Black Swan Shootings: A Model for Predicting the Worst of the Worst Mass shootings



Journal Title

Journal ISSN

Volume Title



Since the late 1990s, mass shootings in the United States, and in particular public mass shootings with high victim counts, have increased in frequency at an alarming and largely unprecedented rate. Most academic studies on mass shootings are descriptive and offender-centric, offering few tangible opportunities for predicting future mass violence. This study takes a different approach. This study develops a county-level spatial threat assessment for identifying locations at high-risk for experiencing or producing ‘Black Swan shootings’, defined here as an attack involving a perpetrator(s) using firearms to kill or injure a significantly large number of innocent or unwitting people, chosen intentionally or at random. Counties are evaluated for their risk of experiencing these attacks or producing attack perpetrators based on community-level measures for social contagion, public safety, demographics, mental health and substance abuse, and weapons availability. Using data from 18 different sources on mass shootings, 44 events since 1998 are identified. Based on statistical anomaly detection for extreme casualty counts across all mass shootings since 2013, Black Swan shootings are events with either eight killed, 13 wounded, or 15 total casualties. The analysis of these events indicates clear contagion effects over space and time: attacks occur in 17 distinct clusters of counties and most of the time occur within one year of the prior attack, with heightened risk in the 35 days immediately following an attack. Using annual county-level data from 1998 to 2018, results from t tests, Cohen’s D effect sizes, and Mann-Whitney U tests indicate that communities which experience Black Swan Shootings or produce their perpetrators have statistically significant higher levels of violence, denser populations, more racial diversity, higher percentages of females, and higher counts of firearm laws compared to areas without such attacks. The spatial threat assessment uses a logistic regression model based on these most significant social factors. Retroactively, each year this model identifies less than 1% of the country as ‘high-risk.’ The model is deemed analytically viable when the metric score for the accuracy and precision of the count and location of Black Swan shootings in a given year is greater than 50%; this occurs in ten different years, all occurring since 2006. This includes nine specific bullseyes, where the model was an exact match for the county and year of either an attack or the offender residence of a Black Swan Shooting. The model’s performance is evaluated, with a discussion on the opportunities for operationalizing these findings to inform on future attacks.