Worsening temperature extremes are among the most severe impacts of human-induced climate change. These extremes are often defined as rare events that exceed a specific percentile threshold within the distribution of daily maximum temperature. The percentile-based approach is chosen to follow regional and seasonal temperature variations so that extremes can occur globally and in all seasons, and frequently uses a running seasonal window to increase the sample size for the threshold calculation. Here, we show that running seasonal windows as used in many studies in recent years introduce a time-, region-, and dataset-depended bias that can lead to a striking underestimation of the expected extreme frequency. We reveal that this bias arises from artificially mixing the mean seasonal cycle into the extreme threshold and propose a simple solution that essentially eliminates it. We then use the corrected extreme frequency as reference to show that the bias also leads to an overestimation of future heatwave changes by as much as 30% in some regions. Based on these results we stress that running seasonal windows should not be used without correction for estimating extremes and their impacts.
Worsening temperature extremes are among the most severe impacts of human-induced climate change. These extremes are often defined as rare events that exceed a specific percentile threshold within the distribution of daily maximum temperature. The percentile-based approach is chosen to follow regional and seasonal temperature variations so that extremes can occur globally and in all seasons, and frequently uses a running seasonal window to increase the sample size for the threshold calculation. Here, we show that running seasonal windows as used in many studies in recent years introduce a time-, region-, and dataset-depended bias that can lead to a striking underestimation of the expected extreme frequency. We reveal that this bias arises from artificially mixing the mean seasonal cycle into the extreme threshold and propose a simple solution that essentially eliminates it. We then use the corrected extreme frequency as reference to show that the bias also leads to an overestimation of future heatwave changes by as much as 30% in some regions. Based on these results we stress that running seasonal windows should not be used without correction for estimating extremes and their impacts.