Background Individuals attending emergency departments following self-harm have increased risks of future self-harm. Despite the common use of risk scales in self-harm assessment, there is growing evidence that combinations of risk factors do not accurately identify those at greatest risk of further self-harm and suicide. Aims To evaluate and compare predictive accuracy in prediction of repeat self-harm from clinician and patient ratings of risk, individual risk-scale items and a scale constructed with top-performing items. Method We conducted secondary analysis of data from a five-hospital multicentre prospective cohort study of participants referred to psychiatric liaison services following self-harm. We tested predictive utility of items from five risk scales: Manchester Self-Harm Rule, ReACT Self-Harm Rule, SAD PERSONS, Modified SAD PERSONS, Barratt Impulsiveness Scale and clinician and patient risk estimates. Area under the curve (AUC), sensitivity, specificity, predictive values and likelihood ratios were used to evaluate predictive accuracy, with sensitivity analyses using classification-tree regression. Results A total of 483 self-harm episodes were included, and 145 (30%) were followed by a repeat presentation within 6 months. AUC of individual items ranged from 0.43–0.65. Combining best performing items resulted in an AUC of 0.56. Some individual items outperformed the scale they originated from; no items were superior to clinician or patient risk estimations. Conclusions No individual or combination of items outperformed patients’ or clinicians’ ratings. This suggests there are limitations to combining risk factors to predict risk of self-harm repetition. Risk scales should have little role in the management of people who have self-harmed.