Machine learning-enhanced modeling approach for optimally predicting household level food insecurity in Ethiopia during COVID-19

Food insecurity remains a critical global challenge, with low-income countries such as Ethiopia bearing a disproportionate burden. In settings where frequent data collection is limited, developing predictive models provides a cost-effective means of anticipating risks to enhance rapid life-saving action, supporting timely evidence-based interventions. This study develops predictive models, applying Machine Learning (ML) approaches, capable of accurately forecasting household-level food insecurity. This study used data from the Ethiopia-High Frequency Phone Survey, collected by World Bank. The performances of Machine Learning models and classical logistic regression model, in predicting households' food insecurity, were compared. Predictive models were trained and validated (internally and temporally) to evaluate model generalizability over time. ML models significantly outperformed the traditional logistic regression model in predicting household food insecurity. Based on the Brier score, the ML models demonstrated higher predictive accuracy and better calibration than the classical model. Regularization through optimized hyperparameters improved model stability and feature selection. While ridge, lasso, and elastic-net regressions produced similar coefficient directions, they differed in the number of selected predictors-the lasso model identified minimum key variables with comparable predictive accuracy. Temporal validation confirmed that the ML models maintained strong predictive performance, demonstrating their reusability and generalizability across time. The lasso regression model demonstrated strong predictive capability by selecting a manageable set of relevant features, resulting in reduced model complexity and improved interpretability. Integrating such ML models into food security monitoring systems can help policymakers design timely and data-driven interventions, enabling proactive responses in resource-constrained environment.