As discussed in Section 7.3.7, the choice of the right evaluation procedure for a given application depends on the accepted level of the bias vs. variance tradeoff, the size of the dataset, the regression algorithm, and the available computational resources. Most evidence suggests that the bias vs. variance tradeoff is best handled by the k-fold cross-validation procedure with k set to 10 or 20, particularly repeated several times. Whereas the .632 bootstrap procedure should also yield nearly unbiased and low-variance performance estimators for algorithms that are not prone to overfitting, in practice it is often hardly possible to completely eliminate the risk of overfitting a priori, which result in .632 bootstrap performance estimates being overly optimistic. The leave-one-out procedure should be rather avoided, except for small datasets where all other evaluation procedures would be considerably biased. The hold-out procedure is, conversely, best suited to very large datasets, for which other evaluation procedures would be too expensive and for which considerably smaller training samples would have to be used anyway due to computational constraints.