Forecasting the future of American mass shootings with Bayesian agent-based model calibration.

Mass shootings represent a persistent and devastating public health crisis in the United States, yet their rarity makes them difficult to model using standard statistical approaches. This is especially true for agent-based models (ABMs), which offer a mechanistic framework for understanding human behavior, but have historically struggled to capture low-probability, high-impact events in a way that is both interpretable and statistically well-calibrated. Here, we develop a high-throughput ABM that models mass shootings as arising from the intersection of three factors: the emergence of a motivated potential perpetrator, access to firearms, and the presence of a suitable target population. Through density estimation of the predicted severity of events from our ABM model, we evaluate the log-likelihood of historical US mass shooting data, enabling Bayesian posterior inference over the ABM parameters. We then apply variational Bayesian Monte Carlo (VBMC) methods to fit ABMs to historical data, while also producing predictive distributions with principled uncertainty quantification. Calibrated models are directly compared through Bayesian metrics and yield estimates of latent societal parameters, such as the baseline rate at which potential perpetrators emerge, that are grounded in a plausible generative process. We find that Federal Firearms License density is the dominant predictor of risk while population density has a positive but quickly saturating, impact. Through Widely Applicable Information Criterion evaluation of models, we find that an ABM accounting for high-lethality weapons, but assuming randomized mass-shooting targets, outperforms both naive Poisson-Weibull baselines and plausible behavioral variants. Our results demonstrate that mechanistic ABMs calibrated by VBMC can match or exceed standard statistical baselines while supporting counterfactual policy simulation.