Predicting police and military violence: evidence from Colombia and Mexico using machine learning models

Armed forces violence has pervasive effects on public trust and population well-being. Such misconduct is not random, making its prevention both crucial and challenging due to the difficulty of measuring and detecting these phenomena beforehand. Recent advances in artificial intelligence offer new tools for this task. This article proposes the use of machine learning models to predict armed forces violence at the municipality level. Focusing on Colombia and Mexico—two countries with a significant number of human rights abuses by armed forces—the analysis draws on comprehensive subnational datasets. In Colombia, the study examines 1255 extrajudicial killing cases in which innocent civilians were misrepresented as guerrillas by the military. In Mexico, it considers 12,437 allegations of severe human rights abuses during militarized policing operations. Separate machine learning models are trained using four canonical algorithms—Lasso, Random Forests, Extreme Gradient Boosting, and Neural Networks—and their predictions are combined through a Super Learner ensemble. Results show high accuracy, specificity, and sensitivity in predicting police and military violence. In addition, feature-importance analysis highlights the most influential variables in the models’ predictions. These findings carry significant policy implications for contemporary law-and-order institutions, particularly in Latin America, where over a quarter of the world’s homicides occur, less than half the population expresses confidence in the police, and more than 9000 police killings are reported in a single year.