I'd be most interested to hear if anyone has had worthwhile results using any particular feature importance metrics for determining feature inputs for machine learning models. I've tried most of the classification ones here: https://mlr-org.github.io/mlr-tutorial/release/html/filter_methods/index.html

...plus the usual suspects such as Boruta, MINE and stability selection

I've found none of them especially robust on a consistent basis (both using and not using feature scaling) as a means of predicting whether a particular feature set will be of use in predicting out of sample values. (For now I'm making do with the information ratio.) Of course given the nature of time series prediction that may just be the sad reality one has to live with. Or I may well be missing something... Many thanks