Not to disagree because you're totally correct, but the winning model used GBMs (a non-linear model) and a bunch of subtracted correlated features to great succ

Not to disagree because you're totally correct, but the winning model used GBMs (a non-linear model) and a bunch of subtracted correlated features to great success. When using GBMs, subtracting two features can add a lot of value. https://storage.googleapis.com/kaggle-forum-message-attachments/41007/1148/Josef_Feigl_Documentation.pdf https://storage.googleapis.com/kaggle-forum-message-attachme...