Compress text files to make loading faster (for "reset and run all")
Add class weights to initial fits to help with imbalanced data? Still do thresholding, but class_weights good.
VERY LONG, and struggles around 4.
Struggles also around change in matrix dimensions going to 2007, usually related to failing to clean missings consistently in categorical vars. Some people "factorize" before cleaning, which maps 9s to arbitrary values they can't clean.
class_weights
good.