The research demonstrates a simple procedure that can accurately estimate the generalisation error with only unlabelled data.