Spatio-Temporal Forecasting of Municipal Firearm Events in Colombia: Machine Learning versus Structured Bayesian Modeling

Abstract

This study evaluates predictive performance in municipal firearm event forecasting in Colombia using a rolling-origin evaluation framework (2011-2024). We compare three approaches: (A) a flexible machine learning model, (B) a Bayesian temporal negative binomial model with AR(1) structure, and (C) a Bayesian spatio-temporal model incorporating BYM2 spatial effects and temporal autoregression. Results show that while machine learning achieves the lowest RMSE overall, structured spatio-temporal modeling substantially improves performance relative to temporal-only models and provides interpretable spatial dependence estimates. The findings clarify when explicit spatial structure contributes to predictive accuracy in high dimensional municipal forecasting.