Learning to predict violence

Abstract

Research summary This study investigates whether deep learning models-specifically feedforward neural networks-can enhance the prediction of new violent criminal arrests among individuals released pretrial. Using data from a large southeastern county, we evaluate four neural network configurations varying in depth, regularization, and class imbalance adjustments. Results show that incorporating class weighting and threshold tuning notably improves recall of rare violent events, with the best performing model identifying 57% of individuals arrested for violent crimes while maintaining stable false positive (FP) rates. We estimate counterfactual risk estimates among detained individuals, revealing that although some violent events may be averted through detention, this comes at the cost of detaining large numbers of FPs.Policy implications Improving the identification of individuals at risk for violent crime during pretrial release is a critical public safety priority, yet prevailing tools offer limited recall for such rare events. This study shows that neural networks, when properly calibrated, can improve detection of high-risk individuals without increasing FP rates. However, algorithmic advances alone are insufficient; progress requires modernizing criminal justice IT infrastructures to enable interagency data linkages and real-time analytics. Most current data systems lack the contextual and behavioral features necessary to capture latent risk. Ultimately, system improvement efforts should move away from mechanisms like monetary bail and toward evidence-based supervision strategies, paired with investments in data modernization and cross-agency collaboration to strengthen the predictive foundation of pretrial assessments.