Background Most structured assessment tools for assessing risk of violence were developed in Western countries, and evidence for their effectiveness is not well established in Asian countries. Aims Our aim was to examine the predictive accuracy of the Historical‐Clinical‐Risk Management‐20 (HCR‐20) for violence in forensic mental health inpatient units in Japan. Methods A retrospective record study was conducted with a complete 2008–2013 cohort of forensic psychiatric inpatients at the National Center Hospital of Neurology and Psychiatry, Tokyo. Forensic psychiatrists were trained in use of the HCR‐20 and asked to complete it as part of their admission assessment. The completed forms were then retained by the researchers and not used in clinical practice; for this, clinicians relied solely on national legally required guidelines. Violent outcomes were determined at 3 and 6 months after the assessment. Receiver operating characteristic analysis was used to calculate the predictive accuracy of the HCR‐20 for violence. Results Area under the curve analyses suggested that the HCR‐20 total score is a good predictor of violence in this cohort, with the clinical and risk sub‐scales showing good predictive accuracy, but the historical sub‐scale not doing so. Area under the curve figures were similar at 3 months and at 6 months. Conclusions Our results are consistent with studies previously conducted in Western countries. This suggests that the HCR‐20 is an effective tool for supporting risk of violence assessment in Japanese forensic psychiatric wards. Its widespread use in clinical practice could enhance safety and would certainly promote transparency in risk‐related decision‐making. Copyright © 2016 John Wiley & Sons, Ltd.