Machine learning for food security: Principles for transparency and usability

Machine learning (ML) holds potential to predict hunger crises before they occur. Yet, ML models embed crucial choices that affect their utility. We develop a prototype model to predict food insecurity across three countries in sub-Saharan Africa. Readily available data on prices, assets, and weather all influence our model predictions. Our model obtains 55%-84% accuracy, substantially outperforming both a logit and ML models using only time and location. We highlight key principles for transparency and demonstrate how modeling choices between recall and accuracy can be tailored to policy-maker needs. Our work provides a path for future modeling efforts in this area.