Abstract
By performing experiments on publicly available multi-class datasets we examine the effect of bootstrapping on the bias/variance behaviour of error-correcting output code ensembles. We present evidence to show that the general trend is for bootstrapping to reduce variance but to slightly increase bias error. This generally leads to an improvement in the lowest attainable ensemble error, however this is not always the case and bootstrapping appears to be most useful on datasets where the non-bootstrapped ensemble classifier is prone to overfitting.