This preregistered study investigates the predictive power of online hate speech for offline hate crimes against migrants and the LGBT community in Spain (2016-2018). Using social media data from X (Twitter) and Facebook, and police records, computational models (VAR, GLMNet, XGBTree) were developed to forecast hate crimes. The best model for migrant crimes achieved an R² of 64%, while that for LGBT crimes reached 53%. Toxic language proved a stronger predictor than hatred or sentiment, with Facebook posts outperforming tweets. While not claiming causation, the study concludes that online inflammatory language could be a leading indicator for detecting potential hate crimes.

Carlos Arcila Calderón, Patricia Sánchez Holgado, Jesús Gómez, Marcos Barbosa, Haodong Qi, Alberto Matilla, Pilar Amado, Alejandro Guzmán, Daniel López-Matías, Tomás Fernández-Villazala