Projections of future fire activity from statistical models are a powerful tool for anticipating 21st-century fire regimes. In previous work, we developed a set of statistical models that predict the likelihood of fires over 30-yr timescales in Alaskan boreal forest and tundra ecosystems. These models reveal that fire-climate relationships are strongly nonlinear, exhibiting distinct climatic thresholds to burning. Driving these models with future climate projections further highlights the potential for fire-regime shifts to occur as climatic thresholds are crossed as climate warms, with some tundra and forest-tundra regions projected to experience at least a fourfold increase in the probability of burning. These projections are also accompanied by significant sources of uncertainty, particularly related to the calibration of statistical models using data spanning the relatively short observational period (i.e., 1950-2009). The goal of this project was to evaluate the ability of our statistical models to predict outside the observational record, and thus identify key strengths and limitations when applying statistical models to predict fire activity under scenarios of 21st-century climate change. We compared statistical predictions with independent fire-history reconstructions spanning the late Holocene (i.e., 850-1850 common era [CE]), developed from sediment charcoal in lake-sediment records. Our statistical models were informed with downscaled Global Climate Model (GCM) data for 850-1850 CE, and predictions were compared to mean fire return intervals estimated for each of 29 published fire-history reconstructions spanning seven Alaskan ecoregions. Our model-paleodata comparisons for 850-1850 CE highlighted varying levels of prediction accuracy among Alaskan ecoregions, with variability strongly related to ecoregion proximity to a summer temperature threshold to burning: regions closer to this threshold exhibited significantly larger prediction errors. In conjunction with this spatially varying prediction accuracy, modifying a modern (i.e., 1950-2009) fire-climate relationship even slightly, resulted in significant changes in prediction accuracy over 850-1850 CE, suggesting future projections would be sensitive to uncertainties in GCMs and/or changes in fire-climate relationships. The sensitivity in future predictions to the presence of a threshold to burning implies that uncertainty will likely shift spatially across Alaska and temporally throughout the 21st century, as different regions approach and surpass climatic thresholds over the course of the century. Our findings provide key information for understanding how fire regimes may respond to climate change in Alaska, valuable to land and fire managers for anticipating and preparing for future ecosystem change. Specifically, we provide spatially explicit projections of potential future fire regimes, based on the climate suitability to burning. These projections highlight which regions are most vulnerable to climatically induced fire-regime shifts. Additionally, we link these projections with our current understanding of spatial and temporal patterns in threshold-driven uncertainty, highlighting regions where our models are most likely to have high uncertainty. By evaluating both spatial patterns in vulnerability and threshold-driven uncertainty, our results allow managers and policymakers to identify and prioritize resources in preparing for future landscape change.