Kudos to Samir for this idea.
Take the different models (mean week, random forests, extra trees) that the tool currently trains, and instead of using the best one (evaluated using R2 value using k-fold cross validation), use a combination of the three.