In this paper, we introduce several approaches for maintaining weights over the aggregate skill ratings of subgroups of teams during the skill assessment process and extend our earlier work in this area to include game-specific performance measures as features alongside aggregate skill ratings as part of the online prediction task. We find that the inclusion of these game-specific measures do not improve prediction accuracy in the general case, but do when competing teams are considered evenly matched. As such, we develop a "mixed" classification method called TeamSkill-EVMixed which selects a classifier based on a threshold determined by the prior probability of one team defeating another. This mixed classification method outperforms all previous approaches in most evaluation settings and particularly so in tournament environments. We also find that TeamSkill-EVMixed's ability to perform well in close games is especially useful early on in the rating process where little game history is available.