LOOCV bias is lower than k-fold CV bias.

LOOCV bias is lower than k-fold CV bias.

• A. False
• B. True
• C. It depends
• D. Not defined

Answer: (B)

Cross-validation bias measures how far the CV estimate of predictive error is from the true error on new data. In leave-one-out CV, for each fold you train on nearly all the data (n−1 observations) and test on the single left-out point. That training set is almost the full dataset, so the resulting model is very close to what you’d get if you trained on all the data. The error you observe on that single hold-out point then tends to be a good reflection of the model’s true generalization performance, making the CV estimate less biased relative to the true error. In contrast, standard k-fold CV with a smaller number of folds uses much smaller training sets. The models trained in those folds are more restricted, and their test errors can deviate more from the true generalization error, introducing greater bias in the CV estimate. The trade-off is that LOOCV has higher variance because each fold is so similar to the others and relies on almost the same data, so its estimates can fluctuate more with the particular sample.

Cross-validation bias measures how far the CV estimate of predictive error is from the true error on new data. In leave-one-out CV, for each fold you train on nearly all the data (n−1 observations) and test on the single left-out point. That training set is almost the full dataset, so the resulting model is very close to what you’d get if you trained on all the data. The error you observe on that single hold-out point then tends to be a good reflection of the model’s true generalization performance, making the CV estimate less biased relative to the true error.

In contrast, standard k-fold CV with a smaller number of folds uses much smaller training sets. The models trained in those folds are more restricted, and their test errors can deviate more from the true generalization error, introducing greater bias in the CV estimate. The trade-off is that LOOCV has higher variance because each fold is so similar to the others and relies on almost the same data, so its estimates can fluctuate more with the particular sample.