For another error metric, the k-fold CV was also calculated, which only holds out m/k, for k = 10, of the training data each time for testing and trains on the remaining and takes the average of these k errors. As expected, logistic regression has lower error in the 9 and 19 features case, since it is generally asymptotically more e" cient and robust with larger data. We note that the k-fold CV error is higher than the hold-out CV error and it also decreases as the number of features increase. For more features than 20, the Hessian becomes ill-conditioned due to the dependency of the features and so that error is not reported.