This study uses an extended version of the daily LArge-SaMple DAta for Hydrology and Environmental Sciences for Central Europe (LamaH-CE; Klingler et al., 2021). LamaH consists of 859 gauged catchments including 21 catchment averaged meteorological variables, with more than 60 static catchment attributes. Since the original version of LamaH only contains meteorological ERA5-Land data and Kratzert et al. (2021) show that leveraging multiple meteorological data sources is beneficial in large sample hydrology, we expanded the data by 15 further variables from five sources. The products used are (i) ERA5-Land (Muñoz-Sabater et al., 2021) as in the original LamaH data, (ii) ECMWF-HRES European Center for Medium Range Weather Forecast - High Resoultion Forecast (ECMWF, 2025), (iii) E-OBS gridded observational data (Cornes et al., 2018), (iv) MSWEP multi-source weighted ensemble precipitation (Beck et al., 2019) and (v) GLEAM global land evaporation Amsterdam model (Miralles et al., 2011). Details of the variables used, including their definitions, units, and sources, are summarized in Appendix A. The data products were obtained as raster data and subsequently aggregated to the LamaH basins. All variables are daily averages (e.g. temperature) or daily sums (e.g. precipitation). For the ECMWF-HRES variables temperature, dew point and sea level pressure, 3 hourly forecast values (8 per day) were calculated as daily averages starting from 0 o'clock (UTC) issue time. A second adaption we made to the LamaH dataset concerns the gauge files. In the daily version of LamaH-CE, there are only the mean daily discharges – we have extracted the daily minima and maxima from the hourly LamaH data for all gauges and extended the daily version with those.

For the conducted experiments, we used a subset of 451 basins with no and low anthropogenic influence at LamaH aggregation level A, which represents the lumped topographic catchment area of a gauge. Level A is comparable to the aggregation of the catchment areas in the CAMELS (Newman et al., 2015) dataset. The catchments are spatially distributed across the entire LamaH domain, with catchment areas including high alpine-, alpine foothill- and lowland areas. The subset includes both headwater and nested catchments, with 72.5 % of the basins being headwater catchments and 27.5 % being nested, meaning they have at least one other gauged basin from the selection upstream.

To comprehensively evaluate the performance of LSTM-based flood prediction models under different data availability scenarios and training strategies, we designed five distinct experimental groups with the primary research question: How to reduce the meteorological forecast induced bias in runoff predictions?

All experiments were, in terms of reproducibility, conducted with the NeuralHydrology (Kratzert et al., 2022) python library and trained on different LSTM architectures to predict maximum daily discharge (qmax). The models incorporated dynamic meteorological inputs using a 365 d input sequence length, and static catchment attributes (33 physiographic, climatic, and land cover characteristics), both processed through separate embedding networks. The embedding networks are fully connected neural network layers that transform raw input variables into learned representations (Ahmed et al., 2023). A primary motivation for their use in this study is to enable transfer learning and consistent processing across experiments with different input dimensions. By projecting varying numbers of input variables from different data sources into a shared latent space, the model can be pre-trained on reanalysis and fine-tuned on forecast data regardless of differences in the feature space between domains. The experimental framework utilized a consistent temporal split with training data from 2003–2009, validation from 2010–2013, and testing from 2014–2017. Model performance was evaluated using basin averaged Nash-Sutcliffe Efficiency (NSE*, Kratzert et al., 2019c) as the primary loss function. A description of the loss function is attached in Appendix C. Model hyperparameters, such as the number of hidden units, were optimized using Bayesian optimization (see Snoek et al., 2012) with NSE* as the objective function. A detailed description of the performed hyperparameter tuning is attached in Appendix D. Hereafter, we use the term “domain” in its machine learning sense, referring to a specific data distribution characterized by its feature space and statistical properties, rather than in its common hydrological meaning of a geographical region. Throughout the remainder of this paper, domain specifications and equations reference only dynamic meteorological forcings for clarity, with the understanding that static catchment attributes remain unchanged across all experimental setups.

The premise of this study is that the distributional differences between reanalysis and forecast meteorological data, as visually evident in the violin plots of Appendix B, lead to substantial performance degradation when transitioning between the two domains. To investigate the spatial patterns of this domain shift and their relationship to catchment characteristics, two complementary analyses are conducted. First, the 1-Wasserstein distance (Villani, 2009) is computed for each of the five shared meteorological variables between the reanalysis and forecast distributions across all 451 basins during the test period (2014–2017), using the input domains from the cross-domain evaluation experiment. The 1-Wasserstein distance measures the minimum effort required to transform one probability distribution into another and is computed in the original physical units of each variable to preserve interpretability. It is defined as: 15W1P,Q=∫-∞+∞FPx-FQxdx where FP and FQ are the empirical cumulative distribution functions of the reanalysis and forecast distributions respectively. The resulting per-basin Wasserstein distances are then correlated with the 33 static catchment attributes listed in Table A2 using Spearman correlation and visualized as a heatmap to identify which physiographic characteristics are associated with larger meteorological domain shifts.

Further, to investigate how catchment characteristics modulate model skill across all experimental configurations a two- fold analysis is conducted. First, all basins of the experimental domains DRA, DED-LSTM,2, DSEQLSTM,1, DSEQLSTM,2 are compared against those of the Forecast Baseline DFCby calculating the per basins difference of the NSE values to assess in which basins the model performance possibly increases or decreases – defined as: 16ΔNSE=(DRA,DED-LSTM,1,DED-LSTM,2,DSEQLSTM,1,DSEQLSTM,2)-DFC Inspired by the work of Seibert et al. (2018), suggesting to compare hydrological model experiments against well-defined benchmarks, we introduce a data plot technique to compare experimental differences in large sample hydrology settings: Plotting a heatmap with the model experiments as rows, ΔNSE per basin as columns and sorting the columns by the mean ΔNSE over all experiments per basin (columns). This plot allows a simple visual comparison of all experimental differences per basin and enables detecting outliers. While Seibert et al. (2018) propose a simple conceptual bucket-type model as the lower benchmark, we use the Baseline Forecast as the lower reference because it represents the most operationally relevant starting point for our specific problem: a LSTM model trained and tested exclusively on operational ECMWF HRES forecast data, without any bias correction strategy applied.

Second, the Spearman correlation between all positive per-basin ΔNSE and all 33 static catchment attributes is computed for each experiment and visualized as a heatmap. The focus on positive error metric differences (ΔNSE>0) was chosen since it indicates those basins where the proposed strategy outperforms the Baseline Forecast, effectively capturing the physiographic controls on forecast skill improvement. It is important to note that higher positive ΔNSE values do not exclusively reflect bias reduction – they also incorporate improvements stemming from increased model complexity, such as the additional architectural components and discharge integration introduced in the more advanced configurations.

A fast and straightforward way to analyse the propagation of forecast uncertainty in predicting maximum daily discharge (qmax) is the cross-domain evaluation (CD) experiment, depicted in Fig. 3. The CD experiment replicates the classical hydrological modelling workflow, where a model is calibrated on reanalysis data and subsequently applied with forecast inputs. To ensure that the input feature space remains identical between training and inference, only the five variables that have a direct equivalent in both reanalysis and forecast are used (ERA5L_2m_temp_mean, ERA5L_2m_dp_temp_mean, ERA5L_surf_net_solar_rad_mean, MSWEP_RR, GLEAM_ETA), in contrast to the Baseline Reanalysis model which uses all 31 available reanalysis variables from ERA5-Land, E-OBS, MSWEP, and GLEAM. CD reveals degradation in hydrological model performance when transitioning from reanalysis to forecast meteorological forcings, despite five identical input variables. The median Nash-Sutcliffe Efficiency (NSE) decreased from 0.58 to 0.33, representing a 0.25 reduction in model skill. This performance deterioration is accompanied by increased uncertainty, with the NSE standard deviation rising from 0.87 to 1.1, indicating that forecast uncertainty propagates through the hydrological model shifting and broadening the NSE distribution. The mean NSE exhibits an even more pronounced decline (0.44 to 0.19), suggesting increased negative skewness due to extreme poor-performing outliers.

The underlying cause of this performance decrease can be attributed to the difference in data distributions between reanalysis and forecast datasets (see Appendix B), effectively representing a domain shift problem. Domain shift occurs when the statistical properties of the training data (reanalysis) differ from those of the target data (forecast), violating the fundamental assumption of independent and identically distributed data that underlies machine learning model generalization (Goodfellow et al., 2016; Hosna et al., 2022). Neural networks are particularly susceptible to domain shifts as they learn to map input-output relationships based on the specific distributional characteristics of their training data, leading to degraded performance when deployed on data from a different distribution. These results demonstrate that it is not feasible to simply substitute reanalysis data with forecast data in neural network-based hydrological modeling applications, as meteorological forecast uncertainty propagates through the model chain, degrading the representation of catchment processes and creating performance risks for operational hydrological forecasting systems.

To address the domain shift challenges identified in Sect. 3.1, we evaluated different neural network architectures and training techniques, presenting here the optimal configurations from each experimental setup.

Transfer learning was implemented through two contrasting approaches: comprehensive parameter updating, where the entire network was retrained on forecast data (TL AllWeights), and selective embedding retraining, where only the input embedding layers were fine-tuned while maintaining the pre-trained weights of the deeper network components (TL EmbeddingNet). Both experiments utilized the reanalysis baseline with a complex embedding layer as the starting point. The selective embedding approach (TL EmbeddingNet) achieved higher performance, with a median NSE of 0.44, representing a 0.11 improvement over the cross-domain baseline (0.33). The advantage of selective retraining becomes evident when comparing the two transfer learning strategies: TL EmbeddingNet shows an improvement in the 10th percentile (0.15) compared to TL AllWeights (0.05), indicating that the retraining approach is especially effective for poorly performing basins. The enhanced performance stems from the model's ability to leverage robust hydrometeorological relationships learned from reanalysis data while adapting the input representation to forecast data characteristics. Reanalysis products provide more complete and physically consistent atmospheric descriptions through data assimilation, enabling the model to learn generalized process representations that are subsequently refined during fine-tuning on forecast data.

The Encoder–Decoder LSTM architecture demonstrated performance improvements over the transfer learning approaches, achieving a median NSE of 0.57. This architecture showed particular strength in the upper performance range, with the 75th and 90th percentiles reaching 0.65 and 0.73, respectively.

The Sequential Forecast LSTM achieved the highest overall performance among the unmodified forecast-based configurations, with a median NSE of 0.63 and notably consistent results across all percentiles. The model demonstrated higher stability compared to other approaches, evidenced by the low standard deviation (0.52). Among the unmodified architectures, the Sequential LSTM is best able to efficiently reduce forecasting bias. We attribute this to its two-phase processing structure, where the hindcast and forecast phases are processed consecutively within the same LSTM, maintaining a single continuous state evolution from hindcast to forecast. The cell state serves as long-term memory that can selectively retain or forget information across time steps, while the hidden state captures current relevant information, enabling the model to maintain both short-term adaptations to recent forecast patterns and long-term memory of systematic biases. Unlike the Encoder- Decoder architecture transforming hindcast states through a fixed handoff network before initiating forecast processing, the Sequential LSTM maintains continuous state evolution from hindcast to forecast, preserving temporal patterns without disruption and potentially enabling better compensation for systematic errors that emerge at the forecast transition.

Given the performance degradation observed when transitioning from reanalysis to forecast data, we investigated whether training models with a combination of reanalysis and archived forecast data could improve forecast performance. In short, the integration of archived forecasts in the hindcast phase demonstrates limited effectiveness across all tested architectures.

In Fig. 5 it is evident that in transfer learning only fine-tuning the embedding net led to higher forecast skill, leading us to only focus in this experimental setup on the selective transfer learning method. As previously discussed, the TL EmbeddingNet approach pretrained exclusively on reanalysis data achieved a median NSE of 0.44. In contrast, when the same transfer learning architecture is pretrained on combined forecast and reanalysis data, the results show deteriorated performance compared to the reanalysis-only pretraining, with a median NSE dropping to 0.41. This configuration exhibits increased variability (NSE standard deviation of 2.29) while the 10th percentile performance remains poor at 0.02, failing to achieve the improvement (0.15) observed with reanalysis-only pretraining. The mean performance also deteriorates significantly (0.21 vs. 0.35 for reanalysis-only), indicating the introduction of more negative outliers.

For the Encoder–Decoder LSTM, incorporating archived forecasts yields marginal improvements in the upper performance percentiles, with the 75th and 90th percentiles increasing from 0.65 to 0.69 and 0.73 to 0.77, respectively, while the 10th percentile improves slightly from 0.21 to 0.24. However, these modest gains are accompanied by increased variability (standard deviation increases from 8.03 to 9.33) while the median NSE remains unchanged at 0.57.

The Sequential Forecast LSTM shows no meaningful benefit from archived forecast integration, with the median NSE decreasing marginally from 0.63 to 0.62. More critically, this architecture experiences an increase in variability (standard deviation from 0.52 to 6.28) and the mean performance drops from 0.57 to 0.19, indicating the introduction of numerous extreme negative outliers that significantly compromise model reliability. A systematic explanation for these outliers has not been identified, but individual inspection of specific basins suggests that residual anthropogenic influences may play a role. For example, Basin 758 shows a persistent and strong overestimation of simulated discharge relative to observations throughout the test period, which is consistent with the presence of water abstractions, retention structures, or other human interventions not represented in the model. Despite filtering the 451 basins for low anthropogenic influence based on the LamaH-CE catchment classification, some residual human influence cannot be excluded.

The results demonstrate that integrating archived forecasts during the hindcast phase does not provide meaningful performance improvements for either architecture. The approach either yields negligible benefits while increasing instability (Encoder–Decoder LSTM) or actively degrades performance (Transfer learning, Sequential LSTM), indicating that this strategy is not effective for addressing domain shift challenges in hydrological forecasting applications.

Previous studies have consistently focused on modeling ungauged basins (Kratzert et al., 2019b; Nearing et al., 2024). However, we argue that when near real-time discharge data are available, as is the case for most gauging stations across the LamaH catchments in Central Europe, the incorporation of discharge observations positively influences forecast accuracy. For example, in the Austrian LamaH basins, discharge data from the eHYD platform (https://ehyd.gv.at, last access: 19 January 2026) are available with a time delay of only 2 h, making the integration of recent discharge observations operationally feasible for real-time forecasting applications.

The integration of past discharge observations demonstrates performance improvements across both tested architectures. For the Encoder–Decoder LSTM, incorporating discharge data yields significant gains, with the median NSE increasing from 0.57 to 0.66. The improvement is particularly pronounced in the lower percentiles, with the 10th percentile rising from 0.21 to 0.37, indicating considerably better performance for poorly performing basins. The 75th and 90th percentiles also show notable improvements (0.65 to 0.77 and 0.73 to 0.83, respectively), while the variability decreases slightly (standard deviation from 8.03 to 6.68).

When discharge is integrated, the Sequential Forecast LSTM achieves a median NSE of 0.71, compared to 0.63 without discharge. The absolute improvement (+0.08) is marginally smaller than for the Encoder–Decoder LSTM (+0.09), but the Sequential Forecast LSTM reaches a higher absolute performance level, surpassing even the reanalysis baseline of 0.69. The 10th percentile shows substantial improvement (0.35 to 0.42), while the upper percentiles reach 0.81 and 0.88 for the 75th and 90th percentiles, also exceeding the reanalysis baseline performance in these ranges.

The performance gains from discharge integration are not surprising given that near-real-time discharge observations constrain the current hydrological state through autoregressive inputs (Nearing et al., 2022), providing the model with direct information about catchment conditions at the time of forecast initialization that meteorological inputs alone cannot capture. These gains therefore reflect a combination of effects: beyond reducing meteorological forecast-induced bias, improved hydrological state estimation directly contributes to forecast skill regardless of meteorological input quality.

Additional transfer learning experiments were conducted to investigate whether discharge information learned during pre-training could be effectively transferred to discharge-free operational scenarios. These experiments included domain adaption configurations where discharge was incorporated during the pre-training phase but excluded during fine-tuning, as well as setups using the Sequential Forecast LSTM with discharge in the source domain and without discharge in the target domain. All transfer learning approaches consistently resulted in performance deterioration and failed to yield improvements that would justify the computational overhead of the transfer learning process. While these experiments demonstrated that information extraction from LSTM cell states is feasible, the learned discharge-related representations could not adequately compensate for the absence of direct discharge observations during operational forecasting. The experiments with discharge incorporated directly into the training data consistently outperformed all transfer learning alternatives, indicating that real-time discharge integration provides irreplaceable benefits that cannot be effectively substituted through knowledge transfer mechanisms.

During experimenting we experienced that the embedding complexity plays an important role in reducing the forecast induced bias in runoff predictions. This circumstance has led us to create this experimental setup, in which a simple (linear) and a complex (non-linear) embedding were generated for all architectures used in the previous experimental settings. In brief, a similar pattern can be observed for all architectures except transfer learning: the more complex the input embedding, the lower the prediction performance.

While the transfer learning experiments did not achieve the high NSE values of the best-performing forecasting approach (Sequential LSTM), notable relationships emerged between embedding complexity and prediction accuracy. In the baseline, encoder–decoder and sequential LSTM experiments, simple linear embedding networks produced slightly higher simulation performance compared to more complex embedding architectures. This pattern reversed in transfer learning experiments, where increased embedding complexity led to improved prediction results. The transfer learning procedure consisted of transferring weights from an LSTM model pre-trained on all available reanalysis data (DRA) to a new network architecture with modified input embeddings, where only the embedding weights underwent retraining while the remaining LSTM parameters remained frozen.

The minimal performance improvement of simple embeddings in non-transfer learning experiments can be attributed to the tendency of complex embedding networks to overfit when trained on the available data. The number of trainable parameters in complex embeddings exceeds the optimal ratio relative to available training data, causing the model to learn specific noise patterns in the training data rather than generalizable hydrological patterns. This overfitting tendency is further amplified by the absence of dropout regularization within the embedding networks, which was set to 0 across all experiments. Additionally, the use of tanh activations in both the embedding networks and the LSTM may compound saturation effects (Acuña Espinoza et al., 2025), potentially contributing to the performance degradation observed with more complex embeddings. This means simpler embeddings capture the relevant input patterns more effectively without introducing unnecessary model complexity comprising generalization.

The improved results with increased embedding complexity in transfer learning can be explained by the requirement for flexible, non-linear transformations necessary for effective domain adaptation. Since only the embedding weights are trained while all other network parameters remain frozen, the embedding layer must perform all adaptation work between the target domain and the representations pre-trained on reanalysis data. More complex architectures can perform richer and more domain-specific feature extraction, compensating for the discrepancy between source and target domains through more expressive input transformations. The contrasting performance patterns between transfer learning and standard training suggest that the optimal embedding complexity depends on whether the model parameters are trained from scratch or adapted from pre-trained weights, though the precise mechanisms underlying this relationship warrant further investigation.

While the previous sections demonstrated the overall effectiveness of the proposed bias correction strategies, a key operational question remains: does model skill improve uniformly across catchment types, or do specific basin characteristics such as aridity, elevation, land cover, or precipitation regime determine which catchments benefit most from the proposed strategies?

Inspired by the benchmark framework of Seibert et al. (2018), who argue that model performance should always be evaluated relative to meaningful reference points, we compute the per-basin ΔNSE between each experimental configuration from Sect. 3.2 and 3.4 to the Baseline Forecast. Figure 9 shows these ΔNSE values with basins sorted by their mean ΔNSE across all experiments in ascending order, providing a clear picture of which basins consistently improve, remain unchanged, or degrade in forecast skill across the proposed strategies.

The results in Fig. 9 reveal that the Sequential Forecast LSTM with discharge integration achieves positive ΔNSE values across the vast majority, with 445 of the 451 basins, confirming that nearly all basins benefit from this configuration relative to the Baseline Forecast. A small number of isolated basins show exceptionally large positive ΔNSE values, visible as dark red columns, which are particularly pronounced in the discharge-augmented configurations and likely reflect basins where near real-time discharge observations provide especially strong anchoring of the initial hydrological state. The Baseline Reanalysis shows a very similar basin-level improvement pattern, consistent with its role as the upper performance reference.

A clear and consistent architectural difference emerges between the Sequential Forecast LSTM and the Encoder–Decoder LSTM. The Sequential Forecast LSTM without discharge integration already improves 435 out of 451 basins, while the Encoder–Decoder LSTM without discharge improves only 361 out of 451 basins and actively degrades 90, nearly 5 times as many as the Sequential Forecast LSTM. Even with discharge integration, the Encoder–Decoder LSTM degrades 28 basins compared to only 5 for the Sequential Forecast LSTM. This persistent instability is consistent with the higher standard deviations reported for the Encoder–Decoder architecture in Sect. 3.2 and suggests that the fixed handoff network introduces fragility for specific basin conditions that the Sequential LSTM's continuous state evolution avoids.

The transfer learning configurations show the most heterogeneous basin-level patterns. TL EmbeddingNet and TL AllWeights improve only 275 and 273 out of 451 basins respectively, while degrading the rest, indicating that both transfer learning strategies fail to provide consistent improvements across the basin population and cannot be considered reliable bias correction strategies across diverse catchment conditions.

Since we now know which catchments benefit most from each architecture in terms of increased forecasting skill, we further investigate which catchment characteristics determine where the largest improvements occur. To this end, we focus exclusively on basins with positive ΔNSE values (red columns in Fig. 9) – that is, basins where the proposed strategy outperforms the Baseline Forecast. Basins with negative ΔNSE are excluded from this analysis. Figure 10 shows the Spearman correlation between the per-basin ΔNSE and the 33 static catchment attributes across all experimental configurations for this subset of basins. Positive values indicate that catchments with larger values of a given attribute tend to benefit more from the proposed strategy, while negative values indicate smaller improvements.

The Spearman correlation between per-basin ΔNSE and static catchment attributes reveals a consistent and striking pattern across all experimental configurations. Aridity index (arid_1) and mean precipitation (p_mean) emerge as the dominant controls on the magnitude of improvement, with aridity showing positive correlations across all experiments (ranging from r=0.31 for TL AllWeights to r=0.50 for the Sequential Forecast LSTM with discharge) and mean precipitation showing consistently negative correlations of similar magnitude (r=-0.31 to r=-0.50). This indicates that drier catchments benefit more from the proposed bias correction strategies in absolute terms, while wetter catchments show smaller improvements over the Baseline Forecast.

When combined with the Wasserstein distance analysis of Sect. 3.1.1, these results reveal a physically meaningful asymmetry. The attributes showing larger meteorological domain shift between reanalysis and forecast inputs, including mean precipitation, mean slope, fraction of snow and aridity, are precisely the same attributes associated with smaller improvements in forecast skill across all experimental configurations. Conversely, the basins associated with smaller meteorological domain shift, characterized by higher aridity and more frequent dry days, are the same associated with larger gains in forecast skill relative to the Baseline Forecast (see Fig. 11). We explain this paradox by the nature of the hydrological signal in different catchment types: alpine, wet, and snow-dominated catchments experience the largest distributional differences between reanalysis and ECMWF-HRES inputs, but their hydrology is dominated by a strong and seasonally predictable snowmelt signal. This signal is captured and retained in the LSTM cell state through its recurrent processing of the 365 d input sequence, enabling the model to compensate for the large meteorological domain shift through its long-term memory mechanism even when forecast inputs are strongly biased. This interpretation is consistent with the findings of Kratzert et al. (2019a), who demonstrated that LSTM memory cells internally learn to represent snow storage dynamics even without snow being explicitly provided as a model input, confirming the ability of LSTMs to encode physically meaningful hydrological states in their cell states. As a result, even the Baseline Forecast achieves relatively good performance in alpine catchments, leaving less room for improvement from the proposed strategies. In contrast, drier and more episodic catchments experience smaller meteorological domain shifts but have weaker and more nonlinear rainfall-runoff relationships that are more sensitive to input quality, leading to lower Baseline Forecast performance and consequently larger absolute improvements when different model architectings or training strategies are applied.

This finding has important operational implications. While the proposed strategies consistently outperform the Baseline Forecast across the vast majority of basins as shown in the CDFs, the largest absolute gains are concentrated in arid catchments where the hydrological response is more complex and the baseline performance is lowest, rather than in the topographically complex alpine catchments where the meteorological input bias is largest.

This study is subject to several limitations. The experiments were conducted on the Extended LamaH-CE dataset, which is restricted to Central Europe; the transferability of the results to other hydroclimatic regions remains to be tested. While discharge observations were shown to substantially improve performance, they are not error-free, and their availability cannot be guaranteed in ungauged or poorly monitored catchments. The focus on a 1 d lead time restricts the conclusions that can be drawn regarding forecast quality decay with increasing lead time, and extension to multi-day-ahead forecasting is the subject of a forthcoming study.

In addition, we only evaluated existing LSTM-based architectures and a modeling setup with spatially lumped data and daily maximum discharges. Applying these approaches in fully distributed settings with higher temporal resolution forecasts may yield different, potentially more pronounced results.

The use of ECMWF-HRES as the only forecast dataset was driven by data availability – to our knowledge, ECMWF-HRES is the only NWP system for which sufficiently long archives of forecast data are openly available to cover the experimental period used in this study. From a theoretical standpoint, however, the approach is not inherently tied to any specific NWP system. A key advantage of using discharge as the target variable is that the model trains on an integrated catchment signal that implicitly captures the combined effect of all meteorological inputs, regardless of their source, meaning the training strategy can in principle be applied to any NWP system by replacing or supplementing the forecast inputs accordingly. Drawing on the findings of Kratzert et al. (2021), who demonstrated that combining multiple meteorological data sources improves LSTM simulation accuracy in a reanalysis setting, we further hypothesize that combining forecasts from multiple NWP systems could yield similar benefits in a forecasting context, enabling the model to learn source-specific bias patterns and potentially improving robustness across different forecast systems

The integration of reanalysis data in the hindcast phase, while useful for training, is also constrained by data latency, which is often critical for operational use, although future improvements in near–real-time reanalysis may alleviate this. However, this limitation depends strongly on the application scale. At national scales, real-time analysis, nowcast, or station-based observed data can serve as the hindcast forcing, bypassing reanalysis latency entirely. Products such as INCA from GeoSphere Austria (Haiden et al., 2011) are a good example of operationally available high-resolution analysis data that could fill this role. In large-sample settings such as ours, however, station-based data could introduce further practical challenges, as the number of stations per basin varies greatly, a large basin may contain dozens of stations while a small one may have only one, complicating consistent spatial aggregation and resulting in varying input dimensions across basins. While input embeddings as used in this study could potentially help map these varying input dimensions into a shared latent space, the implications for model performance and stability in such a setting remain an open question.

Addressing these challenges – testing transferability, reducing dependence on delayed or unavailable inputs, and extending both the spatial and temporal resolution of experiments – will be essential to further investigate and advance robust, operationally viable solutions for flood forecasting.