Low levels of reliability were found among individual judges' forecasts, but high agreement was found for their composite judgements. These professionals' forecasts of inmates' violence showed low accuracy. Lens model analyses suggested that although forecasts and actual violence were both linearly predictable from the independent variables, low accuracy may have resulted because professionals failed to weight these factors optimally. Further analyses suggested that inmates showing violence at follow-up were younger, were more likely to come from nonurban areas, and had a history of more prior arrests and convictions than did inmates who did not show violence in prison. Implications for improving accuracy at predicting short-term violence are discussed. 3 tables, appendix, 5 notes, and 28 references (Author abstract)