It is crucial that the test set is not used during the process to select the best algorithm, nor should it be used to train this final model. If these caveats are followed, then the test set can be used to estimate the generalization performance of this final model on unseen data.
