The regional downscaled model developed here inherits most of the configuration and parameters from the MOM6-COBALT-NWA12 v1.0 model developed by Ross et al. (2023) and further refined for applications to decadal predictions by Koul et al. (2024). The model comprises coupled components for ocean physics, biogeochemistry, and sea-ice. Compared to the previously published decadal predictions, we add the COBALT biogeochemical model (Stock et al., 2020, 2025) and improve the initial physical state estimates by relaxing the downscaled simulation towards the EN4 dataset (Good et al., 2013) at depth greater than 1000 m.

We use a 10-member SPEAR-based global simulation (SPEAR_LO; Delworth et al., 2020; Yang et al., 2021) with a low-resolution configuration (1° ocean and sea ice with near-equatorial refinement; and 1° atmosphere and land) to obtain surface forcing and open boundary conditions for the regional model. We use forcing and boundary conditions from the SPEAR_LO global coupled reanalysis to create the NWA12 simulations used to initialize our hindcasts (i.e., NWA12_EN4_NUDG, Table 1), and forcing from SPEAR hindcasts (SPEAR_HIND) to drive NWA12_HIND. For the SPEAR_LO reanalysis, atmospheric winds and temperature were restored toward the 6-hourly Japanese Reanalysis (JRA3Q; Kosaka et al., 2024), and the oceanic SST is restored toward the interpolated monthly-mean observations from the NOAA Extended Reconstructed Sea Surface Temperature version 5 (ERSSTv5) dataset (Huang et al., 2017). The reanalysis ensemble spread originates through alternative deep ocean initializations and is maintained due to the gentle surface-only assimilation approach used by SPEAR_LO (Yang et al., 2021).

SPEAR_LO reanalysis output was used to force the regional MOM6 model at the ocean surface and at the three open ocean boundaries (see Koul et al., 2024) from January 1960 to December 2024. The 10 members of this regional downscaled simulation (NWA12_EN4_NUDG) are derived from the 10 members of the SPEAR_LO: we simply carry forward the member-wise information present in the ocean initial conditions, atmospheric forcing and open ocean boundary conditions.

The interior solution of the NWA12_EN4_NUDG simulation was furthermore relaxed with a 90 d timescale towards the full vertical profiles of ocean temperature and salinity from the EN4 gridded dataset in regions where the ocean bottom is deeper than 1000 m. This moderate relaxation helps keep the model solution closer to the observation-based sub-surface variability, particularly in the late 1990s when the boundary conditions overestimated the propagation of a cool, fresh signal into the domain. Limiting the relaxation to waters deeper than 1000 m helps avoid rapid oceanic re-adjustment in shallower waters upon initialization of retrospective predictions wherein such relaxation towards EN4 is not carried out and is necessary to avoid nudging on the shelf where the 1° EN4 shows low correlation with observed salinity despite maintaining high correlation with observed temperature (Figs. S2 and S3). We also note that the bottom waters of the Gulf of Maine analysed in this study (the deep basins and the Northeast Channel) lie well inshore of the 1000 m isobath and are therefore not directly nudged towards EN4. The relaxation constrains only the off-shelf slope source waters. We do not relax the regional model towards the parent SPEAR model, since SPEAR has coarse ocean resolution and does not assimilate sub-surface ocean data, making it less constrained towards observations than the EN4-gridded dataset.

Open boundary data for biogeochemical tracers were obtained exactly as explained in Ross et al. (2023). The time varying boundary data for alkalinity and dissolved inorganic carbon were estimated from SPEAR_LO monthly average temperature and salinity as input to the algorithm from Carter et al. (2021). Climatological river freshwater discharge, nutrient, carbon and alkalinity data were also used as described in Ross et al. (2023) and Koul et al. (2024). This climatological forcing, which includes seasonal variability but not interannual variability, was used because coastwide river observations dating back to the 1960s are not available. In the Gulf of Maine, direct river runoff is estimated to be a minor source of freshwater and nutrients (Smith et al., 2001; Townsend, 1998) and most of the interannual variability is driven by changing sources of inflowing water.

Following Koul et al. (2024), we use the regional simulation from the NWA12_EN4_NUDG model to start 10-yearlong retrospective ocean physical-biogeochemical predictions (NWA12_HIND) every year on 1 January from 1965 to 2025. The last 10-yearlong simulation, which shifts from hindcast to forecast, is from 1 January 2025 to 31 December 2034. The 10-member NWA12_HIND ensemble are started from ocean initial conditions taken from the corresponding 10 members of NWA12_EN4_NUDG, and integrated forward using atmospheric forcing and open ocean boundary conditions from the 10 members of SPEAR_HIND. To ease the transition between the global and regional model of the complex inflow and outflow along the northern boundary, each member of the 10-yearlong predictions is relaxed towards the corresponding SPEAR_LO prediction north of 45° N. The strength of the restoration corresponds to a relaxation time scale of 90 d at the northern boundary and decreases gradually to 0 at 45° N. No bathymetric mask is applied, as the relaxation drops to zero north of the targeted shallower regions. The latitudinal damping rate r (inverse of relaxation time scale) was prescribed as: 1r(ϕ)=rmax/2×1+cos⁡πϕmax-ϕ/ϕmax-45 for ϕ (latitude) from 45° N to ϕmax in the regional model, and zero south of 45° N and rmax=1/90 d⁻¹ at ϕmax, where ϕmax in the regional model is ≈58° N.

The NWA12_HIST is a continuous 75-yearlong simulation started on 1 January 1960 and ending in 31 December 2034. The NWA12_HIST is also a 10-member ensemble forced with atmospheric forcing and open ocean boundary conditions from the 10 members of SPEAR_HIST. The SPEAR_HIST dataset spans from 1850 to 2100 and incorporates future greenhouse gas concentrations based on the SSP585 emissions scenario (Riahi et al., 2017). It should be noted that differences between SSP585 and other emissions scenarios are negligible over the near future (next 1–10 years) examined in this study. Both NWA12_HIST and NWA12_HIND share the response of the model to the common prescribed natural and anthropogenic radiative forcing but the phase of natural variability in NWA12_HIST is not expected to match the observed phase since the impact of 1960 initial conditions is not retained beyond a few years.

Two observation-based gridded data sets were used to assess the quality of model simulations for bottom temperature and salinity: the 1/12° GLORYS12 v1 ocean reanalysis and interim analysis products (Lellouche et al., 2018) and the 1° EN4 gridded dataset (EN.4.2.2, Good et al., 2013) which is used to assess a longer period from 1965–2024 (Fig. S1 in the Supplement, Figs. 3 and 5). For bottom O₂ and bottom Ωar, we used the Fishing Industry SHared Bottom Oceanographic Timeseries data (https://erddap.ondeckdata.com/erddap/tabledap/fishbot_realtime.html, last access: 29 May 2025), hereafter called “Fishbot”, and vessel (Jiang et al., 2021) and glider-based (IOOS Glider Data Assembly Center: https://gliders.ioos.us/erddap/index.html, last access: 29 May 2025) datasets (hereafter called “NOAA OBS”). In addition to these two datasets, we also used the observed bottom temperature and bottom salinity time series (hereafter called “NOAA_SOE”) for the GOM Ecological Production Unit (EPU) from NOAA's State of the Ecosystem reports (NOAA Fisheries, 2025).

Area-weighted averages were first computed for the model-simulated ensemble mean and the two gridded observational datasets (GLORYS12 and EN4), excluding grid cells with ocean depths less than 30 m. Annual anomalies were then derived by subtracting the 1994–2023 mean from the annual mean time series. Figures S1–S3 show the comparison of annual anomalies from variables across these gridded datasets.

For all Fishbot and NOAA_OBS observations, point data was first linearly interpolated onto a uniform 1/4°×1/4° grid before area averaging. Monthly anomalies were then computed by subtracting the long-term monthly mean (1994–2023) from each month, provided there were at least three years of data samples. Finally, annual mean anomalies were calculated by averaging monthly anomalies for years with at least four months of data. The uncertainty (error bars in Fig. 2) for that year is σ_monthly/n, i.e. the sample standard deviation of that year's monthly anomalies divided by the square root of the number of valid months (n).

The variability of the Gulf Stream's position was assessed using a method established by Pérez-Hernández and Joyce (2014) and subsequently utilized by Ross et al. (2023). First, the Gulf Stream's mean position is defined by the latitude where sea surface height (SSH) variance is highest along meridional lines, spaced 1° apart, from 72 to 52° W. Then, to quantify the relative position of the Gulf Stream, the Gulf Stream Index (GSI) was calculated as the annual SSH anomaly, relative to the 1994–2023 average, averaged over these mean position points. The proportion of the two water masses, Labrador Slope Water and Warm Slope Water, in the Northeast Channel (Fig. 1) analyzed in this study were computed from water mass definitions as described in Ross et al. (2023).

We used anomaly correlation coefficients (ACC) to quantify the skill in predicting annual mean anomalies of bottom temperature, bottom O₂, bottom pH and bottom Ωar. The ACC for area-averaged quantities was calculated for the Gulf of Maine Ecological Production Units (GOM_EPU, NOAA Fisheries, 2024). For bottom temperature, bottom O₂, bottom pH and bottom Ωar the NWA12_EN4_NUDG simulation was used as the observational reference, and the skill was calculated over the period 1965–2024. The lead-dependent climatology (1994–2023 mean) was removed from the initialized hindcasts before calculating the skill metric. This choice of using NWA12_EN4_NUDG as the observational reference was made because continuous spatio-temporal observational records of bottom temperature, O₂, pH and bottom Ωar, for any of the EPUs or Large Marine Ecosystems (LMEs) were not available. However, to show that the NWA12_EN4_NUDG was a good proxy for in-situ observations for the skill estimation, we carried out a point-by-point comparison of temporally and spatially co-located model output and available observations (Fig. S7), as well as additional evaluation of interannual variations and covariance with better-constrained physical properties (Figs. S1, 3 and 5). A simple persistence-based skill, defined as the lagged autocorrelation of observed anomalies, was also compared with the hindcast skill.

We assessed the statistical significance of the prediction skill from the initialized NWA12_HIND retrospective predictions. This was done by conducting a two-tailed t-test against a null hypothesis of no correlation, adjusting for the reduced degrees of freedom due to autocorrelation. Additionally, we compared this skill to that of persistence forecasts and the uninitialized (NWA12_HIST) simulation. For this comparison, we employed the Williams (1959) method as described in Steiger (1980) to determine the statistical significance between two dependent correlation coefficients.

Probabilistic tercile predictions are developed by post-processing the model forecasts using extended logistic regression (Wilks, 2009). Extended logistic regression fits the probability of the outcome y exceeding a certain threshold q given some data x, and is formulated as 2P{y>q|x}=σ(f(x)+g(q)) where the logistic function σ(z)=1/(1+e-z)=ez/(1+ez). Here, f is a simple linear function of the ensemble mean prediction x: f(x)=β0+β1⋅x, and the extension function g is also a simple function of the threshold q: g(q)=β2⋅q. Maximum likelihood estimation, following Minka (2001) and Pregibon (1981), is used to fit the three beta parameters for each of the four forecast variables in Fig. 7. In this method, the forecast probability is a function of only the ensemble mean, and not the individual ensemble members or the ensemble spread. As is common in seasonal to decadal dynamical model predictions, we find that the raw ensembles tend to be overconfident, with insufficient ensemble spread (Fig. S8). Furthermore, the ensemble spread has low correlation with the actual retrospective forecast error. As a result, extended logistic regression provides a reasonable improvement to the estimates of the forecast uncertainty.

For evaluating probabilistic predictions, we use accuracy as the verification metric for each tercile. The accuracy is calculated as: 3Accuracy=(TruePositives+TrueNegatives)/TotalForecasts. This metric is computed independently for each tercile category. In this context, a true positive occurs when the tercile with the highest predicted probability correctly matches the observed tercile. A true negative occurs when the model correctly assigns the highest probability to either of the other two terciles (effectively predicting that observations will not fall within the tercile being evaluated).

We also calculated the ranked probability skill score (RPSS) for each variable to assess the accuracy of the predictions across all three categories. Persistence of the previous year's category was used as the reference for the RPSS, so positive skill scores indicate that the model forecasts have some ability to predict transitions away from recent conditions.

The RPSS is a skill score based on the Ranked Probability Score (RPS), and the Ranked Probability Score (RPS) for a single forecast is defined as 4RPS=1K-1∑k=1K∑i=1kPi-∑i=1kOi2 where K is the number of categories (3 in our case: below, normal, or above), Pi is the cumulative forecast probability for category i, and Oi is the cumulative observed probability (0 or 1) for category i. A perfect forecast, which assigns 100 % probability to the observed category, will have an RPS = 0. Predicting a 100 % chance of the opposite outcome (i.e., below when the outcome is above) gives an RPS of 1, while predicting normal when an outcome is above gives an RPS of 0.5. The RPS is notable in that it evaluates the distribution of the forecast probability around the outcome, penalizing outcomes that are further away. This makes RPS a more sensitive metric than accuracy.

The ranked probability skill score (RPSS): 5RPSS=1-RPS(forecast)/RPS(persistence) where RPS (forecast) is the sum of the ranked probability scores comparing the forecast probabilities with the observed categories, and RPS (persistence) is the sum of the ranked probability scores comparing the observed categories with the categories observed in the previous year (in other words, we assign 100 % probability to the category observed in the year before the forecast was initialized, and compare that with the actual category for the next 3 years). By the definition of the skill score, if RPSS = 1, the forecasts had perfect accuracy at probabilistically predicting the observations. If RPSS = 0, the forecasts had the same overall score as the persistence forecasts of the previous year's state; in other words, the forecasts would have had no skill at predicting transitions.

Bottom O₂ saturation ([O₂] sat), was computed from the simulated bottom potential temperature and salinity following Garcia and Gordon (1992). DIC and alkalinity were not saved at the ocean bottom and were thus obtained by solving the carbonate system (PyCO2SYS; Humphreys et al., 2022) from the model's bottom pH and Ωar together with bottom temperature and salinity at a representative pressure of 150 dbar. The result is largely insensitive to this pressure choice – recomputing Ωar from the recovered DIC and alkalinity reproduces the simulated Ωar to within 0.01. COBALT additionally applies minor nutrient alkalinity corrections (PO₄ and SiO₄) that are omitted here, with negligible effect on the bottom anomalies.

The contributions of bottom DIC, alkalinity, temperature (T) and salinity (S) to the bottom pH and Ωar anomalies were quantified by perturbing each variable in turn about the 1994–2023 mean state; the contributions sum to the total to within 0.0045 in pH and 0.008 in Ωar. To attribute their anomalies, we used a one-factor-at-a-time perturbation about a fixed reference state defined as the 1994–2023 climatological mean of each driver, (DIC₀, Alk₀, T0, S0). The reference values pH₀ and Ωar0 were obtained by solving the carbonate system at that state.

Four perturbation time series were constructed to estimate the contribution of each driver to interannual pH and Ωar anomalies by allowing a single driver to take its full annual-mean value while the other three were held fixed at their climatological mean. The approximate contribution of each driver to the pH (or Ωar) anomaly in year t is then the departure of its perturbed solution from the reference. For example, ΔpHDIC(t)=pH(DIC(t),Alk0,T0,S0)-pH0.

Because the carbonate system is non-linear, the four single-driver contributions are not guaranteed to add up exactly to the total anomaly; the difference is the residual cross-term arising from simultaneous variation of two or more drivers. In practice this residual is negligible here: the sum of the DIC, alkalinity, T and S contributions reproduces the full bottom pH and Ωar anomalies to within 0.0045 pH units and 0.008 in Ωar over the whole record.

The regionally downscaled retrospective simulations, NWA12_EN4_NUDG, which were used to initialize retrospective forecasts, have an ensemble mean that matches the complex patterns of bottom warming and cooling in the GOM (Fig. 2a). The period from the 1980s to the mid-1990s shows a gradual warming of about 2 °C followed by cooling of the same magnitude until mid-2000s, and intensified warming from the mid-2000s to around the year 2020. This bottom variability in the regional downscaled simulations is consistent with the observed long term surface temperature variations that the Northwest Atlantic Ocean has experienced over the past century, which have been characterized by distinct multi-decadal oscillations and dramatic shifts in water mass properties that have shaped marine ecosystems from the Labrador Sea to Cape Hatteras (Wallace et al., 2018; Townsend et al., 2006; Mountain, 2012). In the late 1990s, after a long period of persistent presence of Warm Slope Water (WSW), cold and fresh Labrador Slope Water (LSW) replaced WSW along the Scotian Shelf and GOM (Drinkwater et al., 1998; Mountain, 2004; Friedland and Hare, 2007). WSW increased in prevalence again during the period from 2004 to 2020 in conjunction with an unprecedented era of accelerated surface and bottom warming in the Northwest Atlantic, making it one of the fastest-warming ocean regions globally (Pershing et al., 2015; Friedland et al., 2020; Kavanaugh et al., 2017; Du Pontavice et al., 2023). Real-time Fishing Industry SHared Bottom Oceanographic Timeseries (Fishbot) data and independent observation-based datasets (Figs. 2a, S1 and S2; see also Record et al., 2024) suggest that the rapid warming has paused since the early 2020s, as predicted by earlier simulations with the downscaled regional model (Koul et al., 2024), and markedly cooler water has returned to the region in 2024.

Simulated bottom salinity closely tracks bottom temperature variability (Figs. 2b and S3; r=0.84); noticeable differences are only seen at the shorter interannual time scales. The correlation between temperature and sparse bottom salinity observations (NOAA_SOE) is weaker but also positive (r=0.58). However, confounding from seasonal and spatial variability limits the inference of annual average salinity anomalies from the sparsely observed salinity field alone. The independent GLORYS12 bottom temperature and salinity estimates, which use a high-resolution model with data assimilation to impute estimates within the spatial and temporal gaps between observations, are highly correlated with our simulation results over the shorter GLORYS record (r=0.86 and r=0.70, respectively) – and have a similarly strong bottom temperature and bottom salinity covariance (Figs. 1b and S3, r=0.84 versus r=0.90).

The water mass variations described above are also accompanied by biogeochemical shifts in both the model and observations (Fig. 2c–e). The annual mean bottom oxygen (O₂) anomalies in the GOM show an inverse relationship with bottom temperature and salinity anomalies. The warming and water mass transitions described above would be expected to influence dissolved oxygen concentrations through differences in O₂ concentration in water masses (Claret et al., 2018), solubility effects, and potential changes in biological oxygen demand. The rise in deep oxygen levels from the mid-1990s, peaking in the latter half of 2004 to 2005, corresponds with the cool, fresh, high LSW conditions well documented in the observational record (Mountain, 2012; Feng et al., 2016). LSW is high in oxygen due to high solubility and rapid oxygen uptake during its formation in the Labrador Sea (Koelling et al., 2023). Since 2005 the bottom oxygen levels have largely dropped, aside from episodic increases in 2015, 2019, and 2024. The relationship between internal (natural) variations in bottom temperature and bottom O₂ confirms that the proportion of WSW/LSW in the Northeast Channel acts as an important driver of this co-variability, though meteorologically linked variations in the winter mixed layer also dictate large episodic variations in bottom O₂ (Figs. 5 and S5). The close correspondence between the observed Fishbot data and the modeled bottom O₂ (Fig. 3c) suggests that our regional downscaled simulation captures the dominant mechanisms controlling bottom oxygen variability.

The bottom pH and aragonite saturation state (Ωar) anomalies in the GOM show an overall declining trend from 1980–2024 consistent with the uptake of steadily increasing atmospheric CO₂ (i.e., predictable large-scale ocean acidification). The simulated Ωar time-series, however, also exhibits more significant inter-annual to decadal variability than the bottom pH (Fig. 2d and e), which is consistent with Salisbury and Jönsson (2018). After removing the externally forced ocean acidification signal, Ωarvariations are positively correlated with temperature (Fig. 3, r=0.75). This is consistent with the carbonate chemistry response to warming and increased alkalinity in warm and saline WSW waters (Salisbury and Jönsson, 2018). As in Salisbury and Jönsson's surface analysis, the pronounced warming and shift to WSW from 2010–2020 was associated with increased Ωar that counteracted the ocean acidification trend seen in pH. The more recent shift toward colder and fresher conditions, however, amplifies the acidification trend and causes a sharp decrease in Ωar.

To isolate the mechanisms underlying this bottom biogeochemical variability, we decomposed each quantity into its controlling factors (Fig. 4). Bottom O₂ was partitioned into a solubility component, the saturation concentration [O₂] sat (T, S), and a non-solubility residual, [O₂] - [O₂] sat, which equals the negative of the apparent oxygen utilisation (- AOU) and reflects ventilation, respiration and advection (Fig. 4a). The non-solubility residual accounts for approximately 70 % of the interannual variance of the bottom-O₂ anomaly (r=0.89 with total O₂), compared with about 21 % for the solubility term. The dominant O₂ variations – including the episodic maxima of 2004, 2015 and 2019 – are therefore driven by ventilation of the deep basins through deep winter mixing (Fig. S5) and by the changing oxygen content of the inflowing slope waters, rather than by temperature and salinity-dependent solubility alone. Solubility instead imparts a smaller, smoother decline as the bottom water warms.

For the carbonate system, we examined the bottom dissolved inorganic carbon (DIC) and total alkalinity, which mix conservatively and respond in straightforward ways to gas exchange, respiration and CaCO₃ cycling. We also examine alkalinity minus DIC (Alk - DIC), a proxy that helps separate the mechanisms driving pH and Ωar (Xue and Cai, 2020; Fig. 4b). In the model the Alk - DIC anomaly is nearly identical to the bottom Ωar anomaly (r=0.996; Fig. 4b and d) but tracks pH relatively loosely (r=0.82, because pH also responds directly to warming). On interannual timescales Alk - DIC (and hence Ωar) is set by water-mass alkalinity changes, while its long-term decline (-2.2 µmol kg⁻¹ per decade), rising DIC outpacing a smaller alkalinity increase, is the anthropogenic acidification trend.

We attributed the bottom pH and Ωar anomalies to DIC, alkalinity, temperature and salinity (Fig. 4b–d; see Sect. 2). Bottom DIC rises steadily at about 4 µmol kg⁻¹ per decade, which is the likely fingerprint of anthropogenic carbon uptake. Alkalinity varies primarily with water mass, being higher during the high-alkalinity WSW phases and lower during Labrador Slope Water phases (Fig. 4b). The long-term pH decline (-0.013 per decade) is led by the rising DIC (-0.010 per decade), with warming contributing a further -0.007 per decade, increasing alkalinity partially buffering the trend (+0.005 per decade), and salinity making up a small remainder (-0.001 per decade) (Fig. 4c). The decline is thus quantitatively consistent with anthropogenic DIC uptake, modulated by warming and water-mass alkalinity.

The long-term Ωar decline (-0.015 per decade) is likewise led by rising DIC, whose direct effect (-0.035 per decade) is more than halved by two opposing influences: increasing alkalinity (+0.016 per decade) and warming (+0.004 per decade) (Fig. 4d). Warming thus raises Ωar even as it lowers pH, which, together with the larger water-mass alkalinity signal, is why the Ωar decline is weaker and less monotonic than that of pH. On interannual timescales, Ωar variability is instead governed primarily by alkalinity (about 57 % of the detrended variance; r=0.91 with the alkalinity contribution), with only a minor direct thermodynamic effect of temperature. The interannual DIC anomaly is small and, because the inflowing WSW in the model does not carry elevated DIC (detrended DIC–alkalinity correlation r=-0.28), the DIC signal slightly reinforces rather than opposes the alkalinity-driven variability (Fig. 4d). The correlation between Ωar and temperature (Fig. 2) therefore arises largely because temperature and alkalinity co-vary through the WSW/LSW water-mass variability, rather than from a direct effect of warming on the carbonate equilibria. Bottom salinity, which co-varies strongly with bottom temperature through the WSW/LSW water-mass variations (Figs. 2a, b, and S3), is retained as an explicit driver of the pH and Ωar anomalies in this decomposition (Fig. 4c and d), where its direct contribution is small.

Decadal prediction skill is derived from both predictable internal variability of the climate system and from predictable natural and anthropogenic external forcing (Meehl et al., 2009). Capturing the latter requires skillful representation of external drivers (e.g., accumulating CO₂) and its impact on the climate, while capturing the former requires skillful initialization of the climate system and the ability to simulate its natural evolution from that state. The analysis in Figs. 2 and 3 suggested that our retrospective ensemble hindcasts (NWA12_HIND) captured observed variations and covariations across bottom habitat properties, and that the prominence of externally forced and internal variability varies between critical bottom habitat quantities. Comparison of extensive retrospective forecasts against these patterns show that some aspects of both signals were successfully predicted, but others were not (Fig. 5).

Bottom temperature predictions from NWA12_HIND exhibit significant positive anomaly correlations in most of the 10 lead years (Fig. 5a); and in the first lead year, the prediction skill exceeds the skill from the climate change signal (indicated by the diamonds and squares). Both results are consistent with a predictable bottom-temperature signal from climate change and contributions from predictable bottom-temperature (i.e., water mass) variability in the first few lead years. Closer inspection of the prediction skill after linear de-trending further supports the system's capacity to predict aspects of variability around the climate change trend over all the lead years, except at lead years 4 and 5 when the skill drops (Fig. S9). The prediction skill is robust when verified against GLORYS as the observational reference for the period 1993–2024 (Fig. S4).

Initializations at states far from the uninitialized state, the 1980s warming, the rapid 2005–2015 warming and the recent shift to cooler/fresher conditions from 2020 onward, are all correctly predicted by the retrospective predictions (Fig. 6). The 2025 forecast shows a slow reversal of the recent cooling/freshening during the next decade. These results could be interpreted as the model simply relaxing toward the uninitialized state. However, the overall good correlation skill indicates that the model accurately captures the timing and magnitude of such transitions.

The prediction skill for bottom O₂ drops to 0 beyond 2 years and much of the signal for this important habitat factor is linked to predictable internal (natural) variability (Fig. 5b). This is also apparent from the time series (Fig. 2c) which shows that bottom O₂ does not have a uniform long-term trend as was the case for bottom temperature. The rapid drop in prediction skill may also be linked to the presence of sharp and likely less predictable meteorologically driven variations in the depth of the winter mixed layer. The winter 2003/2004 is a prime example of this phenomenon when large positive winter MLD anomalies in our regional downscaled model drove ventilation of the bottom water and strong positive bottom oxygen anomalies (Figs. 6, and S5). Similar peaks in bottom O₂coinciding with deep winter mixing and positive MLD anomalies are seen in 2015 and 2019.

The decadal predictions initialized before and after the 2004 peak (in 2002 and 2006 respectively) show that the ensemble mean prediction has difficulties predicting the extent of the rise in bottom O₂, but the subsequent decay of these elevated bottom O₂ levels is well predicted by the model (Fig. 6b). These characteristics of the past variability in bottom O₂ are consistent with the high skill of the initialized predictions, which outperform the persistence-based skill and the skill of uninitialized simulations in the first two lead years (Fig. 5b), and the significant drop thereafter.

Both bottom pH and Ωar exhibit a declining trend (Fig. 2e and f), yet they differ in how much the climate change signal influences their prediction skill. Ocean pH is primarily controlled by increasing atmospheric CO₂, while Ωar is more strongly influenced by ocean water masses with varying alkalinity (LSW versus WSW). As a result, internal variability contributes more to the prediction skill of Ωar, whereas the climate change signal has a greater impact on the prediction skill of bottom pH (Fig. 5c and d). Beyond lead year 2, the climate change signal predominantly determines the prediction skill of bottom pH, allowing it to remain high and significantly better than persistence for all lead years (except lead year 3) which is not the case for Ωar where the skill drops rapidly beyond lead year 2. When the linear trend is removed, the skill of internal variability significantly surpasses the persistence and climate change-driven skill in the first two lead years. The notable decrease in the prediction skill of detrended bottom pH, compared to Ωar, further demonstrates the additional contribution of internal variability to the prediction skill of Ωar. Individual 10-year forecasts provide further evidence of the capability of the model to track past deviation of bottom pH and Ωar from the long-term trend (Fig. 6c and d).

The preceding analyses leveraged over 6500 years of retrospective forecasts to assess the capacity of the ensemble mean prediction to anticipate physical and biogeochemical changes on a decadal time horizon. While results suggested significant capability arising from predictable climate variability and change signals, decision-making benefits from actionable information on uncertainties (e.g., Ramos et al., 2013; Roulston et al., 2006). A coastal resource manager or fishing industry representative might ask, “How reliable are forecasts for anomalously warm or cool conditions?”. To quantify this in more accessible terms, we calculated the accuracy of probabilistic forecasts derived from the model simulations (Tommasi et al., 2017a, see Sect. 2). We used terciles to define above, normal, and below-normal conditions and focused on predicted conditions averaged over the next 3 years, which corresponds to typical stock assessment timelines in actively managed fisheries (Tommasi et al., 2017b). Extended logistic regression (Wilks, 2009) was used to calibrate forecast uncertainty based on the forecast error (see Sect. 2).

Our regional downscaled predictions skilfully distinguish between tercile classes. All four variables have positive ranked probability skill scores, and the forecast accuracy for a given tercile exceeds the accuracy of a random forecast (0.56) and the accuracy of consistently negative prediction (0.67) in all but two cases (Fig. 7). The ensemble confidently predicted the observed cooling in the mid-2000s and the subsequent warming phase that extended from 2005 to approximately 2020 (Fig. 7).

The RPSS for all quantities is >0, indicating the model has some capacity to anticipate changes (i.e., probabilistic skill beyond the persistence forecast, see Sect. 2). Oxygen has slightly lower skill (RPSS 0.46, accuracy 0.67–0.81) compared to bottom temperature (RPSS 0.47, accuracy 0.72–0.90), reflecting both the weak predictable climate change signals and the influence of less predictable atmospheric variability. Similarly, predictions for bottom pH, which has a very strong predictable climate change signal, have higher skill (RPSS 0.6, accuracy 0.76–0.88) than Ωar (RPSS 0.51, accuracy 0.62–0.79).

The high accuracy of prediction for bottom temperature, O₂, pH and Ωar likely reflects both methodological and mechanistic factors. The bottom temperature, O₂, pH and Ωar are validated against the same regional downscaled simulation from which initial conditions were derived (see Sect. 2). While this simulation recreates observed O₂ patterns and covariances with well-constrained physical properties to the extent that they have been observed (Figs. 2, 3 and S7), it should be viewed as an upper limit for prediction skill. Additional autonomous biogeochemical observing platforms capable of observing shelf environments may help address this limitation in the future (Henson, 2014; Chai et al., 2020).

Looking ahead over the next three years (2025–2027), we forecast below-normal bottom temperatures with high confidence (88.3 % probability), supported by 93.3 % retrospective accuracy when issuing similar high confidence forecasts (14/15 hits). We also predict above-normal bottom O₂ levels with high confidence (92.0 % probability), with an 85.7 % retrospective accuracy at comparable confidence levels (6/7 hits). Additionally, we forecast with very high confidence (>99 % probability) both below-normal bottom pH and below-normal Ωar conditions, both of which are supported by perfect retrospective accuracy when issuing very high confidence forecasts, albeit with limited sample size (9/9 hits and 1/1 hit respectively). While these forecasts represent our best estimate, both the confidence levels and the underlying sample sizes should be considered when making decisions based on these predictions, particularly for the forecasts where limited similarly confident historical forecasts constrain our ability to fully validate forecast skill. Nevertheless, the dynamically consistent physical and biogeochemical forecasts provide compelling evidence that the ongoing watermass reorganization will continue to influence bottom water conditions through the 2025–2027 period.