The short instrumental record of about 100–150 yr forces us to use proxy indicators to study climate over long timescales. The climate information in these indirect data is embedded in considerable noise, and the past temperature reconstructions are therefore full of uncertainty, which blurs the understanding of the temperature evolution. To date, the characterization and quantification of uncertainty have not been a high priority in reconstruction procedures. Here we propose a new statistical methodology to explicitly account for three types of uncertainties in the reconstruction process. Via ensemble reconstruction, we directly obtain the distribution of decadal maximum as well as annual maximum. Our method is an integration of linear regression, bootstrapping and cross-validation techniques, and it (1) accounts for the effects of temporal correlation of temperature; (2) identifies the variability of the estimated statistical model and (3) adjusts the effects of potential overfitting. We apply our method to the Northern Hemisphere (NH) average temperature reconstruction. Our results indicate that the recent decadal temperature increase is rapidly overwhelming previous maxima, even with uncertainty taken into account, and the last decade is highly likely to be the warmest in the last millennium.