Whether adjustments to differential prediction observed among sex, racial/ethnic, or age subgroups in a college-level freshman class could be used to improve grade prediction accuracy for these subgroups in future freshman classes was studied. This study is based on the American College Testing program test scores, high school grades, and college freshman grade averages of students from national samples of colleges. The sex data set included 105,502 students from 200 colleges; the race data set included 134,601 students from 256 colleges; and the age data set included 96,522 students from 216 colleges. For older students, dummy variable and separate subgroup prediction equations were found, on cross-validation, to be more accurate than the total group equations. For sex subgroups, dummy variable and separate subgroup equations were only moderately effective in improving prediction accuracy. For racial/ethnic subgroups, dummy variable and separate subgroup equations were more often than not less accurate, on cross-validation, than total group equations. Among all three kinds of demographic subgroupings, shifts over time in colleges' mean grades were found to be a much more important source of prediction bias than differential prediction. Moreover, prediction bias itself, from whatever source, was typically much smaller than error variance. Eighteen data tables and two graphs are provided. (Author/TJH)