Quantitative palaeoclimatic reconstructions based on biological fossils are a major source of information on long-term climatic variability. Such reconstructions typically use some kind of a modern calibration data set describing the variation of the studied biological group in present-day climate space. Here, we explore the effect of calibration data set selection on palaeoclimatic reconstructions, by creating alternate calibration data sets via stratified random sampling to reconstruct mean July temperature (T_jul) for four fossil pollen sequences from northern Europe. We show that palaeoclimatic reconstructions using methods based on taxon-response models can be highly sensitive to the calibration data set used. In particular, the absolute reconstructed temperatures show great sensitivity to calibration data selection, which suggests that the absolute values of palaeoclimatic reconstructions may not be robust. By contrast, we find the relative shapes of the reconstructed curves to be more robust to calibration data selection because taxa tend to occupy similar relative locations along the sampled gradient regardless of calibration data set location. Based on this robustness of relative palaeoclimate curves, we suggest a debiasing procedure in which palaeoclimate values are estimated by fixing the relative curve with the modern observed value, thus correcting biases resulting from calibration data selection.