Does machine learning generalize across political change? Predicting approval of political violence four years later

This study examined whether approval of political violence could be predicted from individual characteristics and whether models developed in one political context would still work four years later. Logistic regression and XGBoost models using 18 demographic, political, and psychological predictors were developed in a Finnish adult sample collected in 2022 and applied without refitting to an independent 2026 sample. Cross-validated ROC-AUC was .676–.677 in 2022. Despite a decline in violence approval from 12.6% to 7.9%, discrimination remained similar or better in 2026 (ROC-AUC = .751 for logistic regression and .707 for XGBoost). The political distribution of violence approval, however, changed markedly. Midpoint-or-higher responses remained nearly unchanged in the left-party bloc (17.4% to 16.8%) but fell from 18.0% to 3.8% in the right-party bloc. Thus, models trained in 2022 continued to predict violence approval in 2026 even though its relationship with political orientation had changed substantially.