The study area was chosen to encompass the focus regions of the KlimaKonform project (Zorn et al., 2024) in East Central Germany (Fig. 1a). It spans 13 210 km² and hosts the upper and central part of the Weiße Elster river basin (Fig. 1b). The twelve enclosed catchments stretch over an area of 2960 km², with the smallest having an area of 107 km² (Dröda) and the largest of 529 km² (Strassberg).

The study area is embedded between the Leipzig lowlands to the North and the Elster Mountains and Ore Mountains to the South, resulting in a topographically varied landscape encompassing both flat plains and steeply sloped terrain (Rode et al., 2008; Miera et al., 2022). Altitudes range from 8 to 1252 m on a 30 m grid, from 85 to 1006 m on a 3 km grid, and from 101 to 852 m on a 12 km grid.

The northern part of the study area falls under the Köppen climate classification of Cfb (temperate, no dry season, warm summer), and the southern part is classified as Dfb (cold, no dry season, warm summer). In the study region, floods occur primarily in summer as a result of high-intensity convective storms, as well as in winter due to prolonged rainfall, sometimes amplified by rapid snowmelt.

The land use in the Weiße Elster river basin is predominantly cropland and pasture (Rode et al., 2008, see Fig. S1 in the Supplement). Chernozem/phaeozem and luvisol soils on loess are prevalent in the northern part of the study area, while cambisols and stagnic gleysols dominate the southern part (Miera et al., 2022).

Hourly observational data from 20 stations for mean air temperature, two stations for global radiation, 20 stations for relative humidity at 2 m, and 11 stations for wind speed at 10 m has been obtained from the German Weather Service (DWD) for the period from 2005 to 2014 (Fig. S2). No spatial interpolation has been performed; the station measurements are compared point-based to the corresponding climate model grid cells. For the assessment of precipitation anomalies, 60-year time series of monthly point-based precipitation data (1963–2022) have been obtained from DWD for five stations in the study area.

For other evaluations of precipitation simulations and hydrological modelling, adjusted radar-based quantitative precipitation estimations are used as reference for the study period from 2005 to 2014 in hourly time steps at a resolution of 1 km. The data set constitutes a gap-filled version of DWD's RADOLAN product (Weigl and Winterrath, 2010) and is provided through the ReKIS portal (Körner, 2022) for the federal states of Saxony, Saxony-Anhalt and Thuringia. RADOLAN is a radar-based precipitation dataset at 1 km resolution, which has been adjusted with measurements from DWD's ground station network (Winterrath et al., 2012). The choice of adjustment procedure is determined based on a performance test with a few control stations (Winterrath et al., 2012). It is unknown whether any, respectively which of these control stations correspond precisely to the stations used in this study. However, the station values can generally be assumed to be preserved.

The data has been upscaled and regridded to the climate model grids using the area-weighted average method (becoming RADOLAN3km and RADOLAN12km). Kreklow et al. (2020) evaluated the performance of the original RADOLAN dataset against rain gauge data and found RADOLAN to overestimate the number of days with daily precipitation sums >1 mm, as well as heavy rainfall days (i.e. daily precipitation sums >20 mm), in the summer half-year. Especially up to the year 2010, extreme high outliers in the time series were not uncommon, with markable improvements through the introduction of additional gauges in 2007, as well as further quality checks and new processing routines in 2010 (Kreklow et al., 2020). However, the average daily precipitation sum for the summer half-year was underestimated, with a similar trend reported for the winter half-year, yet far less pronounced (Kreklow et al., 2020). The mean median of the absolute daily deviations, when comparing RADOLAN data to daily point-based measurements of precipitation for the period from 2013 to 2016, was quantified as 0.761 mm d⁻¹ (DWD, 2021).

The hydrological model was calibrated and validated using Thiessen-interpolated point-based meteorological observations, including hourly precipitation from 20 precipitation gauges in the catchment (see Fig. S2 for a map of the observational network). An evaluation of how the RADOLAN product compares to the Thiessen-interpolated point-based precipitation measurements over the study area is outlined in Fig. S3 by means of a Q–Q plot.

Discharge data along the Weiße Elster has been obtained through the ReKIS project (Kronenberg et al., 2021), based on measurements by the Sächsisches Landesamt für Umwelt, Landwirtschaft und Geologie (LfULG), Thüringer Landesamt für Umwelt, Bergbau und Naturschutz (TLUBN) and Landesamt für Umweltschutz Sachsen-Anhalt (LAU).

Topographic information was obtained from SRTM 1 arcsec global (USGS, 2018), land use is based on Corine Land Cover 2018 (Copernicus Land Monitoring Service, 2020) and used soil data refers to BÜK200 (BGR and SGD, 2018).

The regional climate model ICON-CLM (version 2.6.4, Pham et al., 2021) has been run over the EURO-CORDEX domain at 12 km resolution for hourly output (ICON12km). It is driven three-hourly with data from the global atmospheric reanalysis ECMWF-ERA5 and provides the lateral boundary conditions for the CPRCM. As CPRCM, ICON-CLM (version 2.6.4) is employed in its convection-permitting setup at 3 km resolution over the Central European (CEU) domain providing hourly weather data (ICON3km). The domains are illustrated in Fig. S4. Information on the model dynamics is provided by Zängl et al. (2015).

In ICON12km, deep, mid-level, and shallow convection have been parametrised using the Tietke/Bechtold scheme (Tiedtke, 1989; Bechtold et al., 2008). ICON3km, in turn, operates on a resolution fine enough to explicitly resolve deep and mid-level convection, with only the shallow convection parametrisation needing to remain switched on. Vertical turbulent transfer is parametrised according to Raschendorfer (2008), and cloud microphysics are calculated with the single-moment scheme of Seifert (2008).

Tracer transport and fast-physics parametrisations are computed at time steps of 25 s, while the dynamical core runs at 5 s.

Air temperature, global radiation, relative humidity, and wind speed station measurements within the catchment bounding box (see Fig. S2) are compared to their co-located climate model grid cells. Air temperature estimates by the climate models are projected to the elevation of the measurement stations using the environmental lapse rate, which is of 0.65 °C/100 m. Precipitation simulations for the area are evaluated raster-based against radar precipitation estimations upscaled and regridded to the climate models' grids. Comparing precipitation simulations to observations on the same grid circumvents the need to consider differences in the degree of smoothing of local heavy rainfall fields with resolution (Iles et al., 2020). While the radar precipitation estimations are upscaled to the resolution of the climate models for comparison between the climate model simulations and the observations, the climate models themselves are not compared among each other by scaling the simulations to a common grid spacing. In fact, upscaling the finer climate simulation results to the grid of the coarser climate model would entail partial information loss (Iles et al., 2020), while downscaling the coarser climate model simulations would bring them into resolutions where processes dominate which are on a scale too fine to have been computed by the coarser model (Prein et al., 2016). It is however worth noting that first studies suggest that improvements at higher resolution may still be apparent after upscaling (e.g. Prein et al., 2016; Fantini et al., 2018 for upscaling from ∼ 12 to ∼ 50 km resolution).

When temporally aggregating observational data, time periods are not summarised if more than 20 % of their data is missing (Haylock et al., 2008). Smaller gaps are filled by dividing the totals through the proportion of non-missing observations (Haylock et al., 2008). Frequency analyses have been conducted based on frequency polygons with a number of bins according to Sturges' rule (Sturges, 1926). It should be noted that for large datasets Sturges' rule is prone to oversmoothing (Scott, 2009).

The meteorological time series are of only 10 years and thereby only a third of the length recommended for climatological analyses by the World Meteorological Organization (WMO, 2017). The short time series can only hint to strengths and limitations of the climate models, but longer time series are needed to draw firm conclusions on the hydroclimatic model performance.

Depth-duration-frequency curves have been constructed by sliding a moving window of a fixed duration (2, 4 to 24 h by 2 h-increments for 12 km-resolution data, and by 4 h- increments for 3 km-resolution data) over the time series and fitting the identified annual-maximum precipitation sums with a Gumbel distribution (Koutsoyiannis et al., 1998). The short duration of the time series of only 10 years poses a strong limitation to this approach.

The distributed, physically based deterministic hydrological model WaSiM (Message Passing Interface, model version: Richards 10.00.03; see Schulla, 2021) has been run for hourly time steps for the Weiße Elster river for the study period from 2005 to 2016. The first simulation year has been discarded to ensure adequate spin-up. Meteorological input variables consist of hourly air temperature, global radiation, relative humidity, wind speed and precipitation. The modules in WaSiM for high-resolution hydrological simulation in the temperate zone are for processes of evapotranspiration, snow accumulation and snow melt, interception of snow and precipitation, as well as an unsaturated-zone model and a groundwater model. Potential evapotranspiration is calculated after Penman-Monteith (Monteith, 1975; Brutsaert, 1982), thereby showing dependency on the meteorological input variables of air temperature, global radiation, relative humidity and wind speed. Interception processes are described by an energy balance approach, considering the complete suite of meteorological input variables, just as is the case for the simulation of snow accumulation and snow melt. The unsaturated zone model is built on the Richards-equation (Richards, 1931) and draws on the aforementioned modules, as well as on the groundwater model. The hydrological model was calibrated focussing on (1) the single reservoir recession constant for surface runoff, (2) the single reservoir recession constant for interflow, (3) the drainage density for interflow and (4) the fraction of surface runoff on snow melt.

Topography, land use/land cover (LULC) and soil characteristics have been embedded on a 1 km² grid for all simulation runs. As interpolation technique for the meteorological input data, Thiessen polygons were chosen in order to keep the grid structure of the climate model data unchanged. Radiation and temperature were adjusted considering the impact of topography, according to the implemented scheme by Oke (1987).

Alongside discharge, WaSiM outputs a range of other hydroclimatic variables. Among them are catchment-average total evapotranspiration and catchment-average relative soil moisture simulated for the upper soil section reaching from the soil surface to the root depth. Evapotranspiration and soil moistures estimates are studied in this paper. While snowmelt simulation data is outputted as well, it is not analysed further. CPRCMs are not expected to offer their main added value in winter, when large synoptic weather systems dominate, as these are already well represented by the driving RCMs (Strandberg and Lind, 2021). The focus of the paper presented here is therefore on summer storms, where CPRCMs and RCMs are likely to show their greatest differences (Strandberg and Lind, 2021) as the CPRCMs no longer rely on the parametrisation of deep convection (Ban et al., 2021).

Through the complex interactions of the hydroclimatic variables, the flood generation is likely to show a non-linear behaviour. In a scatter plot of rainfall to runoff volume, the presence of a breakpoint in the linear regression reflects such non-linear behaviour (Buttle et al., 2019). Ross et al. (2021) found the existence of a breakpoint to be most common when considering the relation of runoff volume (Qtot) to the sum of rainfall volume (Ttot) and antecedent rainfall (AT). This relation is therefore used in the study presented here to analyse the effect the use of a CPRCM has on the non-linearity of flood generation compared to the use of a coarser regional climate model with parametrised convection. The temporal delineation of runoff events over a specified threshold of 0.03 mm h⁻¹ at the most downstream catchment (Kleindalzig) and the association of the triggering rainfall events was performed using a modified version of the HydRun tool (Tang and Carey, 2017; with modifications to the runoff peak search algorithm). The months of November to May were discarded to avoid the effect of snowmelt.

Boxplots of computed hourly discharge for a set of catchments are presented in the paper. Boxplots are built according to McGill et al. (1978), i.e. they represent the median with the 25th and 75th percentile, while whiskers extend to the largest/smallest value but no more than 1.5-times the interquartile range.

While both the studied RCM (ICON12km) and the CPRCM (ICON3km) were found to compute mean hourly air temperatures lower than observed, the deviations fall within the measurement uncertainty bandwidth. Hackenbruch et al. (2016) also showed a cold bias for COSMO-CLM at 2.8 and 7.0 km resolution over southwestern Germany, and so did Tölle et al. (2018) for COSMO-CLM at 1.3 and 7.0 km resolution over central Germany. These studies found higher bias in summer than in winter, however our results show the opposite, likely due to the predominance of cropland and pasture in our study area (cf. Fig. S1). Over open land, cold biases are amplified due to higher albedo and shorter roughness lengths, the latter favouring the development of stable atmospheric conditions and allowing for stronger nocturnal cooling (Lind et al., 2020). These results underline the importance of scale and exemplarily show how dominating processes on the catchment scale can differ from those over larger study areas.

The CPRCM (ICON3km) was found to offer improvement for air temperature simulations in the summer months, in keeping with a variety of other studies conducted over Europe (Prein et al., 2015). These improvements have been linked to deep convection being explicitly resolved, which allows for better estimates of deep-convective clouds (Leutwyler et al., 2017; Hentgen et al., 2019) and a better representation of convective precipitation (Liu et al., 2017).

Furthermore, ICON3km exhibited a slightly lower spatial bias variability throughout the year compared to ICON12km, as reflected by the 95 % confidence interval from the station means. This finding is in line with Xie et al. (2025) who saw higher spatial variability in temperature bias by the RCM they used over western Norway compared the use of a CPRCM.

Both ICON3km and ICON12km were found to overestimate mean daily global radiation over the study area, however the deviations were not significant under consideration of the measurement uncertainty. Hackenbruch et al. (2016) found an underestimation of half-year mean sums of global radiation over southwestern Germany (1983–2000) for COSMO-CLM at 2.8 and 7.0 km resolution. In our study an underestimation was registered for the highest quantiles of daily global radiation both in intensity and frequency. ICON3km was able to reduce this bias. Through a better estimation of the onset of convection in the summer mornings and resulting reduced stratification, CPRCMs were shown to be able to dampen the RCMs' positive bias in the estimation of the frequency of high clouds (Keller et al., 2016). The simulation of more frequent clear-sky conditions allows for a reduction in the underestimation of global radiation (Keller et al., 2016; Leutwyler et al., 2017; Hentgen et al., 2019).

The skill of CPRCMs in depicting relative humidity is seldomly studied, however it has been shown that an increase in relative humidity constitutes a critical factor in deep convective initiation (Morrison et al., 2022).

No significant differences in mean hourly relative humidity were found for ICON3km or ICON12km compared to the observations under consideration of the measurement uncertainty. Monthly bias means proved negative for winter and autumn, but positive for spring and summer for both the RCM and CPRCM. Hackenbruch et al. (2016) found median relative humidity biases computed by COSMO-CLM at 2.8 and 7.0 km resolution over southwestern Germany to be positive for both half years, with biases much greater than seen for the ICON-CLM models. Possible explanations for the smaller biases in our study might be improvements to the model physics in the newer ICON-CLM (version 2.6.4) compared to the COSMO-CLM model they used, or differing model parametrisations.

ICON3km was found to estimate higher wind speed extremes than ICON12km, thereby slightly improving on the underestimation of extremes shown by both climate models. This result is in line with Iles et al. (2020), who found an increase in intensity and frequency of high winds with resolution, when comparing climate models of 12.5 and 50 km resolution over the European domain. Similar trends were found by Kunz et al. (2010) when comparing REMO-UBA at 10 km resolution and COSMO-CLM at 18 km resolution over Germany. ICON3km improved on the depiction of the frequency of light winds (0.6 to 1.7 m s⁻¹) computed by its driving model ICON12km, however both show an overestimation of the frequency of high wind speeds. As such, the simulation of wind speed remains a challenge for both climate models. Belušić et al. (2018) advance that it takes grid spacings of a few kilometres to accurately simulate small-scale wind systems. Hackenbruch et al. (2016) prove added value for the use of a CPRCM of 2.8 km resolution for the computation of channelled wind flow in the Neckar Valley and underline the importance of orographic structures for the initialisation of local wind systems.

ICON12km was found to compute too frequent light precipitation, while underestimating the intensity of heavy rainfall. ICON3km improves on the frequency estimation of light precipitation but overestimates the intensity of the highest quantiles. A reduction in the frequency of light precipitation and an increase in the intensity of heavy precipitation by the CPRCM compared to its driving RCM is in line with the consensus from literature (Lucas-Picher et al., 2021), e.g. with Ban et al. (2021) and Adinolfi et al. (2021). Ban et al. (2021) studied a 23-member ensemble of CPRCMs run at 3 km resolution and their driving RCMs at 12 km and found the CPRCMs to be able to alleviate the RCM-ensemble negative mean bias seen in the 99.9th-precipitation-percentile. In fact, the CPRCM-ensemble shows a respective mean bias within the acceptable range considering measurement uncertainty for all study areas and seasons, except winter (DJF) in Switzerland and summer in Italy (JJA) where the negative bias is weakened but remaining (Ban et al., 2021). Ban et al. (2021) furthermore found improvement through reduced wet-hour frequency by the studied CPRCM-ensemble. Adinolfi et al. (2021) show similar improvement by the CPRCM they analysed over an Alpine domain, stretching from central Italy to northern Germany, reducing precipitation frequency and increasing precipitation intensity compared to the driving RCM. Nevertheless, within the 99.9th percentile, the CPRCM intensity estimates remain slightly too low for precipitation values lower than 6 mm h⁻¹, but the CPRCM overestimates heavy precipitation higher than 6 mm h⁻¹ (Adinolfi et al., 2021). In their study over the southern United Kingdom, Kendon et al. (2012) showed an RCM of 12 km resolution to compute heavy rainfall as not intense enough, but as too persistent and widespread. A CPRCM of 1.5 km resolution in turn achieved improvements in the frequency distribution, duration and spatial extent, but estimated too intense heavy rainfall (Kendon et al., 2012). Xie et al. (2025) found an overestimation of annual maximum 1 d precipitation and maximum 1 h precipitation using the HARMONIE climate model in convection-permitting set-up at 3 km resolution compared to running it at coarser resolution (12 km) with parametrised convection. However, they found the CPRCM to outperform the RCM in capturing hourly precipitation above the 95th percentile (Xie et al., 2025).

RCMs with convective parametrisation have shown to underestimate heavy rainfall. In fact, small atmospheric instabilities are sufficient to trigger the convection schemes in the RCMs, leading to the release of light rain and thereby inhibiting a greater computational built-up of the convective cells (Ban et al., 2021). Parametrisation schemes were designed de facto to represent average effects at coarse resolution, however not localised extreme events and individual storms (Kendon et al., 2012).

CPRCMs in turn tend to overestimate the intensity of heavy rainfall, as updrafts are computed as too wide under the coarseness of the model grid (Kendon et al., 2023). Additionally, entrainment is insufficiently captured due to turbulence not being sufficiently resolved in the convective cells (Bryan and Morrison, 2012 cited in Quintero et al., 2022).

ICON3km was found to compute unrealistically high hourly rainfall intensities of up to 157.3 mm h⁻¹ over the study region, which is around twice as high as the maximum seen in RADOLAN3km (73.6 mm h⁻¹). These extremely high rainfall intensities were primarily estimated in the summer, and most were seen to cluster spatially and extend over several consecutive time steps. They are consequently likely the result of shortcomings in the parametrisation of the microphysics, rather than randomness or numerical instability.

ICON12km was found to compute too high mean annual precipitation totals, likely because of an overestimation of wet-hour frequency. This result is in agreement with findings by Strandberg and Lind (2021) for a 12.5 km resolution RCM over the European domain. The underestimation of average daily precipitation sums by RADOLAN compared to ground-truth rain gauge data (Kreklow et al., 2020) further exacerbates this positive bias.

The deviations of ICON3km and ICON12km to the adjusted radar data estimates upscaled to their respective resolution were found to be similar between the climate models during autumn and winter but differing most strongly during the peak of summer. In fact, synoptic weather systems, which are dominant in winter, are well represented even in coarse climate models (Strandberg and Lind, 2021) and the CPRCM is likely not providing much additional information or added value to their depiction. For summer however, when small-scale convective processes are at the forefront, the explicit representation of convection in the CPRCM likely leads to pronounced differences compared to estimates by its driving RCM where convection is parametrised (Strandberg and Lind, 2021).

Monthly negative precipitation anomalies were not found to be better represented by ICON3km than by ICON12km. They are predominantly shaped by large-scale atmospheric circulations, handed over to the CPRCM as lateral boundary conditions by its driving RCM. Taylor et al. (2013) however found a change from positive to negative in the sign of the soil-moisture precipitation feedback over their semi-arid domain when switching from an RCM with parametrised convection to a CPRCM. A negative feedback makes a drought less likely, whereas a positive feedback may prolong or intensify it (Taylor et al., 2013). Kendon et al. (2019) further show over an Africa-wide domain, that the use of a CPRCM (4.5 km resolution) can be of added value for the estimation of the length of dry spells through a more accurate depiction of the triggering of diurnal rainfall and an improved representation of the dying out of westward propagating systems.

Reszler et al. (2018) looked at soil moisture on a monthly basis and found estimates from a hydrological model driven with CPRCM data to be similar in winter and spring to those when using RCM input data, however the estimates deviate considerably in summer and autumn. In fact, during summer and autumn both climate models underestimate soil moisture, but the CPRCM is able to reduce the negative bias seen in the hydrological model driven with RCM data (Reszler et al., 2018). In contrast, ICON3km as input to the hydrological model WaSiM leads to an overestimation of hourly soil moisture throughout the year, while ICON12km leads to an underestimation. However, the results agree with the study of Reszler et al. (2018) on higher soil moisture in summer and autumn by the hydrological model driven with CPRCM data than when driven with RCM data.

Biases in observed and modelled meteorological variables can greatly influence evapotranspiration estimates. Meyer et al. (1989) introduced random and systematic errors to the input data of the Penman method to study the effect of such errors on the estimation of daily potential evapotranspiration. They found that random errors distort potential evapotranspiration estimates in a similar manner than systematic errors. Standard errors in potential evapotranspiration were fairly low when introducing random errors on wind run data, but three times as high when applying random errors of the same scale on temperature data, and four to five times as high when adding the random errors on relative humidity or solar radiation data (Meyer et al., 1989). In climate modelling, however the uncertainty is not equally distributed among variables. Lai et al. (2022) investigated the missing-variable estimation approaches by FAO and how the selection or exclusion of certain climate variables from global climate models shape the uncertainty of projected reference crop evapotranspiration using the Penman-Monteith method. They concluded that air temperature and shortwave radiation should be prioritised given the low projection uncertainty of these variables and high influence on reference evapotranspiration. Kay and Davies (2008) found the uncertainty introduced by the climate model structure to be higher than that by the potential evaporation formulation.

Mendoza et al. (2016) found higher basin-averaged mean annual evapotranspiration estimates in hydrological modelling when using meteorological forcing data from the WRF model at 4 km resolution in convection-permitting set-up than at coarser resolution of 12 or 36 km with parametrised convection. These results are in keeping with the study presented here, where higher mean hourly total evapotranspiration is registered for the hydrological model driven with ICON3km compared to when driven with ICON12km. During summer (JJA), the monthly mean biases of hourly catchment-average total evapotranspiration computed by the hydrological model driven with ICON3km and ICON12km deviate most strongly, with a shift from a negative bias by the driving model ICON12km to a positive bias by the driving model ICON3km. Higher total evapotranspiration by ICON3km, especially in summer, is likely the result of higher water availability through higher precipitation intensities in ICON3km, as well as a reduction of the negative summer bias in temperature and global radiation by ICON3km. Mendoza et al. (2016) also saw a wet bias in the CPRCM they studied. They furthermore analysed the partitioning of basin-averaged mean annual precipitation into basin-averaged mean annual evapotranspiration and found partitioning to be lower when the hydrological model is driven with CPRCM data than when driven with data from the RCMs.

All the discussed meteorological variables are used as input to the hydrological model and shape through their biases the accuracy of the discharge simulations. The hydrological model WaSiM driven with ICON3km simulates higher discharge than when driven with ICON12km. These findings are in agreement with the consensus from literature. Kay et al. (2015) report higher low flow volumes, median flows and high flow volumes when using CPRCM data compared to when using data from an RCM with parametrised convection. Mendoza et al. (2016) also saw higher basin-averaged mean annual runoff when driving their hydrological model with a CPRCM than with an RCM. Reszler et al. (2018) simulated higher maximum annual flood peaks using data from a CPRCM of 3 km resolution than from an RCM of 12.5 km resolution. Higher peak streamflows were also found by Schaller et al. (2020) using CPRCM data to simulate a selected autumn flood event on the West Coast of Norway compared to the use of GCM data. Davis et al. (2022) simulated higher (maximum) water levels in an urban catchment in India driving their hydrological model with data from a CPRCM ensemble than with data from the Global Forecasting System. Ascott et al. (2023) in turn computed greater daily river flows in the Lake Victoria basin when using CPRCM data compared to the use of RCM data. In keeping, Poncet et al. (2024) and Xie et al. (2025) found higher peak-over-threshold historical flood peak discharges using CPRCM data than when using RCM data to drive their hydrological models. Archer et al. (2024) found larger relative flooded areas from the historical intensity-duration-frequencies derived from a CPRCM ensemble than from interpolated rainfall observations.

Given its strong overestimation of precipitation intensity, ICON3km did not allow for improved discharge simulation compared to ICON12km. Kay et al. (2015) and Reszler et al. (2018) also attributed no added value to the use of their studied CPRCM for hydrological impact modelling. In fact, Kay et al. (2015) found worse performance in discharge computation using the 1.5 km CPRCM compared to using its driving 12 km RCM as input to the gridded hydrological model CLASSIC-GB, given the strong positive precipitation bias of the CPRCM when left uncorrected. Reszler et al. (2018) found deficiencies in the depiction of the temporal distribution of rainfall intensities to be another potential source of error. In a later study, Kay (2022) did however find added value in the use of a 2.2 km CPRCM compared to a 12 km RCM (each 12 member ensembles) for the simulation of daily low flow volume, median flow, and high flow volume across Britain. They attribute these improvements to changes in the physics of the used CPRCM compared to the earlier model employed in their 2015 study (cf. Kay et al., 2015). Also Mendoza et al. (2016), Schaller et al. (2020), Davis et al. (2022) and Poncet et al. (2024) found various added value in the use of CPRCMs for hydrological modelling (cf. Table 1). Xie et al. (2025) only found added value for the simulation of extreme flood peaks, characterised by return periods above 10 years.

The 99.5th percentile of hourly discharge was overestimated in the median by the hydrological model driven with ICON3km, as a result of the overestimation of the intensity of heavy precipitation. The input of ICON12km data to the hydrological model also resulted in an overestimation of hourly discharge of the 99.5th percentile in the median for six out of the seven catchments on the main stem, despite an underestimation of precipitation intensity. In fact, the rainfall-runoff-analysis (Fig. 14) showed a greater number of rainfall-runoff events of high return period when using the climate model data for hydrological modelling than when using meteorological observations. This is reflected through a more strongly scattered rainfall-runoff-plot in comparison to the use of observational meteorological data and translates to a higher interquartile range and a higher median of the 99.5th percentile of hourly discharge for the hydrological model driven with climate model data (Fig. S18). An overestimation of high discharges by the hydrological model driven with RCM data is also apparent in the study of Reszler et al. (2018). For four out of the six catchments they studied, the return periods from the hydrological model driven with 12.5 km RCM data overtake or at least come to match the return period curves from the observations for high return periods, while undercutting them for lower return periods.

We showed exemplarily for the flood event of 2013 in the Weiße Elster catchment that a simulated overestimation in precipitation does not always translate into an overestimation in flood peak discharge, as the storm cell might be shifted in space in the climate model simulation and rain off over soils with greater available storage capacity. In fact, in simulations run over vast domains, such as the Central European (CEU) domain in this study, the general events and patterns can be expected to match, but the climate models are unlikely to perfectly capture the actual geographic position of a local storm.

To conclude, no added value could be found in the use of uncorrected climate model data from ICON3km for hydrological impact modelling, given the overestimation of the intensity and frequency of heavy rainfall by ICON3km. The results presented here are drawn from a short time series (10 years) obtained from a single CPRCM run and reflect the hydroclimatic simulations over a single cluster of catchments. Longer time series, climate model ensembles and additional study areas are needed to draw firm conclusions on the general performance of CPRCMs for hydrological impact modelling.