Why aren't significant variables automatically good predictors?