Understanding potential drivers of seasonal prediction skill as well the non-stationarity behaviour of prediction skill itself over time is key to the development of a trustworthy, operational climate forecast system. That said, most prediction systems, either statistical or physical, are tuned on the climate of the last 30-40 years. Going into a new climate state, it is important to evaluate the underlying predictability assumptions over multiple climate states.

We present initial output of a data set version 1.0, which covers the years 1421-2008 C.E., has 100 members for each forecast step, covers the variables sea level pressure, 2m temperature and 500 hPa geopotential height and will be produced for the months January, February, June, July, August and December. This data set is produced using rather simple convolutional neural networks as architecture (same as in the initial WeatherBench approach) and is trained on reanalysis-infused atmosphere-ocean general circulation model data.

Exchanging parts of the model chain, such as model architecture, training data and initial conditions will allow the community to develop better and better versions of this data set eventually.

This data set and its future versions should be understood as an open-science, community-driven project. The code and output data behind this data set will be published openly. An exchange platform for interested community members will be highlighted during the presentation.

How to cite: Wegmann, M. and Brönnimann, S.: The road to 500 years of multi-member, seasonal climate hindcasts, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-5458, https://doi.org/10.5194/egusphere-egu26-5458, 2026.

Numerical seasonal forecasting consists in predicting the expected distribution of several climate variables (e.g. temperature, precipitation) over the next three months, using a global climate model that is initialized with real-time observations. Seasonal forecasts are often communicated as anomalies with reference to the model climatology estimated from forecasts initialized over a past period (hindcasts).

These anomalies are affected by long term trends due to anthropogenic climate change. Consequently, most seasonal forecasts of temperature currently issued by the Copernicus Climate change services (C3S) in the last few years indicate warmer than normal conditions over Europe, regardless of the season. 

Here, we investigate three methods to quantify the contribution of climate change from seasonal forecasts of temperature anomalies, and compare it to the usual reference based on hindcast climatology. First, we use a linear trend fitted on hindcasts. This approach is usually used in the literature to evaluate the forecast skill as it provides an estimate of the  climate change response. However, this method relies on the major assumption that the anthropogenic climate (forced) response is linear, which is not always reasonable. The second method is based on a Bayesian technique which combines CMIP6 simulations and seasonal hindcasts to estimate the forced response within the model, assuming that it is indistinguishable from the CMIP6 ensemble. The third method is based on numerical seasonal forecast experiments initialized in a so-called counterfactual world unaffected by anthropogenic forcings: dynamical initial conditions are the same as for the real, factual, seasonal forecasts, but the thermodynamic initial conditions correspond to a colder climate representative of the hindcast climatology. From this protocol, the climate change contribution can be estimated from the difference between the factual and the counterfactual forecasts. In this work, the three methods are implemented on the operational Météo-France seasonal forecast. While both the Bayesian method and numerical experiments show consistent results in the forced response estimate, results from the linear method might be inappropriate or overly simplistic in some cases.

How to cite: Ledoux--Xatard, L., Specq, D., Qasmi, S., and Giordiani, H.: Assessment of climate change contribution to seasonal forecast anomalies , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-226, https://doi.org/10.5194/egusphere-egu26-226, 2026.

Atlantic multidecadal variability (AMV) has profound climate impacts on both local and remote areas. Traditional analyses mostly concentrated on the AMV impacts on decadal-multidecadal variability. Recent studies show that AMV could also exert significant impacts on El Nino-Southern Oscillation (ENSO) and its connection with the Indian ocean dipole. However,  little attention has been paid to the AMV impacts on seasonal predictability. Based on observations and sets of ensemble hindcast products, for the first time, this study investigates the role of AMV phase on the seasonal predictability of sea surface temperature anomalies (SSTA) in North Atlantic. Our results show that the seasonal prediction skill and potential predictability of spring SSTA over the subtropical North Atlantic (STNA) region is significantly higher in AMV+ than in AMV- period. Similar contrasts between AMV phases are also obtained by the persistence skill of the observed SSTA over STNA at various lead months. Further analyses show that the differed seasonal predictability between different AMV phases are closely connected to the different upper ocean heat content, which is primarily contributed by different heat convergence driven by the Atlantic meridional overturning circulation.

How to cite: Wei, B. and Yan, X.: Seasonal predictability of North Atlantic sea surface temperature under different AMV phases, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-4613, https://doi.org/10.5194/egusphere-egu26-4613, 2026.

Reliable Arctic sea ice forecasts are important, not just for Arctic use-cases (such as determining shipping routes), but also for the potential impact that sea ice has on the midlatitude circulation. However, sea ice forecasts are often highly underdispersive, including in the IFS, the model developed and run by the European Centre for Medium-Range Weather Forecasts (ECMWF). We describe here the implementation of a stochastic parameterization scheme to the sea ice component of the IFS, and the impact it has on seasonal forecasts in the northern hemisphere midlatitudes in summer and winter. We show that sea ice ensemble spread is generally enhanced by around 10%, resulting in a more reliable forecast. We also show that the perturbations result in small but robust mean state change in Arctic air temperatures up to at least 850hPa, as a result of robust changes to the mean sea ice. A seeming consequence of this is a large increase in 500hPa geopotential (Z500) winter forecast skill over the Euro-Atlantic sector, which partially projects onto the North Atlantic Oscillation (NAO). We conclude that sea ice stochastic perturbations can be a valuable contribution to increased reliability of seasonal forecasts of the sea ice itself and can impact seasonal forecasts of the atmosphere at high and mid latitudes.

How to cite: Strommen, K., Mayer, M., Storto, A., Spaeth, J., and Tietsche, S.: How are midlatitude seasonal forecasts affected by stochastic sea ice perturbations?, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18707, https://doi.org/10.5194/egusphere-egu26-18707, 2026.

The rainfall during the austral summer season over vast regions of South America is primarily associated with tropical-extratropical cloudbands. These northwest-southeast oriented clusters of convective clouds trigger widespread rainfall and are influenced by slowly varying tropical and subtropical sea surface temperatures. Remote teleconnections also occur through atmospheric Rossby waves at synoptic to subseasonal timescales. Therefore, to accurately forecast these high impact weather events, state-of-the-art prediction systems need to capture processes at various temporal and spatial scales. An automated cloudband detection algorithm based on outgoing longwave radiation (OLR) is used in this study to examine the ability of various seasonal prediction systems, namely, ECMWF SEAS5, UKMO GLOSEAS6, and CPTEC/INPE BAM v1.2, to forecast cloudband characteristics. We find that these systems can represent cloudband seasonality and climatology well, although biases exist. There is significant spatial variability in cloudband prediction skill; the forecast systems predict monthly cloudband statistics over Southeastern South America and parts of tropical Amazon with some skill, whereas the skill is relatively poor over the core South Atlantic convergence zone region. The spatial variability in skill appears to depend on the cloudband - El Niño Southern Oscillation relationship (ENSO). Prediction skill is relatively higher in the months when ENSO has a larger influence on monthly cloudband count. In addition, the presence of skill over South Brazil possibly indicates that the models represent the underlying Rossby wave dynamics to some extent although the absence of skill over Central and Eastern Brazil potentially suggests the need for improvement in representing these teleconnections. The skill is, however, found to decrease rapidly with an increase in lead time, which might have to do with processes at shorter time scales and intrinsic atmospheric variability as suggested by previous studies. In line with this, the composite evolution of upper-level v-wind anomalies in the lead-up to cloudband events appears to be more zonally oriented in the seasonal prediction systems compared to observation. Despite being continental scale weather regimes, differences in upper-level teleconnections indicate that predicting tropical-extratropical cloudband occurrence at seasonal timescales remains a challenge, although the intense rainfall associated with cloudbands are often more predictable than extreme rainfall occurring on non-cloudband days.

How to cite: Samuel, J. B., Zilli, M. T., and Hart, N. C. G.: Tropical-extratropical cloudbands over South America in state-of-the-art seasonal forecast systems. , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19310, https://doi.org/10.5194/egusphere-egu26-19310, 2026.

Strong El Niño and La Niña events typically produce symmetric impacts on global mean surface temperature (GMST), inducing notable warming or cooling, respectively, from their developing year through the boreal summer of the following year. However, this symmetry in GMST response breaks down in the subsequent autumn and winter, and the underlying mechanism has remained unclear. This study reveals that the opposite transition behaviors of strong ENSOs are key to this breakdown: while strong El Niños commonly transition into La Niña, strong La Niñas more often persist into multi-year episodes, resulting in asymmetric climate trajectories. These divergent evolutions produce asymmetric GMST anomalies since post-summer, including not only the divergent locations and intensities of cold sea surface temperature over tropical Pacific, but also the contrasting land surface temperature dipoles over the Northern Hemisphere’s mid-to-high latitudes, mediated by tropical–extratropical teleconnections. These findings highlight a previously underappreciated source of GMST variability and offer new insight into its predictability on interannual–biennial timescales.

How to cite: Li, K.-X., Zheng, F., Yu, J.-Y., Wang, L., and Zhu, J.: Multi-year La Niñas Break the Interannual Symmetric GMST Responses to Strong ENSO Events, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3266, https://doi.org/10.5194/egusphere-egu26-3266, 2026.

Wind-driven upwelling of subsurface ocean waters to the surface is a fundamental component of ocean dynamics, and ensures nutrient-rich waters reach the epipelagic zone. Weakening or collapse of upwelling can reduce nutrient availability, potentially impacting ecosystem health and fishing activities. In early 2025, the Gulf of Panama experienced an unprecedented collapse of the local upwelling system, indicated by exceptionally weak northerly winds leading to record warm ocean temperatures and reduced nutrient availability. Despite the societal relevance of this local-scale process, the predictability of upwelling strength and in particular collapse, remains poorly understood. 

Here, we explore the predictability of upwelling in the Gulf of Panama on seasonal timescales, and find that the unprecedented collapse of 2025 was accurately predicted a season in advance. We employ the operational seasonal forecasting system CMCC-SPS4 which has a horizontal resolution of 0.25o for the ocean component, 75 vertical depth levels, and outputs 40 ensemble members. Forecasts of sea surface temperatures initialised in November and December of 2024 predicted record values for January to March 2025, indicating considerable weakening of upwelling. Validation against the OSTIA sea surface temperature dataset using hindcasts from 1993 to 2024 demonstrates high probabilistic and deterministic skill, including for predictions of upper-quintile temperature events. Moreover, by validating against the global 1/12o GLORYS12 ocean reanalysis, we also find an increase in temperature forecast skill with depth, making the case for exploiting subsurface information for improved early-warning. 

While high surface temperatures are often used as an indicator of upwelling collapse, we show that in 1998—despite strong winds and active upwelling—extreme temperatures occurred throughout the water column. These results suggest that surface temperature records alone may not fully capture changes in nutrient availability. To ensure that the forecast system captures the collapse of upwelling, we also explore the predictions of regional winds and derived upwelling indicators. 

This study demonstrates the utility of seasonal forecasting in local marine environments and makes the case for future uptake in activities related to the Blue Economy. The work also supports the definition of user-relevant indicators of extreme temperatures (Horizon Europe project “ObsSea4Clim”) and the role of reanalyses in studying subsurface temperature extremes (as part of the ocean reanalysis validation project “GLORAN”).

How to cite: McAdam, R., Sanna, A., and Scoccimarro, E.: Unprecedented suppression of local upwelling in the Gulf of Panama predicted a season in advance, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10491, https://doi.org/10.5194/egusphere-egu26-10491, 2026.

In this study, we apply the Model‑based Analog Forecast (MAF) approach to perform Indian Ocean Dipole (IOD) hindcasts using CMIP6 pre‑industrial simulations. The MAF method constructs forecast ensembles by identifying states in existing model simulations that best match an observed initial anomaly and then tracing their subsequent evolution, without requiring additional model integrations. By optimizing key parameters in the MAF framework, we demonstrate that the MAF‑based IOD hindcasts exhibit skill comparable to that of assimilation‑initialized hindcasts. Utilizing this approach, we investigate the diversity in IOD prediction skill across different climate models, with a focus on the impact of cold tongue bias on forecast performance. Our analysis reveals substantial inter‑model spread in IOD prediction skill within CMIP6 models, with useful predictability extending up to 1–4 months depending on the model. Furthermore, we identify a clear link between cold tongue bias and IOD prediction skill: models with a stronger cold tongue bias show weaker El Niño–Southern Oscillation (ENSO) teleconnections into the tropical Indian Ocean, which consequently reduces their IOD forecast capability. These results offer valuable insights into the sources of IOD prediction diversity and underscore potential pathways for improving IOD forecasting.

How to cite: Wu, Y.: Assessing the Impact of Cold Tongue Bias on IOD Predictability Using a Model-Analog Method, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6147, https://doi.org/10.5194/egusphere-egu26-6147, 2026.

The North Atlantic is a key source of seasonal-to-interannual climate predictability, as Subpolar Gyre (SPG) sea surface temperature anomalies (SSTAs), coupled with the North Atlantic Oscillation (NAO), modulate surface air temperatures over Europe and North America. However, model biases in North Atlantic dynamics and ocean–atmosphere coupling limit the skill of initialized hindcasts. While data assimilation partially constrains these errors using observations, hindcasts initialized during periods of sparse observational coverage may underestimate the true predictive potential of the system. Here, we reassess North Atlantic-driven extended seasonal predictability for the period 1960-2020 using a hybrid machine-learning (ML) assimilation approach, trained during periods with abundant observations (2004-2020) and applied to reconstruct North Atlantic Ocean temperatures during sparsely observed periods (1960-2004). Relative to standard initialization, the hybrid ML approach leads to stronger ocean–atmosphere coupling and a more robust NAO-like atmospheric response. As a result, we find enhanced winter and spring SSTA skill in the SPG during the first lead year in sparsely observed periods, along with improved surface air temperature skill over northwestern North America, southern Greenland, and central to northern Europe. Our results suggest that initialized prediction systems may systematically underestimate North Atlantic-driven predictability, and that initialization improved by hybrid ML can unlock greater forecast credibility than is implied by current standard hindcasts.

How to cite: Beobide-Arsuaga, G., Bader, J., Lentz, S., Brune, S., Kadow, C., and Baehr, J.: Underestimated Extended Seasonal Hindcast Skill in Sparsely Observed Periods Revealed Through Hybrid Machine-Learning Initialization, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9243, https://doi.org/10.5194/egusphere-egu26-9243, 2026.

Reliable seasonal forecasting of near-surface wind speeds is essential for optimizing renewable energy production, particularly in regions with expanding wind power infrastructure. Global seasonal forecast models, despite offering valuable large-scale predictability, are limited by coarse resolution (~1°), which fails to resolve local topographic, land-surface, and boundary-layer influences critical for accurate hub-height wind predictions. This study presents a high-resolution dynamical downscaling framework using the Weather Research and Forecasting (WRF) model to enhance seasonal wind speed forecasts over a target region in India. Initial intercomparison of leading global seasonal systems (ECMWF SEAS5 and NCEP CFSv2) demonstrated superior performance by ECMWF SEAS5 in reproducing observed wind climatology over the Indian subcontinent, leading to its selection as the primary driving dataset. A three-domain WRF configuration (27 km → 9 km → 3 km) was implemented, and comprehensive sensitivity experiments identified the MYNN planetary boundary layer (PBL) scheme as the optimal configuration, yielding the lowest wind speed bias and best representation of vertical wind shear.

Downscaled hindcast simulations were rigorously validated against ERA5 reanalysis across multiple vertical levels, showing substantial improvements in hub-height wind speed skill metrics. To extend forecast skill beyond the 7-month limit of available boundary conditions, a long short-term memory (LSTM) neural network was developed and trained on 40 years of ERA5 wind time series using a sliding-window approach (7-month input → 90-day output). The model was retrained for each sliding window to adapt to evolving patterns, resulting in robust predictive performance from months 8 to 10. Finally, quantile mapping bias correction was applied to the downscaled and LSTM-extended outputs compared to ERA5, resulting in an approximately 38% reduction in root mean square error and a marked improvement in probabilistic reliability. The resulting bias-corrected, high-resolution seasonal wind speed dataset provides enhanced accuracy for wind resource assessment, power production forecasting.

How to cite: Pulipati, Y., Gunthe, S. S., Chakravarthy, B., Vs, S., and Cp, A.: A Hybrid NWP–LSTM Framework for Seasonal Wind Speed Forecasting with Multi-Resolution Downscaling and Bias Correction, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-12228, https://doi.org/10.5194/egusphere-egu26-12228, 2026.

How to cite: Mignot, J., Alkama, R., Castelle, B., Couallier, J., Modou Noreyni Fal, C., Gastineau, G., Ogée, J., Provenzano, E., Raymond, T., Sakarovitch, C., Sultan, B., and Swingedouw, D.: Towards impact-ready decadal climate services: The promise of hybrid approaches, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-23065, https://doi.org/10.5194/egusphere-egu26-23065, 2026.

Renewable energy production is strongly influenced by weather and climate states, making the energy sector highly sensitive to climate variability from seasonal to decadal timescales. Decadal climate predictions, which forecast climate variability over the next 1–10 years, are therefore promising tools for optimising renewable energy deployment. For example, reliable long-term forecasts can support the identification of the most suitable locations for wind farms and solar plants, helping to stabilize energy production and reduce climate-related risks.

This study assesses the predictive skill of decadal climate predictions for energy-relevant climate impact indicators, focusing on forecast years 1-3 over Western Europe. Climate indicators are used to quantify the impact of climate variability on energy production, which is ultimately the most useful information for the energy industry.  

The calculation of the indicators requires different climate variables and temporal resolutions depending on the energy source. For solar energy, daily mean values of near-surface air temperature (TAS), surface solar radiation (RSDS), and surface wind speed (SFCWIND) are used. For wind energy, 6-hourly SFCWIND is required. The indicators are computed using a multi-model ensemble from climate forecast systems participating in the Decadal Climate Prediction Project (DCPP), which is part of the Coupled Model Intercomparison Project Phase 6 (CMIP6). To evaluate the forecast quality of the indicators, the ERA5 reanalysis is used as the reference dataset during the period 1961-2019. Skill is evaluated against ERA5 and compared with non-initialized historical forcing simulations produced with the same models to quantify the added value of decadal initialization.

Three indicators are considered: photovoltaic potential (PVpot) for solar energy, capacity factor (CF) for wind energy, and the number of effective days (Neff) for both renewable energy resources. PVpot quantifies photovoltaic performance relative to nominal capacity and is derived from RSDS, TAS, and SFCWIND. Wind CF represents the ratio between actual and maximum possible energy production and depends on SFCWIND and turbine characteristics. Neff is defined as the number of days meeting efficiency-related thresholds for each resource, based on radiation and temperature constraints for solar PV technology and wind-speed limits associated with CF ≥ 25% and turbine cut-out for wind energy. By expressing production in terms of effective days, the Neff indicator enables anticipating periods when both renewable energy resources are simultaneously scarce, as well as a consistent cross-resources comparison between them.

Results show higher and more seasonally dependent skill for PVpot than for wind CF, with Neff skill varying across regions and seasons. Decadal initialization generally enhances skill in regions where historical simulations already exhibit predictability, while limited additional skill is introduced elsewhere, suggesting that initialization primarily amplifies existing sources of predictability rather than introducing entirely new skill. These results highlight the potential of tailored climate impact indicators to bridge decadal climate prediction science and renewable-energy applications.

How to cite: Moreno Montes, S., Delgado-Torres, C., Olmo, M., Ghosh, S., Torralba, V., and Soret, A.: Decadal predictions of wind and solar power indicators to support the renewable energy sector, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9265, https://doi.org/10.5194/egusphere-egu26-9265, 2026.

The near-surface (10-m) wind speed (hereafter referred to as NSWP) is a key meteorological variable that contributes to the hydrological cycle, the transport of dust and plants, and the energy sector (e.g. wind energy). The NSWP decreased over the Northern Hemisphere (0–70°N) between 1980 and 2010. This decrease in the mean NSWP over the Northern Hemisphere is known as 'global stilling'. Using decadal predictions (DCPP-A, or Decadal Climate Prediction Project, Phase A), we demonstrate the feasibility of predicting the direction of global stilling for forecast lead times ranging from one to ten years. For example, prediction skill (quantified as the anomaly coefficient correlation, ACC) is high for the 2–5 year forecast lead time (ACC = 0.81). We demonstrate that this high prediction skill is due to the impact of changes in atmospheric greenhouse gas concentrations and anthropogenic aerosol emissions. However, the prediction of wind speed variability relative to the long-term downward trend is poor.

How to cite: Monerie, P.-A., Robson, J. I., Schiemann, R., Hutchins, B. W., and Brayshaw, D. J.: Prediction systems can forecast the direction of global stilling., EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2450, https://doi.org/10.5194/egusphere-egu26-2450, 2026.

Previous work has indicated that the subpolar North Atlantic Ocean exhibits particularly high decadal predictability, influenced by both external forcing and predictable internal variability as a result of large-scale ocean processes. The mechanism driving subpolar North Atlantic (SPNA) upper ocean heat content (UOHC) predictive skill identified in the Decadal Prediction Large Ensemble of CESM (CESM-DPLE) is linked to predictable barotropic gyre and AMOC circulations, with the ocean memory linked to the Labrador Sea Water (LSW) thickness, further corroborated by other studies. Here, we investigate whether this mechanism holds in CMIP6 decadal prediction systems with variable SPNA UOHC skill by analysing lagged regressions between initial LSW deep density and AMOC, sea-surface height, the barotropic streamfunction, deep ocean density, and UOHC. We further investigate lagged regressions between the deep ocean density in the Irminger-Iceland Basins (IIB) and these same variables to determine whether some models show a stronger connection between the SPNA UOHC and the IIB density. We have determined that models with higher SPNA UOHC skill tend to exhibit stronger correlations between the SPNA UOHC at later years and the initial LSW density (i.e., the density at the first month after initialisation). However, high model predictive skill in this initial density is not necessarily associated with higher skill in the subsequent SPNA UOHC. In higher skill models, such as CESM2-DP, CESM1-DP and HadGEM3-GC31-MM, densification in the deep Labrador Sea (1000m-2500m) is associated with a near-simultaneous increase in the AMOC strength and spin up of the subpolar gyre (SPG) as well as a subsequent warming in the subpolar North Atlantic, which later spreads to the western SPG as well. In these models, deep density anomalies accumulate between 1000m-2500m and propagate eastwards at 45°N. In low-skill models, such as CanESM5, IPSL-CM6A-LR, FGOALS-f3-L or BCC-CSM2-MR, LSW densification exhibits either no link to AMOC strength or yields only a brief period of strong AMOC, and is not associated with a persistent warming pattern in the SPNA at later years in the simulations. In these models, density anomalies at depth at 45°N appear in the initial years, but dissipate rapidly and do not propagate eastwards.

How to cite: Oldenburg, D., Yeager, S., Danabasoglu, G., Simpson, I., and Kim, W.: Mechanisms driving Subpolar North Atlantic Upper Ocean Heat Content Predictability in CMIP6 Decadal Prediction Systems, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-5956, https://doi.org/10.5194/egusphere-egu26-5956, 2026.

How to cite: Duffield, J. and Byrne, M.: Processes controlling the seasonal cycle of surface temperature: theory and idealised simulations, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2773, https://doi.org/10.5194/egusphere-egu26-2773, 2026.

Decadal climate predictions are sensitive to model initialization and simulation of climate forced response and internal variability. While analogue-based initialization selects initial states matching observations from large climate model ensemble simulations, it neglects differences in model performance. Focusing on sea-surface temperature decadal predictions, we couple analogue-based initialization with performance-based model weighting. Specifically, we favor selection of analogues from models that are statistically more consistent with observations in climate forced response and spatiotemporal variability characteristics. Through this statistical procedure, we demonstrate the effectiveness of a deviance metric that simultaneously evaluates multiple aspects of model-observation consistency and is novel to model weighting practices. We first conduct performance-weighted predictions of pseudo-observations, targeting model realizations instead of observations. Applying this exercise to more than 300 pseudo-observations to ensure robustness, we demonstrate large decadal forecast potential skill improvement compared to unweighted predictions. Second, we apply the same prediction method in decadal hindcasts of 95-year real-world sea-surface temperature observations. We find significant skill gains from performance-based weighting, however at considerably lower levels than in the pseudo-observation configuration. We explain this apparent contradiction by limited intrinsic predictability, similarity between unweighted and weighted ensembles, and inherent skill sampling uncertainties; we diagnose evidence for these three limitations in our results. Our analysis therefore highlights previously unrecognized challenges in validating performance-based model weighting, with implications for model weighting practices for climate predictions and projections across time scales.

How to cite: Verjans, V., Donat, M., Delgado Torres, C., and DelSole, T.: Large potential of performance-based model weighting to improve decadal climate forecast skill, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3095, https://doi.org/10.5194/egusphere-egu26-3095, 2026.

Reliable climate information at multi-year lead times is essential for informed decision-making and long-term planning. Such information helps policymakers and stakeholders prepare for climate-related risks and build resilience to ongoing climate variability and change.

Decadal climate variability (DCV) affects regional climate patterns all over the world on timescales of several years to decades. Skillful prediction of these modes and their impacts can support planning several years in advance. The Tropical Atlantic SST Gradient (TAG) index is one such DCV mode, characterized by differences in sea surface temperature across the tropical Atlantic Ocean. Variations in TAG strongly affect rainfall patterns, circulation, and climate extremes in surrounding regions, including parts of Africa and South America, with important socio-economic consequences. The Decadal Climate Prediction Project (DCPP), conducted under CMIP6, provides coordinated decadal hindcast and forecast experiments to study and predict such variability.

However, traditional statistical approaches often struggle to represent the complex, non-linear, and non-stationary nature of DCV modes like TAG. Deep learning (DL) methods offer a promising alternative, as they are well suited to capturing both long-term trends and shorter-term fluctuations, as well as changes in the phase of variability.

In this study, we aim to strengthen the prediction skill of the CMIP6 multi-model ensemble (MME) TAG index for lead years 1–10 using DL-based post-processing. We apply a recurrent neural network (LSTM) to correct the raw CMIP6 MME TAG forecasts. Our results indicate that DL methods have strong potential to enhance the prediction of TAG variability, particularly in terms of its trend and phase. These findings suggest that DL can serve as a valuable complementary tool to existing dynamical models, improving real-time decadal predictions and increasing confidence in operational climate forecasting systems.

How to cite: Dixit, J., Kodamana, H., Sandeep, S., and AchutaRao, K. M.: Statistical improvement of TAG Index Prediction Skill in DCPP-A Hindcast Experiments Using Deep Learning, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-21421, https://doi.org/10.5194/egusphere-egu26-21421, 2026.

Hot-dry compound extremes have recently gained attention as a result of their potential destructive impacts on environments and societies. To this end, multi-annual predictions of these events could potentially offer useful information for a variety of socio-economic sectors. However, while previous studies have successfully predicted these extremes in some regions, they still struggle to capture much of the interannual variability, with most skill stemming from long-term forcings. Here, we investigate the sources of such limitations by comparing the skill of multi-annual forecasts against a perfect-model setup, using the EC-Earth3 model. While real-world predictions are initialized towards the observed state and evaluated in their ability to predict observed climate, the perfect-model predictions are initialised and assessed against a historical simulation with the same model, ensuring physical consistency between the prediction and the reference, and avoiding the uncertainties tied to the initial conditions. By comparing the perfect-model setup (PerfSet) with the real-world setup (RealFor), we assess to what extent the inconsistencies between real-world climate and the model affect the multi-annual predictability of compound hot-dry extremes.

From a skill perspective, the relative performance of PerfSet and RealFor depends on the region analysed, with neither experiment consistently outperforming the other. Residual correlation analysis, representing the contribution of initialization to forecast skill, indicates that PerfSet generally exhibits larger areas with statistically significant correlations. These regions broadly coincide with areas where PerfSet shows higher skill, suggesting a stronger influence of initialization in this experiment. Further analyses distinguish dry conditions as a key limit to predictability for both experiments, particularly where aridity is mainly dependent on precipitation variability rather than potential evapotranspiration. These results illustrate the inherent limitations of models for multi-annual predictions and highlight how the intrinsically low predictability of precipitation constrains the predictability limits for hot-dry compound extremes, whether predicting real-world observations or a controlled reference dataset.

How to cite: Aranyossy, A., De Luca, P., Mahmood, R., and Donat, M.: Exploring the limits of multi-annual predictability for compound hot-dry extremes, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19534, https://doi.org/10.5194/egusphere-egu26-19534, 2026.