On the biases of error estimators in prediction problems

We conduct a theoretical analysis of the bias of Efron's (1983) "0.632 estimator", and argue from those results that a more appropriate choice might be a "0.667 estimator". The differences in construction are largely unimportant in applications, and hardly affect the already extremely good performance of the estimator. Nevertheless, it is interesting to note that Efron's heuristic argument and our very different, more technical one produce alternative but close recommendations for the "optimal" weights.Bias Bootstrap Error rate Linear model Prediction problem