How Do You Get the Best Predictions?

We've drawn an imaginary line at 250 columns as the upper limit of columns that the model will attempt to figure out the relative patterns contributing between. That's arbitrarily set and on a go-forward basis, as compute capacity increases like it always does, that'll expand too. By the time you're dealing with 500 possible drivers of an outcome, almost certainly there's 10 that are meaningful. There are 100 that give very tiny contributions to your outcome and the remainder are strenuous information that's not particularly relevant.