Potential and limitations of machine learning modeling for forecasting Acute Food Insecurity

Acute Food Insecurity (AFI) remains a highly relevant and persistent challenge. Machine Learning (ML) presents promising solutions to improve predictions and early warning systems by integrating large and diverse datasets and considering multiple drivers of AFI. This review examines target variables and input features in existing ML modeling efforts, providing an assessment of current data availability, accessibility and fragmentation, and improving the understanding of possibilities and limitations of ML for end-users. For modelers, we recommend optimal input variables and outline the modeling workflow by comparing all approaches. We furthermore develop a quantitative comparison of the influence of drivers in studied models’ predictions. We advocate for an increased effort to investigate ML causality and improve usability of ML models. • Target variables for AFI ML modeling remain scattered and with uneven coverage across sources. • A set of optimal input feature sources for actionable models is proposed. • XGBoost model seems to bring the most robust and accurate prediction tradeoff. • Past food insecurity and agroclimatic-economic drivers demonstrate a large influence in the models. • End users need more usable models results to fully harness the increasing amount of information available.