Subjects were 411 boys who had been followed from age 8 to 25 as part of the Cambridge longitudinal study of delinquent development. Variables were derived from data obtained from records, parents, teachers, peers, and the subjects at age 10. Five methods examined included the Burgess point score, the Blueck method, multiple linear regression, predictive attribute analysis, and logistic regression. Half the study population was used as a construction (C) and half as a validation (V) sample. Of the 25 variables examined, the best single predictor of official juvenile delinquency was troublesomeness in the C sample and daring in the V sample. The best predictor of self-reported delinquency was parental criminality in the C sample only. In the C sample, the best predictor of official adult delinquency was official juvenile delinquency; while in the V sample it was self-reported juvenile delinquency. For adult self-reported delinquency, the best predictor for the C sample was self-reported juvenile delinquency; in the V sample, it was official juvenile delinquency. Overall, it was difficult to identify a group with much more than a 50 percent chance of delinquency or to identify more than 50 percent of the delinquents. The more sophisticated techniques were, if anything, worse than the simpler Burgess and Glueck methods. Possible reasons for the relative inefficiency of delinquency prediction are discussed. Tabular data and 35 references are included.