Two satellite-based observational datasets are used to evaluate mean-state fidelity of both the reanalysis dataset and the model outputs on the NES: one for sea surface temperature (SST) and the other for sea surface height (SSH). For SST, we use the NOAA Optimum Interpolation Sea Surface Temperature (OISST; ), which provides high-resolution daily fields based on blended satellite and in situ observations. For SSH, we use the AVISO (Archiving, Validation, and Interpretation of Satellite Oceanographic) gridded altimetry product archived from the data portal Copernicus Marine Environment Monitoring Service (CMEMS), which merges measurements from multiple satellite missions to generate consistent maps of sea level anomalies relative to a long-term mean. Both datasets are provided at a horizontal resolution of approximately 25 km and are used exclusively to validate the climatological mean distributions across the model domain, thereby providing a reference for the mean pattern of ocean circulation and surface water properties on the NES.

To evaluate multi-year prediction skill, we employ the GLORYS ocean reanalysis (Global Ocean Reanalysis and Simulation) version 12V1, produced by Mercator Ocean and distributed through the CMEMS . The CMEMS product is provided on a regular 1/12° (∼ 8 km in the study area) grid with 50 standard vertical levels and spans the satellite altimetry era from 1993 to the present (). It assimilates a wide range of observations, including satellite-derived SST and SSH, sea-ice concentration, and in situ profiles of temperature and salinity from ARGO floats, and ship-based measurements. Independent evaluations have shown, through comprehensive comparisons with available observations, that GLORYS is one of the best global reanalysis products in representing the climatological hydrography, circulation, and their temporal and spatial variability over the NES region . These characteristics make GLORYS an appropriate benchmark for assessing the predictive skill of both the regional and global prediction systems analyzed in this study. One limitation of GLORYS is that it does not explicitly include tidal forcing, which can contribute to residual temperature and salinity biases in strongly tidal regions such as Georges Bank and parts of the Gulf of Maine . However, in this study prediction skill is evaluated using annual-mean anomalies with a lead-time-dependent climatology removed, which reduces sensitivity to mean-state biases and high-frequency variability associated with tidal processes. In addition, GLORYS's representation of shallow coastal waters is limited, which may introduce additional uncertainties in nearshore regions.

The global simulations used in this study are based on the Community Earth System Model version 1.1 (CESM1.1), a fully coupled global climate model that integrates atmosphere, ocean, sea ice, and land components through a central flux coupler . The atmospheric component employs the Community Atmosphere Model version 5 (CAM5) at approximately 1° × 1° horizontal resolution with 30 vertical levels , while the ocean component uses the Parallel Ocean Program version 2 (POP2) at a comparable 1° × 1° resolution with 60 vertical levels . Sea ice is simulated with the Community Ice Code version 4 (CICE4) on the ocean grid , and land processes are represented by Community Land Model version 4 (CLM4) on the atmospheric grid .

To provide large-scale lateral boundary condition and surface forcing fields as well as the initial condition for the dynamic downscaling experiments, and to compare skill directly with a global prediction system, we use the CESM Decadal Prediction Large Ensemble (CESM-DPLE; ). CESM-DPLE is a global prediction system, which consists of 40 ensemble members, each initialized every 1 November from 1954 to 2015, and integrated for 10 years. The system uses CESM1.1 with nominal 1° horizontal resolution in both the ocean and atmosphere. Initialization of the ocean and sea ice is based on an ocean–sea ice-only simulation forced with Coordinated Ocean Research Experiments (CORE) bulk fluxes , while the atmosphere and land components are initialized from the CESM Large Ensemble (CESM-LE; ). The external forcing is prescribed to be identical to CESM-LE.

CESM-LE is a fully coupled companion ensemble of simulations designed to represent the response of the climate system to external forcing (e.g., greenhouse gases, aerosols, volcanic eruptions) in the absence of initialization from observations. Each member is generated by applying small perturbations to the atmospheric initial state. Because CESM-LE is uninitialized, each member exhibits year-to-year ocean internal variability such that the temporal evolution is not necessarily consistent with observations. Instead, the ensemble mean captures the externally forced component of variability. As such, CESM-LE serves two important roles in this work: (1) it provides a baseline for assessing how much of the forecast skill in CESM-DPLE and ROMS-DOWN arises from external forcing versus initialization, and (2) it illustrates the level of multi-year predictability expected solely from forced long-term trends or externally driven decadal variations.

CESM-DPLE has been widely used to assess decadal predictability of the Atlantic and global climate system, and has demonstrated robust skill in the subpolar North Atlantic . In CESM-DPLE, high-frequency atmospheric output (daily or sub-daily) is archived for only a subset of ensemble members; the implications of this constraint for the regional downscaling design are described in Sect. .

To evaluate the performance of our regional ocean model configuration, we use a hindcast simulation (ROMS-HIND). ROMS-HIND is based on the Regional Ocean Modeling System (ROMS), a free-surface hydrostatic, primitive equations numerical model . The model domain encompasses the NES, adjacent slope waters and the Gulf Stream. It extends from 30 to 55° N and from 80 to 35° W (Fig. ). The northern boundary of ROMS-HIND and ROMS-DOWN is set within the subpolar gyre so that the boundary condition from CESM-DPLE may contain highly skillful prediction . The model has a curvilinear grid with a ∼ 5 km spatial resolution and 40 terrain-following vertical levels that are unevenly spaced with a higher resolution near the surface. A minimum depth of 10 m was imposed by modifying the bathymetry prior to the simulation to avoid wetting and drying of grid cells. At the open boundaries, Chapman boundary conditions were used for the free surface, Shchepetkin lateral boundary conditions for the depth-averaged momentum, and all other fields use radiation boundary conditions .

The model bathymetry is derived from the ETOPO1 Global Relief Model (Fig. a and b). To reduce the pressure gradient errors due to the steep bathymetry, the bathymetry was smoothed with the method using a maximum roughness value of Rxo = 0.25 and Rx1 = 7. Several sensitivity experiments were conducted in which Rxo was varied, producing different degrees of bathymetry smoothing and shifts in the shelf break position relative to the coast (Fig.  in the Appendix). The configuration presented here was selected because it yielded the most realistic Gulf Stream path – including separation near Cape Hatteras – while minimizing pressure gradient errors and maintaining a realistic shelf circulation.

Surface atmospheric forcing is obtained from the European Centre for Medium-Range Weather Forecasts (ECMWF) Reanalysis v5 (ERA5) . The ERA5 analysis provides hourly estimates for a large number of atmospheric, ocean-wave, and land-surface quantities. ERA5 combines vast amounts of historical observations into global estimates using advanced modeling and data assimilation systems. ROMS-HIND is forced with hourly data including zonal and meridional wind, shortwave and longwave radiation, cloud, atmospheric pressure, rain, and humidity via COARE 3.5 bulk flux formula . The initial and open boundary ocean conditions are produced from daily 1/12° GLORYS12v1 output. The model is forced with monthly climatological river discharge and temperature at 21 river mouths along the NES, derived from the global river reanalysis dataset of . Tidal forcing is included in the model using the dominant semidiurnal tidal constituents (M2 and S2) from the TPXO09 global tidal model , providing the primary tidal mixing necessary to realistically represent the circulation over the Shelf . The simulation was performed from January 2004 to December 2010 which overlaps with the satellite and reanalysis datasets used for validation and with the ROMS-DOWN predictions described below.

To perform dynamical downscaling for multi-year prediction on the NES, we employ a regional forecast system (ROMS-DOWN) based on the same ROMS configuration used in ROMS-HIND. Unlike ROMS-HIND, which is initialized and forced solely by reanalysis products, ROMS-DOWN is initialized and forced by anomalies from CESM-DPLE. The atmospheric forcing is constructed as the 1993–2024 ERA5 monthly climatology plus daily CESM-DPLE anomaly fields, where the daily anomalies are computed relative to a daily climatology obtained by interpolating the 1993–2010 monthly climatology. Similarly, the ocean initial and lateral boundary conditions are constructed from the 1993–2021 GLORYS monthly climatology plus monthly CESM-DPLE ocean anomaly fields computed relative to the 1993–2010 monthly climatology. The different climatology periods reflect the temporal coverage of the respective datasets. Monthly ocean anomalies are used because the CESM-DPLE ocean variables required to force the regional model are available only as monthly means. This temporal resolution is also appropriate because the large-scale ocean variability relevant to multi-year prediction evolves primarily on monthly to longer timescales, while also reducing computational and storage requirements. River input is prescribed identically to ROMS-HIND using monthly climatological river transport and temperature at 21 river mouths along the NES. Because river discharge is prescribed as a repeating climatology, interannual freshwater variability is not included in ROMS-DOWN. This “anomaly downscaling” strategy reduces the impact of mean-state biases in CESM-DPLE while allowing large-scale signals to propagate into the high-resolution regional domain.

ROMS-DOWN consists of 20 start years (1992–2011), each initialized on 1 January with 10 ensemble members and integrated for 8 years, yielding a total of 200 eight-year forecasts. Our downscaling experimental design is constrained by the availability of high-frequency CESM-DPLE atmospheric output. The required daily (or 3-hourly) atmospheric forcing fields are archived only for a limited subset of CESM-DPLE ensemble members and initialization years; as a result, only 10 members provide the necessary forcing for our study period. This output constraint dictated the choice of a 10-member ROMS-DOWN. To assess whether the ensemble size of 10 members and the number of start dates of 20 are sufficient for NES multi-year prediction, sensitivity tests were conducted using the larger CESM-DPLE ensemble set (62 start years and 40 ensemble members). The results suggest that 10 members and 20 start dates are adequate to capture the ensemble-mean forecast signal over the NES region (Fig. ), indicating that the ensemble design provides a sufficient sample to characterize forecast spread and uncertainty. By construction, ROMS-DOWN preserves the global predictability inherited from CESM-DPLE while resolving smaller-scale shelf processes that are absent in the coarse global simulations (Fig. ).

To ensure consistent comparisons, all datasets are bilinearly interpolated to the ROMS grid and evaluated using a common land-sea mask. For regional skill assessment (Sect. ), metrics are aggregated over the entire NES, defined here as the combined shelf region encompassing the Gulf of Maine (GOM), Georges Bank (GB), and the Middle Atlantic Bight (MAB), as well as over each subregion individually.

We evaluate prediction performance with two deterministic metrics – the Pearson anomaly correlation coefficient (ACC) and the mean square skill score (MSSS) – applied to ensemble-mean annual anomalies, and one probabilistic metric – the Brier Score (BrS) – applied to ensemble probabilities of tercile events following . Skill is assessed using the GLORYS ocean reanalysis. To evaluate prediction skill on multi-year timescales, we focus on annual-mean anomalies which emphasizes interannual-to-decadal variability by reducing the influence of seasonal fluctuations. This choice is consistent with the intended application of CESM-DPLE as an annual-to-decadal prediction system. While seasonal analyses may provide additional insight into season-dependent sources of predictability, such analyses are beyond the scope of the present study and will be the subject of future work.

For each variable, we focus on their anomalies based on annual mean and the ensemble mean for each prediction year, which is calculated as: 1fiτ=1ne∑j=1nefijτ. where f is the annual mean anomaly, i=1,…,N indicates the model initialization years (N=20), j=1,…,ne indicates each ensemble member (ne = 10), and τ is the prediction lead year (LY1–LY8 in this work). Here, LY1–LY8 denote the first through eighth forecast years following initialization. For example, a forecast initialized on 1 January 1996 has LY1 = 1996, LY2 = 1997,…, and LY8 = 2003. Skill at each lead year is evaluated by pooling all forecasts corresponding to the same lead year across the 20 initialization years.

To evaluate prediction skill, we remove a lead-year-dependent climatology from both the model predictions and the observations. This climatology is computed separately for each prediction lead year by averaging across all initialization years. Removing a lead-year-dependent climatology accounts for systematic forecast drift that develops with increasing lead time and isolates the interannual-to-decadal anomalies used to calculate the skill metrics . Specifically, the anomaly is defined as: 2f^iτ=fiτ-f‾τ,f‾τ=1N∑i=1Nfiτ,

Analogously, we sample the observations with 20 initial years and 8 lead years to be parallel to the prediction, with the same observation values repeated for different pairs of initial and lead years, e.g., 1995 observation is used for LY1 of 1995 initialization and LY2 of 1994 initialization. For the observations, oiτ, o^iτ=oiτ-o‾τ. Because the forecasts are initialized over a finite 20-year period, the verification time windows differ among lead years. For example, LY1 spans 1992–2011, whereas LY8 spans 1999–2018. Consequently, differences in prediction skill among lead years may reflect both dependence on the forecast lead time and differences in the underlying verification periods. The implications of this sampling are discussed further in Sect. .

All deterministic skill metrics are computed from (f^iτ,o^iτ) using the N start-year pairs available at each lead years. For probabilistic metrics, we use the full ensemble to form event probabilities (Sect. ) following .

To assess how the regional and global prediction systems reproduce the mean spatial structure of key ocean variables on the NES, we show the 2004–2010 time-mean (and ensemble-mean) distributions of SST, SSH, sea surface salinity (SSS), bottom temperature (Tbottom), and 0–200 m heat content (HC200m). Observational references are provided by OISST for SST and AVISO for SSH, with GLORYS used for all variables. Note that the GLORYS are compared against OISST and AVISO, while ROMS-HIND, ROMS-DOWN, and CESM-DPLE are compared against GLORYS for the quantification of biases using RMSE (Fig. ). Spatial maps of the mean bias (model minus GLORYS) for each variable are provided in Fig.  to illustrate the spatial structure and sign of these errors. Across all variables, both ROMS-HIND and ROMS-DOWN show markedly improved mean structures and reduced mean biases compared to CESM-DPLE. For SST, CESM-DPLE exhibits a pronounced warm bias on the shelf lacking any sign of shelf break front, with a root-mean-square error (RMSE) of 3.35 °C, primarily due to the overshooting Gulf Stream separation near Cape Hatteras e.g.,, whereas ROMS simulations better capture the sharp cross-shelf temperature gradient and Gulf Stream position, with RMSE values below 1.3 °C. SSH patterns reveal that CESM-DPLE overestimates sea level gradient across the shelfbreak (RMSE = 0.22 m), while both ROMS configurations align more closely with AVISO and GLORYS (RMSE ∼ 0.08–0.09 m). Similarly, CESM-DPLE produces overly salty shelf waters (SSS RMSE = 1.89), whereas ROMS-HIND and ROMS-DOWN reproduce the fresher shelf and the offshore gradient (SSS RMSE ∼ 1.3–1.4). Improvements in the regional models are also evident in subsurface variables. For bottom temperature, CESM-DPLE lacks a realistic shelf structure due to its coarse bathymetry and simplified coastlines (Fig. ), resulting in overly warm and spatially smooth bottom fields (RMSE = 2.22 °C), while the ROMS simulations capture the shelf patterns and the transition to slope waters by more realistically reproducing dynamical constraint imposed by the bottom topography. The 0–200 m heat content (HC200m) bias is similarly reduced in the ROMS simulations (RMSE = 1.3–1.9 GJm-2) relative to CESM-DPLE (RMSE = 2.31 GJm-2).

To further quantify errors in temporal variability, the temporal RMSE is first calculated at each grid point using annual anomalies relative to GLORYS and then spatially averaged over the NES and each subregion. The resulting regional-mean RMSE values are shown in Fig. 3, while the corresponding spatial distributions of grid-point RMSE are presented in Fig. . Consistent with the spatial mean biases (Fig. ), CESM-DPLE exhibits substantially larger temporal variation errors across all variables and subregions of the NES. For SST, errors exceed 6 °C on average across the shelf, whereas ROMS-HIND and ROMS-DOWN reduce errors to less than 2 °C in all subregions. For SSH, CESM-DPLE shows RMSE values greater than 27 cm, while both ROMS configurations remain below 8 cm, after removing the spatial means – defined as the average over the analysis region shown in Fig.  – from both simulated and observed fields to eliminate reference-level differences when deriving SSH. Salinity errors are also reduced in the ROMS simulations compared to CESM-DPLE. Performance improvements in ROMS are likewise evident in subsurface variables, including bottom temperature and upper 200 m heat content, where RMSE values are reduced by roughly half compared to CESM-DPLE. One notable exception is the Gulf of Maine, where CESM-DPLE performs comparably to ROMS (Fig. d and e). This may indicate that variability in these subsurface fields during this period is partly influenced by large-scale signals, such as NAO-related changes in basin-scale or regional water properties and circulations , which the global model can capture to some degree despite the unrealistic topography due to its coarse horizontal resolution, although why CESM-DPLE performs as well as ROMS in this region is uncertain.

This model comparison shows that the high-resolution regional model reduces mean-state and variability biases in both surface and subsurface variables across the NES. These improvements are associated with a more skillful representation of shelf–slope bathymetry, slope currents, and Gulf Stream position, which are better represented in the high-resolution model framework . In addition, the anomaly-downscaling framework helps maintain a realistic mean state while reducing large-scale climatological biases present in the global system, improving the representation of regional ocean conditions. Although ROMS-HIND is expected to provide the most accurate reconstruction, the differences between ROMS-HIND and ROMS-DOWN are generally small. Occasional lower RMSE values in ROMS-DOWN likely result from its anomaly-downscaling framework, which used the GLORYS climatological mean state, together with model uncertainties and internal variability over the relatively short evaluation period. In addition, the comparison is based on only the lead-year-1 ROMS-DOWN forecasts, which are reinitialized annually, whereas ROMS-HIND is a continuously integrated simulation.

To examine the predictive skill of water properties in the NES, we first assessed the ACC for SST anomalies for lead years 1–8 (LY1–LY8) in ROMS-DOWN, CESM-DPLE, and CESM-LE. The time series for each lead year is constructed by stringing together the corresponding lead year prediction values from each start year simulation. The ensemble-mean annual SST anomaly time series averaged over the entire NES shelf (Fig. ) shows pronounced interannual and lower-frequency variability, including strong warming anomalies between the early 2000s and mid-2010s. ROMS-DOWN generally tracks the GLORYS anomalies more closely across all lead years, whereas CESM-DPLE tends to exhibit weaker amplitude and less agreement in the timing of extremes, particularly at the shorter lead times. Different from ROMS-DOWN and CESM-DPLE, the uninitialized CESM-LE captures much of the long-term warming signal but severely underestimates interannual and decadal variability, as expected by design.

Figure  shows the spatial distribution of ACC over the shelf, while Fig.  summarizes the NES-wide and subregional (GOM, GB, and MAB) mean ACC as a function of lead year, computed from spatially averaged anomaly time series. ROMS-DOWN exhibits enhanced skill generally at early lead years, with widespread positive and statistically significant correlations across the shelf in LY2–LY5. Improvements are particularly notable in the GOM and MAB, where ROMS-DOWN consistently outperforms CESM-DPLE (Fig. b and d). However, a consistent feature across both the regional and global prediction systems is the relatively weak ACC of SST at lead year 1 compared to longer forecast years. This reduced skill likely reflects a stronger interannual SST variability and the relatively smaller contribution of low-frequency signals during the forecast year-1 period (Fig. a), and may also be related to initialization error in CESM-DPLE . At longer lead times (LY6–LY8), CESM-DPLE retains greater skill, which is largely attributed to the increasing influence of externally forced low-frequency variability (Fig. f–h). The uninitialized CESM-LE also shows surprisingly high correlations across much of the shelf. This elevated skill arises because SST variability on the NES is strongly dominated by the externally forced warming trend ; thus, CESM-LE – of which the ensemble mean contains primarily the forced response and limited internal variability – can still reproduce much of the observed low-frequency signal. As a result, CESM-LE provides an estimate of the predictable externally forced component that both CESM-DPLE and ROMS-DOWN inherit at longer lead times.

In addition to ACC, we also evaluate forecast performance using the ensemble-mean mean square skill score (MSSS), which measures error reduction relative to a climatological reference. The spatial distribution (Fig. ) and regional averages (Fig. ) of MSSS for SST show patterns broadly consistent with the ACC results, with generally positive skill across much of the NES except at the earliest lead years. ROMS-DOWN exhibits modest improvements over CESM-DPLE, particularly at shorter lead times, while skill at longer leads reflects the influence of low-frequency variability. Overall, the ACC and MSSS scores suggest that much of the SST forecast skill on the NES is linked to predictable large-scale signals and long-term trends, while ROMS-DOWN improves the prediction skill by better representing shelf-scale variability and captures higher-frequency fluctuations, particularly at shorter leads.

Because the long-term warming trend dominates SST variability in the NES, both ACC and MSSS show relatively high scores in the coarse-resolution CESM-DPLE, with ROMS-DOWN providing modest improvements. However, the SST time series (Fig. ) also reveals that, even when the overall correlation or variance agreement with GLORYS is comparable, the higher resolution and improved representation of the Gulf Stream and fine-scale processes in ROMS-DOWN enable it to track warm and cool extremes more skillfully. Previous studies have shown that meridional shifts of the Gulf Stream strongly influence temperature variability over the NES, particularly in the Gulf of Maine . Therefore, the improved representation of the Gulf Stream mean position and associated shelf–slope exchange in ROMS-DOWN likely contributes to the enhanced SST prediction skill. To evaluate this event probabilistic skill, we calculate the BrS for upper- and lower-tercile SST anomalies (Fig. ) for CESM-DPLE and ROMS-DOWN. Consistent with the deterministic metrics, ROMS-DOWN generally produces lower (i.e., better) BrS than CESM-DPLE across most of the NES, indicating higher reliability in predicting probability of extreme events.

At longer lead times (LY6–LY8), BrS improves as externally forced low-frequency variability becomes the dominant source of skill – especially for upper-tercile warm SST events in both ROMS-DOWN and CESM-DPLE – but ROMS-DOWN still outperforms CESM-DPLE. Recall that, in terms of ACC and MSSS, CESM-DPLE shows comparable or even higher values on these leads (Figs.  and ), largely because it captures the low-frequency warming trend. The SST anomaly time series also highlights pronounced interannual variations during these long lead years (e.g., strong cooling around 2004 and warming around 2012) superimposed on the trend. Although CESM-DPLE retains high ACC and MSSS on these leads, its failure to reproduce these interannual fluctuations is revealed by poorer Brier Score performance for tercile events (Fig. ). Furthermore, compared to warm events, CESM-DPLE produces systematically higher (i.e., poor) BrS for lower-tercile (cold) events than for upper-tercile (warm) events (Fig. ). This asymmetry suggests that much of the apparent event-level skill in CESM-DPLE arises from its ability to capture the externally forced warming trend, which favors the prediction of warm extremes. And cold events may be more strongly influenced by regional processes such as shelf–slope exchange, episodic intrusions of subpolar waters, and wind-driven vertical mixing and entrainment associated with mixed layer deepening. Because such processes depend on realistic bathymetry, vertical resolution, and mixing parameterizations, they are better represented in the high-resolution regional model than in the coarse-resolution global system. As a result, CESM-DPLE is less able to reproduce cold extremes, leading to poorer probabilistic skill (higher BrS) for lower-tercile events.

Building on the SST prediction skill assessment, we next examine whether dynamical downscaling improves prediction skill for other key NES water properties, including SSH, SSS, Tbottom, and HC200m. Figure  shows the differences in ACC, MSSS, and BrS between ROMS-DOWN and CESM-DPLE for these variables, summarized across the entire shelf and its subregions as a function of lead year. To facilitate comparison among variables and skill metrics, Table  summarizes the mean prediction skill over the NES during the early (LY2–5) and late (LY6–8) forecast periods. Unlike the other variables, SSH exhibits generally weak prediction skill in both ROMS-DOWN and CESM-DPLE throughout the forecast period (Fig. ). Although ROMS-DOWN substantially improves the mean-state representation and temporal variability of SSH relative to CESM-DPLE (Figs.  and ), these improvements do not translate into enhanced SSH prediction skill. This limited improvement likely reflects the difficulty of accurately predicting offshore Gulf Stream path variability and associated mesoscale activities, which strongly influence SSH variability over the shelf .

For SST, the skill differences (Fig. a and e) further confirm that ROMS-DOWN's improvements are most pronounced during the first five lead years, reflecting our regional model's ability to better resolve shelf dynamics, Gulf Stream position, and mesoscale variability. Similar enhancements are evident for SSS in lead years 1–6 to a generally greater extent than SST (Fig. b and f), consistent with the pronounced interannual variability in SSS (e.g., Fig. ), which is better represented in ROMS-DOWN. The BrS difference indicates that ROMS-DOWN provides higher reliability in predicting cold and freshening events in SST and SSS (Fig. m and n), whereas CESM-DPLE consistently shows weaker skill for lower-tercile anomalies that are influenced by unresolved regional processes mentioned above in Sect. . For subsurface properties, ROMS-DOWN again improves skill relative to CESM-DPLE. For Tbottom, improvements are most evident in the lead years 1–5, when bottom-layer variability is strongly influenced by shelf bathymetry and stratification, processes that are not well represented in CESM-DPLE. For HC200m, the advantage is most pronounced in MSSS, with widespread positive differences persisting across all lead years. The better agreement of the high-resolution regional model with GLORYS may be related to its finer vertical resolution, more realistic shelf bathymetry, and improved representation of vertical mixing and entrainment processes which together likely contribute to a more accurate simulation of the vertical structure and timing of variability.

The magnitude and persistence of these skill differences also vary across subregions (Fig. ), particularly for variables and lead times where regional processes are expected to play a larger role. In the Gulf of Maine, skill improvements are generally more modest – especially for subsurface variables – likely reflecting the strong influence of upstream slope-water variability that is partly captured by CESM-DPLE. In contrast, the Middle Atlantic Bight tends to exhibit larger improvements in ROMS-DOWN for salinity and subsurface variables, especially at early lead times. This likely reflects the poorer performance of CESM-DPLE in this region, where Gulf Stream overshooting (Fig. ) and the narrow shelf bathymetry (Fig. ) are not well represented. Skill over Georges Bank is more variable and generally weaker, possibly due to strong tidal mixing and locally driven circulation variability that limit deterministic predictability and reduce the contribution of low-frequency, predictable signals.

Across metrics, the complementarity of ACC, MSSS, and BrS provides important insights into the sources of predictability. ACC reveals where models capture (or fail to capture) the phasing of anomalies. MSSS shows reductions in mean bias and error amplitude, while BrS directly evaluates the reliability of probability for discrete events. Together, these metrics suggest that while CESM-DPLE has skill at longer leads (LY6-LY8) through its representation of externally forced low-frequency variability, ROMS-DOWN tends to improve prediction skill at shorter leads and in shelf regions where regional dynamics matter most. As summarized in Table , ROMS-DOWN generally provides higher deterministic (ACC and MSSS) and probabilistic (BrS) skill during LY2–5, whereas the differences become smaller for several metrics during LY6–8. The improved representation of shelf–slope exchange, freshwater pathways, and vertical stratification in ROMS-DOWN may contribute to reduced mean-state biases, enhanced skill in bottom and upper-ocean heat content, and improved reliability for extreme events. While CESM-DPLE retains an advantage during periods when externally forced signals dominate – such as the shelf-wide warming observed over the analysis period (Fig. ) – ROMS-DOWN appears to provide greater regional skill by better representing interannual-to-decadal variability.

To further investigate the sources of improved prediction skill in ROMS-DOWN, we decompose the anomaly correlation coefficient (ACC) into contributions from the linear trend, low-frequency variability (periods > 10 years), and higher-frequency variability. This decomposition is performed by sequentially removing components from both the predicted and observed anomaly time series (Fig. ) and recomputing ACC at each step. Specifically, ACC is first calculated using the full anomaly time series. To quantify the relative contributions to total forecast skill, we use the squared ACC as a measure of explained anomaly variance. The contribution from the linear trend is estimated from the reduction in squared ACC after removing a linear trend from both the predicted and observed series. The detrended series is then high-pass filtered using a Butterworth filter with a 10-year cutoff period, applied independently to the predicted and observed lead-year time series using reflection padding to reduce edge effects. The squared ACC of the filtered series represents the contribution from variability on timescales shorter than 10 years. The remaining contribution, attributed to decadal and longer-timescale variability, is defined as the residual difference between the squared ACC of the detrended series and that of the high-pass filtered series. This decomposition provides a quantitative assessment of the relative roles of externally forced trends versus internal climate variability in shaping the overall skill.

Taking SST as an example, the decomposition of the time series (Fig. ) and the corresponding ACC contributions (Fig. ) show that a large fraction of the prediction skill in ROMS-DOWN is attributable to the linear trend, particularly at longer lead times (LY6–LY7). This suggests that much of the correlation skill derives from the externally forced warming signal, consistent with the finding that CESM-DPLE and CESM-LE also perform well when evaluated at longer leads. However, the linear trend over this short period may also reflect low-frequency internal variability rather than purely external forcing, which may partly explain the reduction of trend-related ACC at LY8 in the GB and MAB regions (Fig. h). At shorter lead times (LY1–LY5), however, interannual-to-decadal variability contribute more to SST forecast skill, suggesting that the regional downscaling adds value by better resolving internal variability driven by regional processes that are not well represented in the coarse-resolution global model. Even though the contribution from high-frequency variability is generally small compared to the trend component, it plays a non-negligible role in certain subregions. In the MAB and GB, for example, where shelf–slope exchange, Gulf Stream and mesoscale activity strongly influence shelf water conditions, high-frequency contributions exceed 20 % of the total ACC in the early lead years (Fig. q–t). This highlights the capacity of the high-resolution model to represent fine-scale dynamics that are not resolved in the coarse global system. Overall, this decomposition emphasizes that while much of the skill in SST predictions arises from predictable large-scale and trend-driven signals, the improvement of ROMS-DOWN lies in its ability to capture interannual-to-decadal variability and some higher-frequency processes that influence the shelf. It also suggests that high ACC or MSSS values do not necessarily imply that the models are skillfully capturing regional variability at all timescales, but may instead reflect dominance of low-frequency trend signals. The added value of regional downscaling may become more pronounced on shorter timescales, such as seasonal predictions, where regional dynamics and improved initialization could play a larger role.

Unlike SST, whose forecast skill primarily arises from externally forced low-frequency variability, the sources of SSS forecast skill reflect complex contributions among trend, low-frequency, and high-frequency variability. Because SSS in the NES is strongly influenced by smaller-scale shelf dynamics and regional freshwater input , salinity anomalies exhibit pronounced interannual-to-decadal variations on the shelf, with relatively smaller contributions from the long-term trend or low-frequency variability, particularly during lead years 1–4 (Figs.  and ). Both CESM-DPLE and the uninitialized CESM-LE exhibit weak SSS variability and do not reproduce the observed shelf-wide freshening and salinification events (Fig. ). In particular, CESM-LE exhibits very weak SSS variability. For example, the freshening event captured in GLORYS in 1998  – associated with advection of freshwater along the Slope Sea (not shown) – was skillfully represented only in ROMS-DOWN in all lead times including 1998, i.e., LY1-LY6. These events are absent in CESM-DPLE and CESM-LE, suggesting that although the large-scale interannual variability is present in the global model, its expression on the shelf is not well represented. The improved performance in ROMS-DOWN likely reflects its more realistic representation of shelf bathymetry and coastal processes, which allow large-scale salinity anomalies to be more effectively translated onto the shelf. Variability in Gulf Stream position and meandering modulates salinity in the Slope Sea, which subsequently influences shelf salinity through water mass exchange. Although freshwater inflow from rivers and estuaries contributes to the mean salinity structure through the prescribed climatological river forcing, the absence of interannual river variability in our experimental design indicates that the enhanced SSS skill in ROMS-DOWN primarily arises from its ability to dynamically translate basin-scale salinity anomalies onto the shelf. This mechanism is further discussed using the 1998 freshening event in Sect. . The enhanced ACC skill of SSS in ROMS-DOWN is particularly evident in the MAB (Fig. b), where the interplay between coastal freshwater plumes and slope water intrusions drives relatively strong interannual variations in salinity , consistent with the notable contribution from high-frequency variability to the overall ACC skill in the MAB (Fig. q–x). Although a larger contribution from higher-frequency variability does not necessarily imply higher overall ACC, it indicates that the model effectively represents a broader spectrum of salinity variability, including processes that the coarse-resolution global model fails to capture. The larger skill improvement in this region likely also reflects limitations in CESM-DPLE's representation of shelf bathymetry and offshore Gulf Stream structure, which reduce the fidelity of shelf salinity variability, whereas the higher-resolution regional configuration provides a more realistic representation of these processes.

Skillful multi-year prediction of physical conditions on the NES has the potential to support ecosystem-based management by providing advance information on environmental variability that influences marine habitats, productivity, and fisheries resources. Here we discuss the prediction of two key physical phenomena on the NES, MAB cold pool and the GOM slope-water mixture, because they translate basin-scale signals into shelf conditions that have direct ecological and management importance. The cold pool regulates summer bottom temperatures and stratification across major demersal fishing grounds, while the relative fractions of different slope-water sources entering the GOM through the Northeast Channel control deep hydrography and nutrient supply . Both features are tracked as ecosystem indicators in regional assessments and are central to understanding habitat for commercially important species . Evaluating prediction skill for these two process-based indicators therefore tests whether regional downscaling not only improves large-scale skill metrics but also provides information that is decision-relevant for fisheries and ecosystem management.

The improved multiyear prediction skill of SSS in ROMS-DOWN suggests that regional processes play an important role in translating large-scale salinity variability onto the NES shelf. To illustrate this mechanism, we consider the pronounced freshening event observed in 1998 (Fig. ). The spatial distribution of SSS anomalies in GLORYS (Fig. a) shows a strong negative anomaly extending along the shelf and slope, particularly in the MAB and GOM. ROMS-DOWN reproduces both the magnitude and spatial structure of this anomaly (Fig. b), whereas CESM-DPLE fails to capture this freshening event (Fig. c). This is consistent with the improved interannual prediction skill of SSS in ROMS-DOWN and suggests that the regional model better represents key processes associated with this event.

Analysis of the preconditioning leading up to the event (not shown) reveals that a fresh anomaly develops in the slope waters one to two years prior to 1998 in both GLORYS and ROMS-DOWN. This indicates a large-scale origin of the signal perhaps from outside of our ROMS-DOWN domain. Consistent with this, the fraction of LSW at the entrance of the GOM (Fig. d) as discussed in Sect.  shows a pronounced increase in both GLORYS and ROMS-DOWN in 1998, which suggests enhanced southward transport of subpolar waters (currents shown in Fig. a and b) as a key driver of the event. However, CESM-DPLE does not capture the % LSW variation (Fig. d) nor the equatorward shelf and slope currents (Fig. c), indicating that although large-scale signals may be present, their expression on the shelf is not well represented in the coarse-resolution global system. The strong relationship between % LSW and salinity anomaly in 1998 further suggests that basin-scale forcing, potentially linked to NAO variability , provides an important source of predictability for NES properties.

To assess whether the predictability of this event is attributable to large-scale anomalies advected from the subpolar gyre – which are prescribed as the northern boundary condition for ROMS-DOWN – we examine SSS anomalies over a narrow area near the northern boundary (Fig. e). ROMS-DOWN closely follows CESM-DPLE in both phase and magnitude of boundary SSS anomalies, indicating that the large-scale salinity signal is inherited from CESM-DPLE forcing by design. Despite this, CESM-DPLE fails to reproduce the corresponding shelf freshening due to lack of equatorward advection along the slope and shelf break (Fig. c). The large-scale salinity signal weakens as it propagates into the coastal region in CESM-DPLE, likely due to its limited representation of shelf–slope bathymetry and along-shelf transport. In contrast, ROMS-DOWN more effectively preserves and translates the inherited large-scale signal into a realistic regional response.

This case study indicates that the improved interannual prediction skill in ROMS-DOWN arises from a combination of inherited large-scale predictability and enhanced representation of regional processes, which together enable a more accurate prediction of basin-scale anomalies onto the NES.