margaretlubega/wfp-food-security

Machine learning pipeline forecasting district-level food insecurity across four African countries, in collaboration with the UN World Food Programme. # Predicting Food Insecurity Across Four African Countries > A machine learning pipeline that forecasts food-security risk at district level, built as a UC Berkeley MIDS capstone in collaboration with the UN World Food Programme's analytics team, and adopted into their humanitarian-response modeling. > ### ⚠️ Confidentiality note > This repository is a **descriptive overview only**. The project was developed with the World Food Programme's VAM (Vulnerability Analysis and Mapping) unit and delivered to WFP for operational use. The underlying datasets, trained models, prediction outputs, and final report are **confidential and are not included here**. No project data or source code is published. This page describes the problem, the engineering, and the methodology, none of which is WFP's to hold. ## The problem Food aid is finite, and getting it to the wrong place costs lives. The teams that decide where aid flows need to see where hunger is likely to worsen *before* it does. The hard part is that the warning signs are scattered across conflict events, food prices, rainfall, currency movements, and historical food-security assessments, all in different formats, on different schedules, across different countries. ## By the numbers | | | |---|---| | **Countries** | 4 (Cameroon, DR Congo, Mozambique, Nigeria) | | **Districts modeled** | ~1,150 admin-2 districts | | **Observations** | 60,000+ district-month records | | **Engineered features** | ~70, synthesized from 5 data sources | | **Data sources fused** | conflict, food prices, rainfall, currency, displacement, food-security assessments | | **Forecast horizons** | current, 3-month, and 8-month ahead | | **Models compared** | XGBoost, LightGBM, CatBoost, neural nets (RNN/CNN), linear baselines | | **Geospatial layers** | administrative boundaries from national to sub-county level (admin 0 to 3) | ## What we built An end-to-end pipeline that turns scattered public signals into district-level forecasts of fo …