Predicting Food Crisis

Early Warning Systems Under Strain

The ability to anticipate and prevent hunger crises is facing growing challenges due to disruptions in key data and early warning systems. Many traditional tools used by humanitarian agencies and governments to monitor food insecurity have experienced reduced capacity, limiting the availability of timely and reliable data. As a result, food crisis forecasting systems that previously provided early alerts are now under significant strain, and essential data streams are becoming increasingly scarce.

Artificial intelligence is emerging as a promising tool to help address this gap. An initial model for predicting food crises, developed by the Food Security Portal, utilizes passively collected big data and historical occurrences of food crises to estimate the prevalence of people experiencing acute food insecurity across each IPC phase. The model leverages high-frequency datasets and machine learning algorithms to assess and forecast food crisis risks across 60 countries.

Predictive Modeling Approach

The presentation below, led by Yanyan Liu and collaborators from IFPRI, UNICEF, and the University of Maryland, introduces a predictive modeling approach for anticipating food crises using passively collected big data. By leveraging over 100 real-time indicators—including climate conditions, economic stressors, infrastructure gaps, and conflict intensity—the model forecasts acute food insecurity (measured by IPC phases) up to 12 months in advance.

Trained on data from 29 countries between 2017 and 2021 and validated on 2022 IPC outcomes, this initial model achieves high sensitivity and reasonable precision, correctly identifying the majority of severe food insecurity cases. Key findings highlight the critical role of food price volatility and conflict in driving crisis-level outcomes. The study demonstrates the feasibility of using machine learning for early warning in fragile contexts, with significant implications for reducing assessment costs and improving the timeliness of humanitarian responses.

Expanded Framework & Global Alert Performance

An updated analysis, also led by Yanyan Liu and collaborators, scales the IPC acute food insecurity forecasting framework to 52 countries. This expanded iteration evaluates performance using 2025 as a strict holdout year, using temporal training based on the three preceding years with time-decay weighting to prioritize more recent observations.

The system nowcasts and forecasts crisis-level food insecurity (defined as IPC Phase 3 or worse) across multiple horizons, maintaining remarkably strong early-warning sensitivity even as the forecast window increases:

Key Insight: The stable F₂ scores (0.83–0.84) demonstrate that the system successfully prioritizes critical crisis detection even over extended timelines, while precision and overall accuracy actively improve at the 3- and 6-month horizons compared to the baseline nowcast.

Global Actual vs. Predicted Alerts

The actual-versus-predicted alert maps validate that predicted IPC Phase 3+ alerts remain highly spatially consistent with observed real-world crisis patterns across Africa, the Middle East, and targeted regions in Latin America.

Feature Drivers and Future Directions

• Primary Drivers: Global SHAP summaries indicate that weather and food price indicators act as the dominant feature groups, followed by agriculture, geography, and broader economic variables. Surprisingly, conflict factors contribute less when viewed in this aggregated global feature importance perspective.
• Operational Value: These findings demonstrate that the expanded model delivers operationally viable early-warning signals several months ahead of official IPC classification releases.
• Next Steps: Ongoing refinements will focus on improving spatial calibration, maximizing country-specific performance, and incorporating higher-frequency household and market indicators to optimize precision without sacrificing the framework’s high sensitivity.