Dynamic femicide prediction model based on reported variables in Peru

Abstract

This study analyzes the predictive dynamics of femicide in Peru using a departmental data panel (2015-2023) to strengthen early warning systems and optimize state response. Methodologically, the performance of an advanced machine learning model, Gradient Boosting Regressor (GBR), was contrasted against the traditional econometric ARIMAX(1,0,0) approach. A rigorous expanding window validation and the Diebold-Mariano test were implemented to evaluate the temporal robustness and statistical significance of the forecasts. Results validated the superiority of the GBR with a Mean Absolute Error (MAE) of 0.313, achieving a 45% reduction in predictive error compared to classical methods. Interpretability analysis using SHAP values identified that the historical inertia of the phenomenon constitutes the main predictor (62%); however, prior reports of sexual and psychological violence emerge as more critical early warning signals than physical violence. It is concluded that non-linear modeling detects specific windows of preventive opportunity that linear models ignore. These findings suggest an urgent reorientation of public policies toward the containment of sexual violence as a lethal precursor. By allowing for an evidence-based algorithmic resource allocation, these models can effectively interrupt the cycle of violence before a fatal outcome occurs, closing the gap between formal reports and effective victim protection.