An earth system model of intermediate complexity (CLIMate and BiosphERe – CLIMBER-2) and a land surface model (JSBACH), which dynamically represent vegetation, are used to simulate natural fire dynamics through the last 8000 yr. Output variables of the fire model (burned area and fire carbon emissions) are used to compare model results with sediment-based charcoal reconstructions. Several approaches for processing model output are also tested. Charcoal data are reported in Z-scores with a base period of 8000–200 BP in order to exclude the strong anthropogenic forcing of fire during the last two centuries. The model–data comparison reveals a robust correspondence in fire activity for most regions considered, while for a few regions, such as Europe, simulated and observed fire histories show different trends. The difference between modelled and observed fire activity may be due to the absence of anthropogenic forcing (e.g. human ignitions and suppression) in the model simulations, and also due to limitations inherent to modelling fire dynamics. The use of spatial averaging (or Z-score processing) of model output did not change the directions of the trends. However, Z-score-transformed model output resulted in higher rank correlations with the charcoal Z-scores in most regions. Therefore, while both metrics are useful, processing model output as Z-scores is preferable to areal averaging when comparing model results to transformed charcoal records.