Skilful prediction of cod stocks in the North and Barents Sea a decade in advance

The study addresses how climate variability and change affect marine resources and whether decadal predictions of fish stock abundance are feasible and useful for management. Marine ecosystems and commercially important stocks, such as Atlantic cod (Gadus morhua), are strongly influenced by climate variability on interannual to decadal timescales, interacting with fishing pressure to shape productivity. Despite this, environmental variables are rarely integrated into management-oriented fish population forecasts due to non-stationary climate impacts, non-linear stock dynamics, and limitations of global climate models (GCMs) in representing shelf seas. The North Atlantic exhibits significant and predictable decadal variability, suggesting potential to leverage decadal climate predictions for fisheries. The paper focuses on two contrasting cod stocks: North Sea cod, near the species’ upper thermal limit, historically over-exploited and in a low-productive state, and Northeast Arctic cod in the Barents Sea, near the species’ lower thermal limit with recent record-high biomass. The research question is whether a combined dynamical-statistical approach, linking GCM-based decadal sea surface temperature (SST) predictions to cod total stock biomass (TSB), can provide skilful 1–10 year forecasts under different fishing scenarios, and to identify the sources of forecast skill for these stocks.

The paper synthesizes prior work showing strong climate impacts on fish populations via recruitment, growth, fecundity, migration, and food availability, and highlights the combined, potentially non-linear effects of climate and fishing on stock dynamics. Previous approaches include bioclimate envelope models and complex ecosystem/population dynamics models driven by GCM projections for long-term (>30 years) change, but these are computationally intensive and less suited for decadal forecasts. GCM limitations in resolving shelf-sea processes and trophic interactions have constrained forecast utility. Notably, the North Atlantic’s subpolar gyre (SPG) shows pronounced and predictable decadal SST variability, which has been linked to Barents Sea ecosystems and cod. Prior statistical studies have related SPG SST to Northeast Arctic cod with multi-year lead times, while North Sea cod has been negatively impacted by warming and overfishing. This work builds on and extends such findings by exploiting initialized decadal predictions from MPI-ESM within a unified dynamical-statistical framework and explicitly testing skill sources (external forcing trends vs initialization).

Overview: A unified dynamical-statistical prediction system is constructed. Statistical linear regression models link cod total stock biomass (TSB) to sea temperature and fishing mortality, driven by decadal climate predictions from the MPI-ESM-LR decadal prediction system. Forecasts are issued for 2020–2030 under multiple fishing scenarios. Cod stocks and predictors: - Stocks: North Sea cod; Northeast Arctic cod (Barents Sea). - Predictand: Total stock biomass (TSB) from ICES assessments. - Predictors: For North Sea cod, local North Sea SST; for Northeast Arctic cod, North Atlantic subpolar gyre (SPG) SST. Fishing mortality (F) considered in multiple regression models (North Sea: F for ages 2–4; Northeast Arctic: F for ages 5–10). - Observational/assimilated temperature: Area-mean SST from MPI-ESM-LR assimilation experiment first model layer (~6 m), validated against AHOI (North Sea) and HadISST (SPG, Barents Sea Opening). Dynamical model and decadal predictions: - Model: MPI-ESM1.2-LR with MPIOM ocean (40 z-levels; GR15 grid with enhanced resolution north of 50°N), HAMOCC biogeochemistry, ECHAM6 atmosphere (T63; 47 levels), and JSBACH land module. - Initialized hindcasts: MiKlip decadal predictions; 10-year integrations initialized annually on 1 Nov from 1960–2019; 16 ensemble members. Initialization via oceanic ensemble Kalman filter (assimilating EN4 monthly T/S profiles) and atmospheric nudging to ERA40/ERA-Interim (excluding SST and lower-tropospheric temperatures to maintain consistency across the air-sea boundary). Forcings include observed solar, volcanic; greenhouse gases follow CMIP6 (RCP4.5 from 2006). - Historical simulations: 16-member MPI-ESM-LR Grand Ensemble (1850–2005) under observed natural and anthropogenic forcings; extended with RCP8.5 from 2006 to provide a non-initialized baseline for comparison. Not phase-locked to observations, so internal variability timing can differ. Statistical models: - Simple linear regression: TSB(y) = β0 + β1·T(y−LT) using anomalies (mean 1970–2019 removed). - Multiple linear regression: TSB(y) = β0 + β1·T(y−LT) + β2·F(y−LF), using anomalies. - Lags (LT, LF) chosen as those maximizing correlation with TSB based on observed/assimilated records: North Sea: SST has a contemporaneous/1-year effect (detrended SST impacts TSB at lag 1, r ≈ −0.48). Northeast Arctic: SPG SST leads TSB by 7 years (r ≈ 0.78); F leads TSB by 2 years (r ≈ −0.88). Model training and cross-validation: - Training data: Assimilation-run SST and ICES TSB (and F where applicable), 1970–2019 anomalies. - 80/20 cross-validation: Random block sampling with replacement to select 80% training and 20% testing sets; coefficients estimated on training, applied to testing; repeated 1000 times; report median skill and 95% confidence intervals for train/test correlations. Dynamical-statistical hindcasts and forecasts: - Apply trained regression coefficients to dynamically predicted SST: - For North Sea: use lead-year 10 SST predictions (extend predictability horizon). - For SPG: use lead-year 4 SST predictions (consistent with best decadal skill and 7-year lag to TSB). - Multiple-regression forecasts incorporate fishing scenarios (Northeast Arctic: Fso=0.42, FMSY=0.40, FLIM=0.74; North Sea: Fso=0.50, FMSY=0.30, FLIM=0.54). Temperature predictor for scenario forecasts taken from initialized hindcasts; for comparison, also from historical+RCP8.5. - Output as ensemble distributions using 16-member climate ensembles and 1000 bootstrap samples of regression coefficients. Bias correction and skill assessment: - Dynamical hindcasts are corrected for lead-time-dependent drift; lead-year-dependent climatology (1970–2019 mean) removed before computing anomalies and feeding statistical models. - Skill metrics: Anomaly correlation coefficient (ACC) and Mean Square Error Skill Score (MSESS = 1 − MSE/MSEREF, with climatology as reference). Uncertainties via 6-year overlapping block bootstrap in time and across ensemble members (1000 resamples) to provide 95% confidence intervals. Data and code: - Data sources: AHOI, HadISST, ICES assessments, MPI-ESM Grand Ensemble, DKRZ archive for assimilation/decadal predictions. Plotting and post-processing scripts available on request from corresponding author.

- Observed variability and drivers: - North Sea: SST warmed by 1.68 °C over 1960–2019, exceeding interannual variability (σ ≈ 0.65 °C). TSB decreased over decades, with persistently low levels since early 2000s. Detrended SST negatively correlates with next-year TSB (r = −0.48, p = 0.0025). Fishing mortality (ages 2–4) shows weak correlation with TSB (r = −0.19, ns), consistent with warming-driven low productivity inhibiting recovery. - Northeast Arctic: TSB shows multi-annual to decadal variability with recent record highs; strong linkage to SPG SST at 7-year lead (r = 0.78, p = 0.0435; detrended r = 0.77). Fishing mortality (ages 5–10) strongly anti-correlated with TSB at 2-year lead (r = −0.88, p = 0.0035). Temperature has an opposite effect compared to North Sea (positive impact on biomass). - Statistical model performance (cross-validated): - North Sea: Fishing-only model has no predictive power; Temperature-only model has significant skill; Adding fishing does not improve skill relative to temperature-only. - Northeast Arctic: Fishing-only (adj. R2 ≈ 0.77) and Temperature-only (adj. R2 ≈ 0.62) both skilful; Combined model best fit (adj. R2 ≈ 0.84), but forecast horizon limited by predictor lead times (F leads 2 years; SPG SST leads 7 years). For long-horizon predictions, temperature-only chosen; forecasts complemented with combined T+F scenarios. - Dynamical SST prediction skill: - North Sea: High ACC out to 10-year lead for both initialized hindcasts and non-initialized historical simulations, indicating skill dominated by externally forced warming trend. - SPG: Initialized hindcasts substantially outperform historical simulations in ACC; skill arises from initialization capturing decadal cooling/warming episodes; historical simulations under-represent decadal variability amplitude. - Dynamical-statistical hindcast skill for TSB: - North Sea: Robust correlation and MSESS using temperature-only model; initialized and historical-driven reconstructions yield comparable skill, consistent with trend-dominated predictability. - Northeast Arctic: Initialized hindcasts reproduce 1970s decline and post-2005 increase; higher MSESS than historical; historical-based reconstructions suppress variability and fail to capture recent decadal shift, yielding MSESS not significantly better than climatology. - Forecasts for 2020–2030: - North Sea cod: Continued unfavorable oceanic conditions projected; no significant recovery under any fishing scenario (Fso=0.50, FMSY=0.30, FLIM=0.54). Forecasts largely reflect the long-term warming-driven decline rather than interannual variability. - Northeast Arctic cod: Climate-driven decline in TSB relative to recent high levels across all scenarios (Fso=0.42, FMSY=0.40, FLIM=0.74), primarily due to delayed (advective) impact of 2010–2016 SPG cooling. Initialized forecasts indicate a larger decline than historical+RCP8.5-based projections, suggesting scenario-based forecasts may underestimate the downturn if relying on non-initialized SST. - Source of predictability: - North Sea: Externally forced SST trend dominates skill. - Northeast Arctic: Decadal SPG variability captured by initialization provides main source of skill; local Barents Sea SST explains little of TSB variability compared to SPG SST.

The study demonstrates that decadal prediction of cod biomass is feasible when robust climate–ecosystem linkages exist and the physical environment is predictable at multiyear leads. For North Sea cod, strong negative sensitivity to warming, combined with trend-dominated SST predictability, enables projections indicating continued low productivity and limited recovery even under reduced fishing mortality. This highlights the role of climate constraints on management outcomes for stocks near thermal limits. For Northeast Arctic cod, positive temperature effects and strong linkage to SPG decadal variability enable skilful hindcasts and forecasts of TSB when using initialized climate predictions. Identifying the differing sources of skill—forced trend in the North Sea versus initialized decadal variability in the SPG for the Barents Sea—clarifies when and why decadal fishery forecasts can be trusted. These results are directly relevant to adaptive fisheries management, enabling planning of exploitation rates that consider expected climate-driven productivity fluctuations on multi-year to decadal horizons. The analysis also underscores that non-initialized projections may under-represent near-term decadal swings critical for management decisions.

This work bridges climate prediction and fisheries forecasting by developing and validating a dynamical-statistical system that provides skilful decadal predictions of cod biomass in the North and Barents Seas. The approach leverages initialized decadal SST predictions from a global climate model and statistical models linking temperature (and fishing) to TSB, incorporating uncertainty via ensemble climate predictions, cross-validation, and bootstrapping. Key contributions include: (1) Demonstration of 10-year predictive skill for cod biomass; (2) Identification of distinct predictability sources—externally forced SST trend for North Sea cod versus initialized SPG decadal variability for Northeast Arctic cod; (3) Practical forecasts (2020–2030) under multiple fishing scenarios, indicating continued low productivity for North Sea cod and a likely decline from recent highs for Northeast Arctic cod. Future research should assess stationarity of climate–ecosystem relationships under ongoing climate change, better represent shelf-sea processes and open ocean–shelf exchanges in Earth system models, integrate additional ecosystem variables (e.g., prey dynamics, carrying capacity), explore non-linear or state-dependent statistical/dynamical models, and extend the framework to other regions and species.

- Low-frequency climate variability assumption: Skill estimates assume the training period represents the relevant spectrum of SPG and regional variability; regime shifts could reduce forecast skill. - Use of ICES assessments: TSB and F are model-derived quantities and not fully independent observations, potentially introducing dependencies in training and evaluation. - Model simplicity: Linear regression models may not capture non-linear dynamics, state dependence, or changes in carrying capacity and lifetime reproductive output. - Implicit ecosystem processes: Species interactions and ecosystem feedbacks are not explicitly modeled; their omission may affect forecast robustness. - Regional heterogeneity: North Sea cod shows spatial structure (e.g., stronger declines in the south), which is not resolved in the basin-scale predictor–response framework. - Non-initialized scenario limitations: Historical+RCP simulations under-represent decadal variability amplitude, potentially underestimating near-term changes if used for forecasts.