Thanks for the PR.
I would just change the number of uncorrelated features to e.g. 20, because there is little need to use so many in this example.
However, I would keep the number of permutations to 1000 to have a reasonable p-value estimate.

@TomDLT Allright, agreed

Even better time improvement from 1:44 min (before) to 0:11 min (after)

Could you also check if using another model other than SVC could speed up the process?

@thomasjpfan re number of unrelated features, you happy with this?

@adrinjalali Yes, I also checked One-vs-One Logistic Regression, One-vs-All Logistic Regression and Multinomial Logistic Regression...all three were significantly slower than SVC.

I'd also check HistGradientBoosting*

@adrinjalali Just checked it, also way worse (12 sec for LogReg vs 10 min HGB)...so far Logistic Regression seems unreached by other models