Nonparametric confidence bounds are obtained for a wide class of statistics using bootstrap. These results improve the errors in the probability estimates of the confidence intervals over the ones obtained by the normal approximation theory unconditionally.