Dividing the data and the ROC

If one uses the entire dataset to build a model, it is possible that we might have over trained the model. A consequence is that the true performance of the model goes unnoticed for the unknown cases. Essentially, we need to build a good model for the credit problem and if the performance is unknown for the new or unforeseen cases, skepticism is bound to creep into our minds. A good practice then is to divide the available in three regions: (i) data for building the model, (ii) data to validate the model, and (iii) data to test the model. Thus, a set of models is built for a problem and then they are evaluated over the validated part of the data, and the model that does best at this stage is chosen for the test portion of the data. Data partitioning in three regions can be easily performed and we can quickly show how it is done on the German credit...