Although many studies have been conducted on behavioral profiling of homicide offenders, none of these have empirically shown that this technique has scientific validity and solid accuracy. Based on data from 839 Japanese homicide cases, we constructed multivariate logistic regression models to infer offenders' characteristics of sex, age, stranger offender, criminal record, single offender, member of organized crime, and residential area. We evaluated validity with a tenfold cross-validation procedure and a receiver operating characteristic (ROC) analysis. The results indicated that the area under the ROC curve ranged from .70 to .87, suggesting moderate and sufficient accuracy. Such a multivariate approach would be useful to reduce costs and minimize errors in the early stages of an investigation, especially for jurisdictions that possess large homicide databases.