We used archival data to examine the predictive validity of a prerelease violence risk assessment battery over 6 years at a forensic hospital (N = 230, 100% male, 63.0% African American, 34.3% Caucasian). Examining "real-world" forensic decision making is important for illuminating potential areas for improvement. The battery included the Historical-Clinical-Risk Management–20, Psychopathy Checklist–Revised, Schedule of Imagined Violence, and Novaco Anger Scale and Provocation Inventory. Three outcome "recidivism" variables included contact violence, contact and threatened violence, and any reason for hospital return. Results indicated measures of general violence risk and psychopathy were highly correlated but weakly associated with reports of imagined violence and a measure of anger. Measures of imagined violence and anger were correlated with one another. Unexpectedly, Receiver Operating Characteristic curve analyses revealed that none of the scales or subscales predicted recidivism better than chance. Multiple regression indicated the battery failed to account for recidivism outcomes. We conclude by discussing three possible explanations, including timing of assessments, controlled versus field studies, and recidivism base rates.