Chronic food insecurity remains a challenge globally, exacerbated by climate change-driven shocks such as droughts and floods. Forecasting food insecurity levels and targeting vulnerable households is apriority for humanitarian programming to ensure timely delivery of assistance. In this study, we propose to harness a machine learning approach trained on high-frequency household survey data to infer the predictors of food insecurity and forecast household level outcomes in near real-time. Our empirical analyses leverage the Measurement Indicators for Resilience Analysis (MIRA) data collection protocol implemented by Catholic Relief Services (CRS) in southern Malawi, a series of sentinel sites collecting household data monthly. When focusing on predictors of community-level vulnerability, we show that a random forest model outperforms other algorithms and that location and self-reported welfare are the best predictors of food insecurity. We also show performance results across several neural networks and classical models for various data modeling scenarios to forecast food security. We pose that problem as binary classification via dichotomization of the food security score based on two different thresholds, which results in two different positive class to negative class ratios. Our best performing model has an F1 of 81% and an accuracy of 83% in predicting food security outcomes when the outcome is dichotomized based on threshold 16 and predictor features consist of historical food security score along with 20 variables selected by artificial intelligence explainability frameworks. These results showcase the value of combining high-frequency sentinel site data with machine learning algorithms to predict future food insecurity outcomes.
