Why Bagging Hits a Wall More Trees Can't Break — and Randomness Is the Fix
Add enough trees to a bagged ensemble and error plateaus. The culprit isn't variance, which averaging already attacks, but correlation between tree predictions — a wall no number of trees can break. Feature subsampling is what targets it, and this post walks through the variance equation and an experiment that shows the ceiling in action. It reframes Random Forest's design from arbitrary hyperparameters into a single argument against correlated errors.