Assessing the utility of machine learning for predicting food sufficiency: a case study in Malawi

Abstract This study explores the potential of applying machine learning (ML) methods to identify and predict areas at risk of food insufficiency using a parsimonious set of publicly available data sources. We combine household survey data that captures monthly reported food insufficiency with remotely sensed measures of factors influencing crop production and maize price observations at the census enumeration area (EA) in Malawi. We consider three machine-learning models of different levels of complexity suitable for tabular data (TabNet, random forests, and LASSO) and classical logistic regression and examine their performance against the historical occurrence of food insufficiency. We find that the models achieve similar accuracy levels with differential performance in terms of precision and recall. The Shapley additive explanation decomposition applied to the models reveals that price information is the leading contributor to model fits. A possible explanation for the accuracy of simple predictors is the high spatiotemporal path dependency in our dataset, as the same areas of the country are repeatedly affected by food crises. Recurrent events suggest that immediate and longer-term responses to food crises, rather than predicting them, may be the bigger challenge, particularly in low-resource settings. Nonetheless, ML methods could be useful in filling important data gaps in food crises prediction, if followed by measures to strengthen food systems affected by climate change. Hence, we discuss the tradeoffs in training these models and their use by policymakers and practitioners.