In response to concerns that some of the most methodologically rigorous predictive studies of criminal offender characteristics may yet be less generalizable and applicable than advertised or assumed, this research first tests how well seven regression analysis models (represented by 28 equations) predict characteristics across three conditions: familiar cases (used to create the regressions), less familiar cases (native to the sample used to create the regressions) and foreign cases (from a similar but novel sample). Here a linear trend shows overfitting of the models to their own sample: a drop-off in prediction accuracy relative to simple mean-based prediction as cases become more foreign (ηp 2 = .646). In response to hopes that subjective input from expert police investigators could be integrated into the models to correct for this overfitting bias, this research also tests an algorithm combining expert ratings with the regression equations. Here moderate and significant improvement in novel-case prediction is observed overall (p = .036, r = .44) and equations for all twelve expert participants are shown to improve prediction to varying degrees. These results suggest that current best methods would perform poorly in the field, but can be improved by expert insight.