Background: Risk scales are used widely in the management of patients presenting to hospital following self-harm. However, there is evidence that their diagnostic accuracy in predicting repeat self-harm is limited. Their predictive accuracy in population settings, and in identifying those at highest risk of suicide is not known. Method: We compared the predictive accuracy of the Manchester Self-Harm Rule (MSHR), ReACT Self-Harm Rule (ReACT), SAD PERSONS Scale (SPS) and Modified SAD PERSONS Scale (MSPS) in an unselected sample of patients attending hospital following self-harm. Data on 4000 episodes of self-harm presenting to Emergency Departments (ED) between 2010 and 2012 were obtained from four established monitoring systems in England. Episodes were assigned a risk category for each scale and followed up for 6 months. Results: The episode-based repeat rate was 28% (1133/4000) and the incidence of suicide was 0.5% (18/3962). The MSHR and ReACT performed with high sensitivity (98% and 94% respectively) and low specificity (15% and 23%). The SPS and the MSPS performed with relatively low sensitivity (24-29% and 9-12% respectively) and high specificity (76-77% and 90%). The area under the curve was 71% for both MSHR and ReACT, 51% for SPS and 49% for MSPS. Differences in predictive accuracy by subgroup were small. The scales were less accurate at predicting suicide than repeat self-harm. Conclusions: The scales failed to accurately predict repeat self-harm and suicide. The findings support existing clinical guidance not to use risk classification scales alone to determine treatment or predict future risk.