A statistical procedure to assess how well a model, fit using a training dataset, predicts outcomes in a holdout dataset, drawn from the same population as the training dataset.