

Case 1: Conflict in Eastern Democratic Republic of Congo

The model predicts change in conflict levels in the North Kivu, South Kivu, and Ituri provinces, achieving up to 87% accuracy at a 150-day forecast horizon. It identified social perception - including sentiment about politics, armed groups, and mining - as the strongest driver of change, alongside economic pressures and conflict history.

The model is informing the World Bank's Risk and Resilience Assessment and the Stabilization and Recovery in Eastern DRC project, which envisages a risk financing mechanism that would trigger disbursements in response to observed or forecast conflict levels.

Case 2: Population Change in the Horn of Africa

In the data-scarce borderlands where Ethiopia, Kenya, and Somalia meet, the team used satellite imagery to track built structures across 56 towns and cities as a proxy for population change - achieving over 99.9% accuracy in structure identification. The model forecast population change up to three months ahead with up to 74% accuracy, identifying conflict events, climate pressures, and economic conditions as the principal drivers.

The findings are informing the $330 million DRIVE project on pastoral economies in the Horn of Africa, and the design of a displacement risk forecasting model for Ethiopia.

Case 3: Crime in a Small Island Developing State (SIDS)

Where official crime statistics were unavailable, the team used an agentic large language model to produce crime data from online news reporting. The resulting model predicted daily changes in crime with 87.1% accuracy, identifying food prices and sentiment about women, youth, public services, and governance as among the key drivers.

The model supported the World Bank Country Office in assessing governance indicators and has informed portfolio development.

How Can These Models Be Adapted and Applied?

The Social Policy and Disaster Risk Finance team delivered a model to support the world's first displacement risk financing mechanism enabling the Government of Uganda to scale public service capacity in advance of refugees' arrival rather than in response to it.

These models can be adapted to different risk types, geographies, and operational questions - from risk screening in project design, to social safeguard assessment, to targeting of social protection transfers. In the future, they could also be applied to predict and explain other development challenges, such as poverty levels, labor market shocks, or macro-fiscal pressures.