The Météo France seasonal prediction system SYS8 MF system 8; is based on the high-resolution version of the coupled CNRM‐CM6-1 global climate model used for CMIP6 . It contributes to the seasonal forecast component of the Copernicus Climate Change Services (C3S).

The streamflow forecast is derived from the interaction between the atmosphere component, ARPEGE-Climat 6.3 ; the land surface component (ISBA), which simulates the runoff; and the advanced river routing (CTRIP), which simulates the streamflow river discharges . In ISBA, the soil is discretized in 14 vertical layers, accounting for the soil hydraulic and thermal properties, while the multi-layer snow model simulates water and energy budgets separately in the soil and the snowpack. ISBA in one grid cell is tiled into 12 patches of soil and vegetation, which aggregates 500 land cover units at 1 km resolution present in the ECOCLIMAP-II database , where mean seasonal cycles of snow-free albedo and leaf area index are prescribed from Moderate Resolution Imaging Spectroradiometer (MODIS) products at 1 km spatial resolution and the normalized difference vegetation index product from the SPOT/Vegetation. The soil textural properties (clay, sand and soil organic carbon content) are given by the Harmonized World Soil Database (http://webarchive.iiasa.ac.at/Research/LUC/External-World-soil-database/HTML/, last access: 17 September 2025) at a 1 km resolution. Topography is derived from the 1 km Global Multi-resolution Terrain Elevation Data 2010 (https://topotools.cr.usgs.gov/gmted_viewer/index.html, last access: 17 September 2025). Heterogeneities in precipitation, soil infiltration capacity, topography and vegetation are considered through a comprehensive sub-grid hydrology scheme . In CTRIP, the result of rainfall excess, effective river aquifer exchange, open-water evaporation and inflow from the upstream cell is routed by a river model in which the streamflow velocity is solved dynamically via Manning’s formula (kinematic approach) and assuming a rectangular river cross-section in a grid resolution of 0.5o .

CNRM‐CM6-1 incorporates an explicit two-way coupling between ISBA and CTRIP via the SURFEX and OASIS-MCT interface . The coupling allows us to consider (i) the dynamic river flooding in which floodplains interact with the soil and the atmosphere through infiltration, open-water evaporation and precipitation interception and (ii) the fact that a two-dimensional diffusive groundwater scheme represents unconfined aquifers and upward capillarity fluxes into the superficial soil . The latter contributes to capturing active groundwater–river connections that are crucial in representing groundwater-sustained baseflow during dry seasons . More details on the model parameterization and structure can be found in and .

Most previous works evaluate the “potential” streamflow predictability of a forecast system by adopting the perfect-model assumption, in which the streamflow forecast is compared to simulated streamflow (from a model driven by meteorological observations) instead of observed streamflow. Here, we compare the forecasts against observations because, in addition to the IHC and FSC, these incorporate the uncertainty associated with model error (due to structure, physics, and parameter uncertainty) and provides actual (as opposed to potential) streamflow predictability, which is more valuable for end-users or the development of climate services. A database of 1755 flow gauge stations has been created, compiling the global streamflow open-access datasets presented in Table . We have filtered the full dataset to remove stations with relatively small drainage areas poorly represented by the model resolution and those stations with more than 25% of missing streamflow records in the season of concern.

We conducted a correlation analysis to select the minimum drainage area considered in the study. For basins with an area higher than a certain threshold (Athreshold), Fig.  shows the correlation between the basin area (Abasin) and the area estimated for the CTRIP routing model (ACTRIP). With increasing Athreshold, the correlation increases, but the number of available basins is reduced. The threshold is set to 6×103 km2 (about two CTRIP cells per basin in mid-latitudes) to maintain a balance between the number of basins analysed and their geometrical representation and to avoid considering basins inside the spurious oscillating correlation curve (Fig. ). There are 1451 gauged basins with Abasin≥Athreshold, with an Abasin|ACTRIP correlation of 0.9886.

From the 1451 streamflow stations, we only consider those with less missing data than 25% of the total data in the analysed season. Figure S1 in the Supplement shows the distribution, in terms of space and frequency, of the full database and the selected stations. The final dataset has 1071 stations in JJA and 1043 stations in DJF, distributed across North America (≈82%), Europe (≈13%), South America (3.5%), Africa (1.7%), Asia and Australia (0.4 % = 4 stations). In Sect. , we remove 14 stations where the mean bias magnitude exceeded the maximum machine number in double precision. This can occur in basin outlets, where backflow or strong regulation can lead to a near-zero mean or standard deviation of observed streamflow (in the denominator of bias, the simulated-to-observed mean and standard deviation ratios, as shown in Table ).

Typically, statistical post-processing methods are applied to compensate for errors in model structure or initial conditions, correct biases, and improve ensemble dispersion . Such bias correction can be applied to atmospheric forecasts (such as precipitation, temperature and evaporation) and/or to hydrological forecasts like runoff and streamflow e.g.. Our study uses an online atmosphere–ocean–land–river coupled model, for which bias correcting the atmospheric forcing is irrelevant. Instead, we correct the streamflow forecast bias for each flow gauge station using the empirical quantile mapping (EQM) method. To ensure consistent comparisons, we apply streamflow bias correction to both offline and online forecasts.

Unlike adjusting parametric distributions, the EQM method removes bias using empirical cumulative distribution functions (ECDFs) from observations and forecast percentiles. Roughly, the approach replaces the forecast values with observed values corresponding to the same non-exceedance probability (i.e. it calibrates the forecast distribution with the observed distribution by fitting the forecast values). Analogous to , the bias-corrected streamflow Qc is calculated as follows: 1Qc=Fo-1Ff(Qf), where Ff and Fo are the ECDFs of forecast Qf and observation streamflow Qo, respectively.

Table  presents the deterministic and probabilistic scores used to evaluate the new forecast system performance. The thresholds for the Brier score computation are based on the 3-month average of observed streamflow exceeded 66 % (the lower tercile Q₆₆), 95 % (Q₉₅) and 10 % (Q₁₀) of the time. These thresholds characterize low, very low and high flows . The skill of the online approach is relative to the performance of the Offline_ICL benchmark.

The significance of the precipitation correlation is calculated using the parametric Student's t test. All other significance tests and confidence interval computations use the bootstrap approach, where 1000 random sub-samples are created from the full sample to establish the probability distribution of the statistical estimator being analysed (e.g. the anomaly correlation coefficient or the Kling–Gupta efficiency (KGE) score). An estimator is considered to be significant if the p value is less than or equal to 0.05.

We assess the global performance of the river streamflow simulated by the initialization runs (ICL and ICL_nud) against historical streamflow observations. For this purpose, we compare the initial-month mean streamflow (May for JJA and November for DJF) against the observed streamflow over the 1993–2017 period.

Figure  presents three performance metrics of the comparison (bias, root mean square error and anomaly correlation coefficient). Note that only stations with less than 25 % of missing data during the corresponding month are considered in the following analysis.

For May, the streamflow bias of the ICL tends to be positive in the driest regions (Fig. a), particularly in the west of North America, northeastern Brazil, southern Africa, the Iberian Peninsula and Australia. The higher concentration of red markers in Fig. b suggests a reduction in bias from the ICL to ICL_nud. This reduction is more pronounced for negative bias, as indicated by the shift of the negative peak towards zero bias in the frequency distribution shown in Fig. c. Besides this, the RMSE is generally smaller with ICL_nud, particularly over regions with large RMSEs in the ICL (Fig. d–f). In seasonal forecasts, the temporal correlation between the forecasted and the observed anomalies is crucial since it indicates the capability of capturing the interannual variability of streamflow departures from the mean value. The spatial distribution of the difference in the anomaly correlation coefficient |ACCICL-1|-|ACCICLnud-1| in Fig. h shows that the soil moisture nudging improves the temporal dynamics of the simulated streamflow in May over most of the 1067 gauging stations. The result is verified in Fig. i, which reports up to 20 % more stations with ACC>(0.4-0.6).

The performance of the river initialization in November (used for DJF forecasts) is presented in Fig. S3. ICL_nud tends to reduce the mean bias of stations displaying a high positive bias in the ICL (Fig. S3a–b). The global distribution of bias in Fig. S3c confirms a reduction in high positive bias, favouring the concentration of bias values closer to zero than in the ICL. However, unlike in JJA, in DJF, ICL_nud enhances the number of basins with higher RMSE and lower ACC. In Sect. , we show and discuss the impact of the initial hydrologic condition (IHC) degradation on the hindcasts in boreal winter.

One way to bring out the influence of the land–atmosphere coupling is to assess the impact of different land IHCs on the atmospheric forecast. The performance of the atmospheric seasonal forecast is presented in Figs.  and , particularly for two of the most important water cycle drivers: precipitation and near-surface temperature. Precipitation is compared against the Multi-Source Weighted-Ensemble Precipitation (MSWEP v2, ), and the temperature is compared against the Climatic Research Unit gridded Time Series (CRU TS v4.05, ).

A global view does not reveal marked changes in terms of ACC for the atmospheric predictions. However, from a continental to regional scale, differences are noticeable. In boreal summer (Fig. ), enhanced initialization ICL_nud tends to increase precipitation correlation in the middle region of South America, including the Paraná River basin and the southern Amazon basin (red box), with degradation in the northeast of Brazil, Australia, and some areas of North America and Asia north of 40 ° N (cyan boxes). Notably, Europe experiences improved precipitation predictions. Temperature predictions are less sensitive to the land initialization in summer, but degradation is concentrated in higher latitudes (north of 40° N and south of 20° S). In winter, regions with reduced performance for both precipitation and temperature predictions are primarily found in North Africa, Europe and Asia (Fig. ).

We have found that the ICL_nud initialization can have a detrimental effect on the accuracy of precipitation and temperature seasonal forecasts. This is due to soil moisture nudging, a technique intended to enhance the variability of soil water content and to improve the forecast of river streamflows. However, it can also lead to adverse effects on the land–atmosphere coupling simulated by the model. The initial soil moisture conditions introduced by the offline nudging technique may shift the coupled system away from its equilibrium state. When the forecast integration begins, the nudging constraint is deactivated, and the model is adjusted to its equilibrium, potentially generating misleading heat and water fluxes at the land–atmosphere interface. This could ultimately disrupt the atmospheric circulation and reduce the accuracy of the temperature and precipitation forecasts.

We have shown evidence of the impact of land IHC on the performance of seasonal atmospheric forecasts as proof of the importance of land–atmosphere feedback. In the following section, we will explore the sensitivity of the SSF to enhanced IHCs in a fully coupled global forecast system.

The hindcast performance of the ESP benchmark (Offline_ICL) is compared against the hindcasts of the fully coupled configurations with two different land initializations (Online_ICL and Online_ICL_nud) to determine the contributions of initialization and land–atmosphere coupling. Unlike online configurations, where the model forecasts the atmosphere, in Offline_ICL, the atmosphere forcing is based on climatology without land–atmosphere feedback. More details on the three configurations can be found in Sect. .

The enhanced land–river initialization was designed to capture the spatial and temporal dynamics of soil moisture and thus improve water storage variation. For this purpose, soil moisture is nudged to reconstructed fields of a reanalysis based on a surface model driven by ERA5 atmospheric forcing.

In boreal summer, enhanced land initialization is more critical than using a fully coupled GCM-derived forecast system since only the former approach led to improved forecasts. For this season, the highest impact of initialization on forecast skill occurs in large basins of the driest regions, where streamflow is strongly sustained by the water storage naturally released during low flows (i.e. baseflow).

In boreal winter, the new IHC negatively affected basins at high latitudes, probably due to the potential disruption of the energy, ice and liquid water budget in the soil induced by the lack of nudging of soil temperature e.g., which could lead to spurious model adjustment through excessive or reduced runoff. Confirming this would deserve a dedicated evaluation beyond the scope of this study. In South America, the coupling is beneficial (for both IHCs) in JJA and DJF, which is consistent and directly related to the strong ACC of precipitation and temperature provided by the online systems in most of the basins analysed in South America (see Figs.  and ). For Africa, no robust conclusions can be drawn from the results due to the reduced sample of stations. Still, predictions from all model configurations in DJF were poor, with fewer than 40 % of stations exhibiting a positive KGE.