The accurate assessment of risk of sexual reoffense for juveniles has the potential to inform many actions, including the segregation of low risk from high risk offenders, allocation of limited resources, matching treatment assignments and other programming with risk, and implementation of various state and federal regulations (e.g., registration and community notification laws) (Epperson, Ralston, Fowers, DeWitt, & Gore, 2006). The Juvenile Sexual Offense Recidivism Risk Assessment Tool -- II (JSORRAT-II; Epperson et al., 2006) is the only known fully actuarially-derived risk assessment tool designed specifically for juveniles who have offended sexually (JSOs). However, it has yet to be validated. To test the predictive validity of the JSORRAT-II, the present study employed an archival file review of an exhaustive and representative sample of JSOs ages 11 through 17 from Utah (n = 566) who entered the juvenile justice system for a sexual offense, their index offense, in 1996 or 1997. Juvenile sexual recidivism data, defined as a new formal charge for a sexual offense prior to age 18, was obtained for all JSOs in order to determine the predictive validity of the JSORRAT-II. The results indicated that the tool can be scored with a high degree of reliability (ICC = .96) when research assistants received extensive training and oversight. Additionally, the JSORRAT-II predicted juvenile sexual recidivism at greater than chance levels (ROC = .64, 95% CI from .58 to .71), and this level of predictive accuracy was the same regardless of time at risk. Because this level of accuracy was substantially below the values reported on the original development sample, several additional analyses were carried out to attempt to find its cause. First, several coded variables were examined to determine their impact on the predictive validity. Of these, it appeared that the JSORRAT-II did not perform well for JSOs who had exclusively offended against siblings (ROC = .58; 95% CI from .43 to .73). However, when thes JSOs were removed from the predictive validity analysis, the ROC did not substantially improve (ROC = .66; 95% CI from .59 to .73). The amount of missing data was also examined vis a vis the predictive accuracy. The results of those anlayses indicated that missing data did not substantially impact the indices of predictive validity. Confirmatory factor analysis was used to test the four-factor solution found with the original development sample. The results of the original model did not appear to fit the data adequately; however, model fit improved when factors and several residuals were allowed to correlate. Some of these correlated residuals appeared logically or theoretically justified, while many others did not. Consequently, an exploratory principle-components analysis with Varimax rotation was employed to compare the original, development factor structure with a four-factor structure from the validation sample. Three items loaded highest on different components between the two samples; however, two of these items exhibited similar cross-loading patterns in the two samples. Only item 6 (Use of Deception and Grooming) did not follow the same pattern of loading. Finally, several possible explanations for the reduced level of accuracy were discussed, as well as the implications of this reduced accuracy for informing a variety of risk-related decisions.