One criticism that is often leveled against using resampling methods (such as cross-validation) to measure model performance is that there is no correlation between the CV results and the true error rate.
Let's look at this with some simulated data. While this assertion is often correct, there are a few reasons why you shouldn't care.
The Setup
First, I simulated some 2-class data using this simulation system. There are 15 predictors in the data set. Many nonlinear classification models can achieve an area under the ROC curve in the low 0.90's on these data. The training set contained 500 samples and a 125 sample test set was also simulated.
I used a radial basis function support vector machine to model the data with a single estimate of the kernel parameter sigma
and 10 values of the SVM cost parameter (on the log2 scale). The code for this set of simulations can be found here so that you can reproduce the results.
Models were fit for each of the 10 submodels and five repeats of 10-fold cross-validation were used to measure the areas under the ROC curve. The test set results were also calculated as well as a large sample test set that approximates the truth (and is labeled as such below). All the results were calculated for all of the 10 SVM submodels (over cost). This simulation was conducted 50 times. Here is one example of how the cost parameter relates to the area under the ROC curve: