The asynchronous modelling framework was previously explored by Ricard et al. (2019, 2020), mainly focusing on testing calibration metrics and implementing a more complex description of hydrological processes. Asynchronous modelling is analogous to a signature-based modelling in the way it aims to identify parametric solutions by optimizing the statistical properties of the simulated hydrograph, capturing the broad hydrologic behaviour of a catchment instead of the precise sequence of hydrometeorological events observed at the outlet. The purpose of asynchronous modelling is however different from signature-based modelling since it proposes constructing hydrologic scenarios according to a specific reconfiguration of the conventional hydroclimatic modelling chain, circumventing the requirement for meteorological observations typically used for post-processing raw climate model outputs and calibrating the hydrologic model.

The proposed asynchronous modelling workflow (Fig. 2) follows three main steps: (1) translating raw climate model outputs into corresponding hydrologic responses using asynchronous modelling, (2) computing change factors derived from a reference and a future simulated hydrologic response, and (3) constructing hydrologic scenarios by applying correction factors to the available streamflow observations. Asynchronous hydroclimatic modelling (Ricard et al., 2020) refers to an alternative configuration of the hydroclimatic modelling chain for which the calibration is performed on climate model outputs (over a recent past reference period) and not on meteorological observations. Since climate models cannot reproduce the observed sequence of meteorological events, we expect correlation-based calibration metrics (such as Nash–Sutcliffe efficiency – NSE – and Kling–Gupta efficiency – KGE) to mislead the identification of calibrated parameters within the asynchronous framework. The parameters of hydrologic models are thus optimized according to an objective function that purposely excludes the day-to-day temporal correlation (Ricard et al., 2019).

The calibration loop trains the hydrologic model in reproducing the statistical properties of the streamflow regime, such as the form of its cumulative distribution, quantiles, or moments, without taking into account the temporal match between them. We propose here a normalized score inspired by the continuous-ranked probability score (CRPS) (Matheson and Winkler, 1976), where the distribution of simulated streamflow is compared against the distribution of observations – the CRPS is commonly used to assess ensemble prediction systems. More specifically, the proposed score is defined such that 1nCRPSF,xobs=∫-∞∞Fx̃-Fx̃obs2dx, where x̃ and x̃obs are respectively the normalized simulated and observed streamflow time series, and Fx̃ refers to the temporal cumulative distribution of the streamflow. 2x̃=xmax⁡x,xobs3x̃obs=xobsmax⁡x,xobs In simple terms, the nCRPS is the squared difference between the normalized observed and simulated cumulative distribution functions, integrated with respect to the normalized streamflow. A perfect similarity between the two distributions indicates that the simulated values share the same statistical properties as the observations. In such a case, the area between the two curves would be null, and the nCRPS equals 0. The calibration loop being completed, raw climate model outputs are translated into corresponding hydrologic responses by forcing the calibrated hydrologic model over an application period, typically including both reference and future ones.

Hydrologic scenarios are constructed by applying a non-parametric quantile perturbation (Willems and Vrac, 2011) to the streamflow observations. Assuming stationarity of climate model biases, quantile perturbation (see also Willems, 2013; Sunyer et al., 2015; Hosseinzadehtalaei et al., 2018) typically modifies meteorological observations according to relative changes in the corresponding distributions projected by raw climate model outputs, preserving the simulated meteorological trends in all quantiles, including their tails (Cannon et al., 2015). In the proposed workflow, change factors are defined by relating quantiles of the simulated reference and future hydrologic responses produced by asynchronous modelling. Change factors (ϕ) are defined here as the ratio between the simulated streamflow values (x, associated with the exceedance probability p) of a future period (Fut) and a reference (Ref) period. Change factors encrypt projected trends for each streamflow quantiles such that 4ϕp,t=xFutp,txRefp,t, where t refers to a given temporal resolution, i.e. a prior sub-sampling of the annual cycle for which ϕ is evaluated (e.g. bi-annual, seasonal, monthly).

At this point, the future hydrologic regime can be assessed in terms of relative changes by analysing change factors for streamflow quantiles of interest. Hydrologic scenarios (xsce) are constructed by applying change factors (ϕ) to the available observed streamflow series (xobs) such that 5xscep,t=xobs⋅ϕp,t. The resulting hydrologic scenarios stand for plausible trajectories of the water regime conditions arising from a given climate simulation ensemble, statistically equivalent to the observed recent past that is affected by physically based long-term trends.

Table 3 lists the seven lumped conceptual hydrologic models used for simulating the hydrologic response corresponding to the nine NA-CORDEX simulations. The models are derived from various scientific and operational sources available from the HOOPLA open-source MATLAB^® toolbox (Thiboult et al., 2019). Models can be categorized as being of moderate complexity, the number of open parameters ranging from 6 to 9. All the models are combined with the Oudin evapotranspiration formulation (Oudin et al., 2005) and the snow module developed by Valéry et al. (2014), for which the two parameters, the thermal inertia of the snowpack (Ctg= 0.25, dimensionless) and a degree-day melting factor (Kf= 3.74 mm d-1), are fixed to default values that are relevant to the region. The selection of hydrologic models is based on the diversity of their structures and their combined performance for short-term streamflow forecasting (Valdez et al., 2022). The main idea here is to select a pool of heterogenous models in order to avoid that the simulated hydrologic responses are tainted by a single model structure.

Hydrologic models are calibrated according to an asynchronous modelling framework, i.e. being forced with raw climate model outputs and excluding the day-to-day temporal correlation (Ricard et al., 2019). The calibration loop is run from 1970 to 1979 with the shuffle complex evolution algorithm (Duan et al., 1993) using 10 complexes. A 10-year period is usually considered sufficiently long for calibration, offering a sound trade-off between identifying representative parametric values and computational requirements.

The proposed asynchronous workflow is compared to a conventional top-down hydroclimatic modelling approach. The latter is typically implemented to produce hydrologic scenarios from GCM or RCM simulations following three main steps (e.g. Poulin et al., 2011; Seiller and Anctil, 2014; Seo et al., 2016). Raw climate model outputs are first post-processed to correct systematic errors (or biases) according to available meteorological observations or a reference product describing the climate system over a recent and sufficiently long past period. Simulated 2 m minimum and maximum air temperature and precipitation are corrected using a quantile mapping approach (Lucas-Picher et al., 2021) combined with daily local intensity scaling (Schmidli et al., 2006). Quantile mapping is implemented every month using 100-node transfer functions interpolated linearly. The wet-day frequency is corrected using a 0.1 mm threshold. Hydrologic models are calibrated separately, forced with gridded meteorological observation datasets to optimize the performance of the simulated hydrologic response according to available streamflow observations. Hydrologic scenarios are finally constructed by forcing the calibrated hydrologic models with post-processed climate model outputs. For the sake of comparison, hydrologic modelling within the conventional hydroclimatic approach is implemented equivalently to the asynchronous workflow as described in Sect. 3.2, using the same pool of hydrologic models and calibrated parameters and the same objective function, calibration period, and configuration of the optimization algorithm.

Figure 3 illustrates the annual cycle of the 2 m mean air temperature (2mt) simulated by the nine NA-CORDEX simulations from 1970 to 1999. Only sub-catchment 2 is shown, considering that it represents most of the area of the Chaudière River catchment, but also because sub-catchments 3 and 4 are nested within. Corresponding observations issued by interpolation of in situ measurements and biases are also illustrated. Most climate simulations overestimate 2mt from November to March, the median bias of the ensemble reaching roughly +5 ∘C in January. NA-CORDEX simulations generally provide a reasonable representation of temperature from May to September, individual biases then ranging from -2 to +2 ∘C from one simulation to another. 2mt biases appear to be linked to the forcing GCM simulations. CanESM2-driven simulations (crx1 to crx3) lead to similar annual profiles marked by an alternation of high warm winter biases and subsequent moderate warm summer biases. EC-EARTH-driven simulations (crx4 and crx5) show a similar annual profile to CanESM2 but are affected by marked cold spring and summer biases, reaching -5 ∘C in April in the case of crx4. GFDL-ESM2M-driven simulation (crx6) is affected by a quasi-systematic cold bias. MPI-ESM-LR-driven simulations (crx7 to crx9) finally show a constant cold bias from May to November. The winter warm bias carried by crx7 (CRCM5-UQAM, positive) differs however from the winter cold biases of crx8 and crx9 (RegCM4 and WRF).

Figure 4 illustrates the mean annual cycle of the precipitation simulated by the nine NA-CORDEX simulations from 1970 to 1999 over sub-catchment 2. The ensemble mean overestimates precipitation by roughly +0.5 mm d-1 (∼ +27  %). In contrast to 2mt, biases in precipitation are fairly constant throughout the whole annual cycle, except for a brief period in autumn (August to October) when simulations are less biased. Biases typically range between -1 and +2 mm d-1 depending on the period of the year. Part of the wet bias in winter precipitation can be explained by solid precipitation undercatch, which can reach 20  % to 70  % (Pierre et al., 2019). Also, in contrast to 2mt, biases in annual profiles are not as clearly related to the driving GCM.

Figure 5 illustrates seasonal changes (2040–2069 relative to 1970–1999) for sub-catchment 2 for the mean 2mt and precipitation from the nine NA-CORDEX simulations. An increase in 2mt generally falls between +2 and +4 ∘C. Also, most simulations anticipate precipitation increasing in winter (+10  % to +25  %), spring (up to +20  %), and autumn (up to +15  %) but decreasing in summer (down to -10  %). Some simulations reveal outlying trends, especially crx3 and crx4, which display, respectively, a +44  % increase in winter precipitation and almost no change in 2mt from September to November.

Figure 6 displays the observed mean annual hydrographs at site 2 over a recent past reference period (1970–1999). The hydrograph shows typical seasonal fluctuations marked by spring flood (∼ 9.5 mm d-1) in April and a second peak (much smoother, ∼ 1.8 mm d-1) in November. Figure 6 also compares mean annual hydrographs simulated by the asynchronous framework and the conventional hydroclimatic approach for each NA-CORDEX simulation. The results show the capacity for the conventional approach to provide a more accurate representation of seasonal streamflow fluctuations over the reference period. Although slightly delayed and underestimated, the peak flow simulated in spring by the conventional approach is typically more accurately synchronized with observations relative to the asynchronous workflow. The inter-model variation (indicated by the envelopes in Fig. 6) related to the conventional approach also tends to be smaller and more centred around the observations, noticeably during summer, autumn, and winter. The shape of the simulated hydrographs remains finally quite similar from one NA-CORDEX simulation to another.

Hydrographs simulated by the asynchronous workflow are in some cases affected by notable flaws in representing seasonal streamflow fluctuations. The shape of the simulated hydrographs also differs notably from one NA-CORDEX simulation to another. This can be related to biases affecting raw forcing NA-CORDEX simulations (see Figs. 3 and 4). In many cases (crx4, crx6, crx8, and crx9), the spring flood is notably delayed and occurs in late spring. This could be explained by cold biases affecting simulated air temperature in spring, combined in some cases with an overestimation of solid precipitation in winter. The inter-model variation also tends to be larger relative to the conventional approach, more noticeably during summer and autumn (crx1, crx2, crx3, and crx5), but also in winter (crx1).

Hydrographs simulated at sites 1, 3, and 4 are given in Appendix A and lead to equivalent conclusions. Hydrographs simulated by the conventional approach at site 3 however produce an atypical two-fold spring flood that can be related to a specific hydrologic model. The inter-model variability is also more marked in the case of site 4 for the asynchronous framework.

Figure 7 compares the hydrologic performance of NA-CORDEX simulations (crx1 to crx9) issued by the conventional modelling approach and the proposed asynchronous workflow at sites 1 to 4. Performance is sorted according to the root-mean-square-error (RMSE) value between simulated mean annual hydrographs and corresponding observations over the 1970–1999 reference period. The median RMSE value out of seven hydrologic model simulations is presented here. The results first confirm the systematic capacity of the conventional modelling approach to provide a more accurate representation of the inter-annual hydrograph, corresponding RMSE values ranging from ∼ 0.9 to 1.2 mm d-1. The performance issued by the conventional approach is also notably comparable from one site to another. On the other hand, the asynchronous workflow produces a systematically less accurate representation of the mean annual hydrograph. The most-performing simulations (ranks 1 to 5) are affected by RMSE values ranging from ∼ 1.3 to 1.6 mm d-1, which is comparably performant relative to the conventional approach. A marked degradation is however observed for other less-performing simulations (ranks 6 to 9, RMSE reaching ∼ 2.5 to 3.0 mm d-1 depending on the site). Sorting NA-CORDEX simulations according to their hydrologic performance systematically points to the same discrimination between the pool of the most-performing simulations (crx1, crx2, crx3, crx5, and crx7) and the less-performing ones (crx4, crx6, crx8, and crx9).

Figure 8 compares projected changes in the seasonal mean flows simulated by the conventional hydroclimatic modelling approach and the proposed asynchronous workflow at sites 1 to 4. Changes are expressed in relative terms (percentage) for the nival (DJFMAM) and pluvial (JJASON) regimes and are grouped according to the driving NA-CORDEX simulation (boxes). The top five most-performing NA-CORDEX simulations identified for the asynchronous workflow are highlighted in blue. A group of outlying changes (all related to hydrologic model 1) projected by the conventional approach at site 3 is also identified in red. NA-CORDEX simulations being analysed separately, Fig. 8 shows discrepancies in change values from one modelling approach to another. Site 2 being given as an example, the spread of changes projected by crx2 is noticeably reduced using the asynchronous workflow in comparison to the conventional approach. A shift in the projected direction of change for the pluvial mean flow (ΔJJASON) can also be observed, from a plausible decrease in the case of the conventional approach to a very likely increase for the asynchronous workflow. On the other hand, Fig. 8 also shows that both approaches lead to comparable interpretation if the projected changes are analysed as an ensemble. Site 2 once again being given as an example, both approaches strongly agree in projecting an increase in the nival mean flow (ΔDJFMAM). Both approaches also agree in projecting a decrease in the pluvial mean flow (ΔJJASON), except for a portion of projections mostly related to the crx2 simulation. The fact that the interpretation of the projected changes remains equivalent for both approaches can be generalized to all the sites.

Figure 8 shows that the asynchronous workflow tends to provide more outlying changes values in comparison to the conventional approach. For all the sites, numerous projections indicate very strong increases in the nival mean flow (ΔDJFMAM), reaching up to ∼ +100  %. Such outlying projected changes are however systematically related to the less-performing NA-CORDEX simulation identified in Fig. 7. The sub-ensemble of change values resulting from the selection of the most-performing simulations (blue boxes) provides a reliable interpretation of the hydrologic changes with regard to the conventional approach, here considered the benchmark. The conventional approach can also produce notable outlying change values of nival mean flow in the specific case of site 3, all related to hydrologic model 1. Changes in the seasonal high flows and low flows projected by both approaches are presented in Appendix B.

Table 4 summarizes the distributions of change values projected by the conventional modelling approach and the proposed asynchronous workflow. Results are displayed for site 2 using six hydrological indices describing seasonal (DJFMAM vs. JJASON) mean, high, and low flow. High- and low-flow indices are computed based on annual maximal (and minimal) values considering a 2-year return period. Distributions are composed of all possible combinations between NA-CORDEX simulations and hydrologic models (n=63) for the conventional approach and by the selection of the top five most-performing simulations for the asynchronous workflow (n=35). Change distributions are described using the “direction” of change (the portion of values pointing to an increase in a given index), the median value, and its standard deviation.

The results first confirm a strong agreement between both approaches in interpreting the changes in mean flow indices, the projected increase being equivalent in terms of direction (98 % vs. 100 %), median values (+15 % vs. +16 %), and standard deviation (9 % vs. 8 %). Both approaches also agree, but to a lesser extent, on the projected decrease in pluvial mean flow. The direction (10 % vs. 23 %), the median change value (-17 % vs. -10 %), and the standard deviation (14 % vs. 17 %) lead to a comparable interpretation of the change signal.

Modelling approaches do not agree as strongly in projecting high flows. While the asynchronous workflow indicates a probable increase in nival high flows (direction = 77 % and median =+10 %), the conventional approach rather provides a blurred signal. The direction of change (41 %) indicates a weak consensus among projections, and the median change value is small (+1 %). In this case, the standard deviation is comparable between both approaches (11 % vs. 12 %). The opposite situation is observed for pluvial high flows where the conventional approach projects a probable decrease (direction = 21 % and median = -13 %) and the asynchronous, distorted, and vague change signal (direction = 40 %, median = -2 %, standard deviation = 23 %).

Modelling approaches agree on the direction of change for low flows. They both indicate a probable increase in the nival lows flow (79 % vs. 77 %) and a probable decrease in the pluvial low flow (6 % vs. 16 %). The conventional approach however suggests a more severe increase in the nival low flow (median value =+56 %) relative to the asynchronous workflow (+21 %). The spread of the distribution is notably high in the case of the conventional approach (+132 %). Both approaches finally roughly agree on median change values (-18 % and -12 %) and standard deviations (18 % and 12 %) for pluvial low flows.

Figure 9 illustrates change factors (ϕ) computed as prescribed by Eq. (4) issued by the asynchronous modelling framework, displayed for each streamflow quantile at site 2. Change factors are computed on an annual basis (all data, no sub-sampling of the annual cycle) and for the nival (DJFMAM) and pluvial (JJASON) regimes that both experience low- and high-flow periods. Change factors are defined from percentile 0.005 to percentile 0.995 by increments of 0.01 (100 nodes), interpolated linearly. Results are shown for all NA-CORDEX simulations and for the selected most-performing ones, respectively. Annual factors show few little projected changes in streamflow quantiles from the reference to future periods. They confirm an increase for lower quantiles (ϕ roughly ranging between 0.9 and 1.5), while no clear change signal can be observed for higher quantiles. On the other hand, nival change factors (DJFMAM) show much more marked projected changes from the reference to the future. While all simulations agree on an increase for smaller streamflow quantiles (ϕ ranging between 1 and 2), ϕ reaches the value of 2.9 for quantile 0.8. ϕ abruptly decreases for quantiles above 0.9, ranging between 0.9 and 1.3. Pluvial change factors (JJASON) are not as marked as nival factors. They confirm however a consensual decrease for quantiles below 0.8. The consensus weakens for higher quantiles, corresponding ϕ values being centred around 1 and affected by a larger spread. One must notice that the selection of NA-CORDEX simulations based on hydrologic performance typically agrees with the ensemble composed by all simulations, except for projecting a nival streamflow quantile from 0.5 to 0.9. In this case, the selected simulations provide a much smaller increase in nival high flows, ϕ typically being below 1.5.

Figure 10 displays the hydrologic scenarios over the 2040 to 2069 period according to RCP8.5, produced over the Chaudière River sub-catchments by applying the quantile perturbations to the observed streamflows. Observed and projected hydrographs are shown for the selected year 1982, given as an example. The hydrologic scenarios reflect the relative changes embedded within the distributions of change factors shown in Fig. 9. Future winter low flows are systematically higher relative to the observations. Mid-amplitude spring high flows are also affected by notable increases, which is not systematically the case for high-amplitude peak flows. Summer low flows tend to decrease, while summer and autumn high flows are affected by moderate increases and decreases, depending on the climate scenario.

Nowadays, the quantification of climate change impacts on water resources mostly resorts to the implementation of top-down modelling cascades, translating climate model outputs into simulated hydrologic time series at the catchment scale. Typically, a statistical post-processing is applied to the raw climate model outputs in order to reduce biases imbedded in the simulated climate variables. Hydrologic models are also typically calibrated when forced by meteorological observations aiming to identify optimal parameter sets minimizing errors between simulated and observed discharge at a given catchment outlet. Assessing the impact of climate change on the hydrologic regime of a catchment using this conventional modelling approach presents drawbacks documented in the scientific literature: (1) the statistical post-processing of climate model outputs may disrupt the physical consistency between the simulated climate variables and even alter the corresponding trends from a reference period to a future period, (2) the modelling work flow relies highly on the availability and quality of meteorological observations in order to conduct the statistical post-processing of climate model outputs and the calibration of the hydrologic model, and (3) it also requires a high level of expertise and computing capacities to post-process the outputs and uses non-trivial statistical methods, restraining the participation of end-users in interpreting and attributing confidence in the simulation results.

In this study, we propose a simplified modelling workflow that enables the production of hydrologic scenarios without resorting to the statistical post-processing of climate model outputs. This asynchronous approach is conducted by calibrating the hydrologic model forced directly with raw climate model outputs instead of meteorological observations, using an objective function that excludes the temporal correlation between the observed and simulated hydrologic responses. Calibrated hydrologic models allow for the conversion of a raw climate model into corresponding reference and future simulated hydrologic responses. Hydrologic scenarios are subsequently produced by applying quantile perturbations to available streamflow observations, with the perturbation factors identified by relating simulated reference and future hydrologic responses for each streamflow quantile. Quantile perturbation is applied to simulated climate variables such as precipitation or reference evapotranspiration (Ntegeka et al., 2014) but never, to our knowledge, to the simulated hydrographs resulting from a hydroclimatic modelling cascade. To our knowledge, two approaches help preserve the physical consistency of climate model outputs and their trends: to apply trend-preserving multi-variate methods or to use raw model outputs straightforwardly for impact analyses, accepting biases. The proposed asynchronous framework is based on calibrating a hydrologic model using raw model outputs, assuming a consistent relative change (within climate simulations) from the reference to future periods. We acknowledge (and discuss below) the requirement for calibration as a limitation of the proposed framework, considering that it may disrupt the consistency of simulated hydrologic processes at the catchment scale.

We assessed the proposed asynchronous workflow by comparing its projected hydrologic regime with a conventional hydroclimatic modelling approach. As shown by others (e.g. Muerth et al., 2013), our results confirmed that the post-processing of raw climate model outputs increases the performance of the simulated hydrologic response over the historical reference period. On the other hand, our results demonstrated that the projected changes in the seasonal mean flows, taken as ensembles, converged to equivalent conclusions, regardless of the chosen modelling approach. The concordance between both approaches did not occur as sharply for high- and low-flow indices, suggesting that further investigations would be required to clarify how and to which extent the projection of high- and low-flow events is sensitive to the selection of the hydroclimatic modelling approach. We here emphasize the fact that the asynchronous workflow is vulnerable to strong biases affecting raw climate model outputs and is consequently more prone in producing outlying projections of hydrologic indices. However, the performance of the simulated response over the reference period provided a functional criterion to identify less-performing NA-CORDEX simulations. Based on the results shown in Sect. 4, we would advocate for the exclusion of these simulations for the analysis of the simulated projections of the hydrologic regime using an asynchronous modelling framework.

Although the proposed asynchronous framework does not completely solve the weaknesses of the traditional modelling approach, it presents the following benefits. (1) It increases confidence in the hydrologic scenarios since it is conducted with raw climate model outputs, thus preserving physical consistency between simulated climate variables and original trends simulated by the climate models (although it requires the calibration of a hydrologic model discussed below) – some authors also foresee that raw climate model outputs will improve in resolution and reliability with time (e.g. Teng et al., 2015; Chen et al., 2017). (2) It does not resort to meteorological observations, not for operating statistical post-processing or for calibrating the hydrologic model, facilitating the assessment of climate change impact on water resources for regions afflicted by observation scarcity (a significant benefit since most of the earth system is affected by data scarcity) – we would also argue that our approach does not inject uncertainty into the modelling cascade from the intrinsic limitations of post-processing methods (Laux et al., 2021) or from poor-quality observations or reference products describing the reference climate system (Hwang et al., 2014; Kotlarski et al., 2017). (3) It is simple to implement and is lighter in computing requirements – post-processing is exclusively applied to streamflow instead of numerous climate variables.

We believe however that the proposed workflow should be used wisely in areas where meteorological observations are abundant and reliable, rather as a complement to than as a substitute for conventional hydroclimatic modelling. In such cases, we would definitively encourage a sound use of all meteorological observations.

Further works could explore the use of bias correction within the asynchronous workflow, aiming to maximize the use of observations while producing hydrologic scenarios at the regional scale or modelling more complex physical processes at the catchment scale. Another assessment scheme could compare the performance of both modelling frameworks with intentionally degraded (scarcer) forcing data. Such comparison could confirm under which conditions the use of a given framework would be preferable over another.

Statistical post-processing of climate model outputs implies a necessary trade-off between key methodological benefits and drawbacks in the scope of providing reliable and supportive information for adaptation to climate change. On the one hand, simulated climate variables are corrected to fit statistical properties of the observed climate system. On the other hand, statistical post-processing disrupts physical consistency and alters trends in the simulated climate variables. While designing statistical post-processing, a decision is implicitly taken on how these benefits and drawbacks are weighted. In a pure top-down perspective, statistical post-processing is applied according to climate-oriented prerogatives, the end-user rarely being involved in deciding upon which benefit to be prioritized and which drawbacks to be limited. Moreover, not communicating source biases affecting raw climate model outputs constrains the capacity of impact modellers and end-users have in assessing the climate model representativeness and attributing confidence to resulting climate scenarios. Nowadays, solutions explored by the scientific community mostly resort to the development of sophisticated post-processing methods. Even though such approaches present undeniable benefits in terms of post-processed physical consistency and trend preservation, we would argue that they further enlarge the gap between climate specialists and water resource end-users.

The approach proposed in this study remains in essence a top-down modelling workflow. Through notable simplifications and straightforward constructions between raw climate model outputs and impact models, this alternative framework creates a space for an increased participation of climate model experts, impact modellers, and end-users in interpreting climate change impacts on water resources (Ehret et al., 2012). It is thus compatible with integrated and transdisciplinary environmental assessments and modelling frameworks in support of decision and policy making (Hamilton et al., 2015; Rössler et al., 2019). By translating raw climate model outputs into the corresponding simulated hydrologic responses, the representativeness of climate models can be assessed in a language further understandable for impact modellers and end-users. Based on the simulated hydrologic responses over the reference period (see Mudbhatkal and Mahesha, 2018), key methodological questions can be addressed and debated through an open and empowered dialogue with climate specialists. These questions can be the following.

Are climate model outputs representative enough to assess the impacts of climate change on water resources?

We believe that decisions on such questions require a sound understanding of simulated climate forcing but also an in-depth awareness of the local specificities of the hydrologic system exposed to climate change. Considering the above arguments, one could rather use the proposed asynchronous workflow as a hybrid analytical framework to evaluate the vulnerability of water resource systems instead of as a pure top-down predictive assessment tool.

The assessment of the proposed asynchronous workflow indicated that the simulated hydrologic response is affected by systematic errors (or hydrologic biases), mostly notable in terms of synchronism of the mean annual hydrograph during spring flood. Considering this, we would advocate that the proposed workflow should be used with caution when focusing on analysing high- and low-flow events. To formally assess the impact of climate change on a given domain, however, a larger ensemble of climate simulations should be considered. Since the workflow does not involve statistical processing of climate model outputs, we would recommend the use of high resolution over coarse gridded climate simulations in order to rely on an improved representation of local-scale processes. The use of seven conceptual lumped hydrologic models can also be considered a limitation to our approach. Although they provide a diversity in modelling structure, no formal evaluation of this specific source of uncertainty has been considered in this study (calibration metric, calibration period, structure complexity).

Assessing the impact of climate change on water resources within the proposed framework implies that the resulting hydrologic scenarios are inevitably tainted by (hydrologic) biases. These biases emerge from raw climate model outputs but also from the limitations imposed by the structures of the hydrologic models. We believe that further work should focus on evaluating how these two sources are intertwined. We also acknowledge that the proposed approach may disrupt the physical consistency of the processes simulated at the catchment scale through parametric compensation affecting the calibration of the hydrologic model. Further work is required to assess how parametric compensation may affect the trade-off between hydrologic scenarios fitted to observations and the preservation of the hydrologic change signal embedded within raw climate model outputs. Such analysis could also clarify the impact of parametric compensation relative to bias correction of raw climate model outputs. In the meantime, we would argue that parametric compensation should be minimized as much as possible to preserve the hydrologic change signal. This could be achieved, for example, by restraining parametric spaces during calibration as closely as possible to realistic boundaries or favouring physically based descriptions of hydrologic processes. Even if climate models constantly improve, their biases can still be important, and a judgement must be made in order to attribute confidence to the resulting hydrologic scenarios. Chen et al. (2021) explicitly raised the idea of an optimal selection of climate simulations before producing hydrologic scenarios to cope with their limitations in representing local hydrometeorological patterns. We do not propose here any specific guidelines, except that such an attribution must consider the scope and objectives of the conducted study and should involve, as much as possible, climate specialists, impact modellers, and end-users.

The proposed workflow is not limited by available meteorological observations, but to available streamflow observations. To assess the impact of climate on ungauged water resources, modellers can translate the hydrologic perturbation signals under the assumption of representativity of available discharge observation with regards to the ungauged domain. If ungauged streamflow is estimated before applying a change factor (using area ratio, hydrological modelling, or optimal interpolation), the corresponding uncertainties must by considered.

Constructing hydrologic scenarios using quantile perturbations, our results demonstrated the necessity of identifying a suitable time period to define change factors. Such resolution must consider specificities of the local flow regime magnitudes. The identification of an optimal duration remains an open question, keeping in mind that the use of a moving window could become necessary to compensate for a breakpoint in the hydrologic scenarios. We acknowledge that the quantile perturbation assumes a comparison between two stationary periods (reference vs. future) and does not consider potential rupture in future trends. We believe that shifts in the seasonal cycle could theoretically be more precisely assessed by applying sub-annual perturbation factors. Even considering the relative change for each streamflow quantile, the capacity of quantile perturbation to preserve mean flows and seasonal budgets should be explored and assessed further.