Out-of-bag (OOB) error

As we already have resample the data several times, we don’t to do it again to estimate the test error if we have enough resamples \(B\).

The out-of-bag (OOB) error estimation is approximation to the leave-one-out cross validation by taking advantage of the \(1/3\) observations that weren’t used to train each base learner which opens the possibility to calculate \(B/3\) predictions, aggregate them and calculate the test error.

As your data sets become larger and your bagging iterations increase, it is common to use the OOB error estimate as a proxy for predictive performance.

The authors explained an example estimating the RMSE of 200 trees

Method	RMSE	Needed Time
10-fold cross validation	26,957	26 minutes
OOB error	26,462	0.96 minutes