A general method to define confidence limits for model predictions based on validations