Page 214 in the book: One-standard-error-rule

Here is how I understand it to work (for k fold CV): within_one_se.R (the script pulls the Credit dataset from the website of ISLR).

Basically, for best subset and CV k fold = 10, you estimate the test MSE for each k-fold 1 <= k <= 10, for each of the best predictor subsets.
Each set of MSEs allows you to calculate a mean MSE and standard error of the mean MSE for each subset of predictors.

I calculated the standard error by taking the standard deviation of the MSE for each predictor set and dividing it with sqrt(k), where k is the highest k fold.

Lemme know if you all think this solution is right!

~ Alex

I did something similar (played with the Hitters set instead). But instead of using Ariel's lines, I drew horizontal dashed red lines representing the oneSE area (the lower one falling in is the selected), which may be clearer !

abline(h=mean.cv.errors[min]+sd.cv.errors[min]/sqrt(k), col="red", lty=2)
abline(h=mean.cv.errors[min]-sd.cv.errors[min]/sqrt(k), col="red", lty=2)