Subject: Re: P123 ML for MI
RAM ..... I find this statement a bit strange

"I’m still in the experimental stages, experimenting with
Random Forest, Extra trees and SVM. Nothing has worked for me as well as I can do with traditional
screens in the SP500 arena with GTR1 or P123"

All traditional screening methods are nothing but intersection rulesets - RF ( Random Forest) is a ruleset ensemble. So its almost impossible for it not to reproduce any GTR1 based screener.

The KEY differences are that in ML based methods - there's inbuilt Train, Test, Validation samples to ensure your model is not OVERFITTING. While on Screens there's NONE. You are using the entire Lookback period to test the efficacy of your method.

In that case - I would PERSONALLY trust RF ( Its literally one of the most robust MLs out there , handling noise based data) more than your screen ie its pointing out a major Overfit risk.

Best