Limitations

Our climate projections are produced with peer-reviewed methods and validated at the global scale, then provided as they are. For a local application, it is up to the user to check that they fit the purpose (Noël et al., 2022). Each step of the processing chain has limits worth knowing before using the data; the main evidence on the method comes from Vrac et al. (2025).

The method: what a perfect-model test shows

Bias adjustment corrects the statistics of a model, not its physics, and adds its own assumptions to the uncertainty of the projections (Ehret et al., 2012; Maraun, 2016). The main evidence on the limits of our method, CDF-t, comes from a “perfect model” test (Vrac et al., 2025). A regional climate simulation over Europe at 0.11° stood in for the observations, a coarse 2° copy of it stood in for the climate model, and CDF-t had to rebuild the fine-scale data up to 2100 under a high-emission scenario, where the right answer is known. The main messages:

• It keeps the climate change signal. Calibrated on 1981–2010, CDF-t rebuilt the temperature and precipitation of 2071–2100, their means, variability and extremes, with biases close to zero. Its precipitation changes can still differ from those of the raw model (wetter and more intense in our CMIP6 data; Noël et al., 2022).
• Links between variables and across space come from the climate model. CDF-t adjusts one variable and one grid cell at a time, so the correlation between temperature and precipitation and the spatial patterns stay as the model has them, with any bias they carry (François et al., 2020). For precipitation a clear spatial bias remains. Multivariate methods that also adjust these links exist (Vrac, 2018; Vrac and Thao, 2020) but still need testing in a changing climate.
• Extremes are kept in check, at a small cost. The variant used in our datasets adjusts the few most extreme values so they cannot become outliers; in exchange it slightly underestimates the day-to-day variability of precipitation.
• Settings matter. Describing the distributions too coarsely degrades the results, especially for precipitation.
• The test has its own limits. It used one regional model driven by one global model, over Europe, and a model is not the real world: repeating it with other models would make the conclusions more general.

The reference dataset

The projections are adjusted to the ERA5-Land reanalysis, so their quality is bounded by it. ERA5-Land is not of equal quality for every variable and region: it depends on how many observations it rests on. Compare it with local observations over the historical period before relying on projections built on it (Noël et al., 2022).

The climate models

• Five models. The projections use the five CMIP6 models selected by the ISIMIP project, chosen as representative of the whole ensemble: three with a low and two with a high climate sensitivity (Noël et al., 2022).
• An ensemble of opportunity. Even the full CMIP6 ensemble was not designed to sample every possible future, so values outside the range of the five models are possible (Tebaldi and Knutti, 2007).
• Known shortcomings. Climate models have improved but still have weaknesses, for example in extremes. A review of impact studies for your region is a useful first step (Noël et al., 2022).

In practice

• Use all five models and more than one scenario: see Which scenario should I use?
• Check the historical period against local observations where you have them.
• For decisions that rest on extremes at a site, combine the data with local studies, and ask us about the best approach.

References

• Vrac, M., Loukos, H., Noël, T., & Defrance, D. (2025). Should we use quantile-mapping-based methods in a climate change context? A “perfect model” experiment. Climate, 13(7), 137. https://doi.org/10.3390/cli13070137
• Noël, T., Loukos, H., Defrance, D., Vrac, M., & Levavasseur, G. (2022). Extending the global high-resolution downscaled projections dataset to include CMIP6 projections at increased resolution coherent with the ERA5-Land reanalysis. Data in Brief, 45, 108669. https://doi.org/10.1016/j.dib.2022.108669
• François, B., Vrac, M., Cannon, A. J., Robin, Y., & Allard, D. (2020). Multivariate bias corrections of climate simulations: which benefits for which losses? Earth System Dynamics, 11, 537–562. https://doi.org/10.5194/esd-11-537-2020
• Vrac, M., & Thao, S. (2020). R2D2 v2.0: accounting for temporal dependences in multivariate bias correction via analogue rank resampling. Geoscientific Model Development, 13, 5367–5387. https://doi.org/10.5194/gmd-13-5367-2020
• Vrac, M. (2018). Multivariate bias adjustment of high-dimensional climate simulations: the Rank Resampling for Distributions and Dependences (R2D2) bias correction. Hydrology and Earth System Sciences, 22, 3175–3196. https://doi.org/10.5194/hess-22-3175-2018
• Maraun, D. (2016). Bias correcting climate change simulations: a critical review. Current Climate Change Reports, 2, 211–220. https://doi.org/10.1007/s40641-016-0050-x
• Ehret, U., Zehe, E., Wulfmeyer, V., Warrach-Sagi, K., & Liebert, J. (2012). Should we apply bias correction to global and regional climate model data? Hydrology and Earth System Sciences, 16, 3391–3404. https://doi.org/10.5194/hess-16-3391-2012
• Tebaldi, C., & Knutti, R. (2007). The use of the multi-model ensemble in probabilistic climate projections. Philosophical Transactions of the Royal Society A, 365, 2053–2075. https://doi.org/10.1098/rsta.2007.2076