Abstract

Effective poverty reduction and food security policies depend on accurate identification of vulnerable populations; however, conventional survey-based targeting approaches are often limited by high costs, infrequent updates, and coarse spatial coverage. This study develops a policy-oriented machine learning framework that integrates household survey, geospatial, and administrative data to improve poverty and food insecurity measurement and targeting efficiency. The framework combines socioeconomic information from household surveys with spatial indicators derived from remote sensing and infrastructure data, alongside administrative records capturing service access and social protection participation. We apply supervised machine learning models in R, including regularised regression and ensemble tree-based methods, to predict household and small area poverty and food insecurity outcomes. Model performance is evaluated against survey-only benchmarks using predictive accuracy metrics as well as policy-relevant targeting indicators, such as inclusion and exclusion errors. The analysis demonstrates that integrating multisource data substantially improves identification of high-risk households and regions, particularly in areas with limited survey coverage.
Simulation of social assistance targeting scenarios shows that machine learning based predictions can significantly increase targeting efficiency relative to conventional methods, enabling better allocation of limited public resources. Feature importance analysis further provides interpretable insights into the structural drivers of poverty and food insecurity, highlighting the roles of spatial accessibility, environmental conditions, and administrative coverage in shaping vulnerability. By translating machine learning outputs into actionable policy metrics, this study bridges the gap between advanced data science and applied development policy. The proposed framework offers a scalable and cost-effective tool for governments and development agencies seeking to enhance evidence-based targeting, monitor welfare dynamics, and design more responsive poverty and food security interventions