Comment by amypetrik8

just to go off of this there is also stochastic random overfit retraining process (SRORP). Idea behind SRORP is to avoid overfitting. SRORP will take data points from -any- aspect of the past process with replacment and create usually 3-9 bootstrap models randomly. The median is then taken from all model weights to wipe out outliers. This SRORP polishing -if done carefully- is usually good for a 3-4% gain in all benchmarks