Objective Periodic structured violence risk assessment is the principle method underlying treatment planning for mentally disordered offenders but little is known about how risk changes over time. We aimed to determine whether hospitalised patients underwent reliable clinical change in assessed risk. Method We used a pseudo-prospective longitudinal study design. Demographic, clinical and risk assessment data of adult inpatients (N = 480) who had been routinely assessed with the HCR-20 on two to four occasions over a mean period of 17 months (SD = 2) were collated. Linear mixed models regression was conducted to determine change over time on total, subscale, and individual item scores, and relative change between clinical and demographic groups. The Reliable Change Index was calculated to examine whether change was greater than that expected by measurement error; clinically significant change was defined as the extent to which HCR-20 scores reduced below previously reported scores for patients not requiring hospitalisation. Results HCR-20 total score (Estimate -0.42, 95% CI = -0.84, -0.01, p <.05; d =.20) and clinical score (Estimate = -0.42, 95% CI = -0.64, -0.20, p <.001; d =.36) reduced over assessments. Significant differences in change were evident between clinically and demographically defined groups. A maximum of 3% of individuals showed clinically significant reliable reductions in HCR-20 total scores. The scores of patients whose overall level of risk was judged to have decreased did not reduce between assessments. Conclusion Violence risk changes very little over the course of treatment although there is some variation between groups. Most change cannot be demonstrated to be reliable or clinically significant. Important clinical management decisions should not depend solely on evidence from changes in HCR-20 risk assessment.