Resumo: Brazil’s food insecurity reversed sharply during the COVID-19 pandemic, underscoring the need for effective forecasting tools to protect vulnerable households. Using nationally representative data on food insecurity from Brazil’s PNADC (2023–2024), this paper compares logistic regression, machine learning algorithms, and LASSO regression to predict food-insecure households. Machine learning models outperform the traditional logit in prediction: XGBoost identifies about 76 per cent of food-insecure households at 71 per cent accuracy, with the best discrimination of any model (AUROC 0.810). Notably, LASSO regression performs comparably to the non-linear models, indicating that regularised feature selection captures most of the available predictive gains, while algorithmic complexity adds a further, smaller margin of discriminative power. Rainfall data, contrary to expectations, added no predictive value. These findings inform the potential design of a national system to support food security policy targeting.