A
A Targeted Correction for Arctic Oscillation Biasesin Seasonal Forecasts
Ji-Han Sim1, Baek-Min Kim1, Ha-Rim Kim2, Gaeun Kim3, A-Young Lim3,
Yoo-Rim Jung3, Ju Heon Kim1, Geun Young Kim1, and Steven Cocke4
1Division of Earth Environmental System Science, Pukyong National University, Busan, South Korea
2Supercomputer Center, Institute of Environmental and Marine Science Technology, Pukyong National University, Busan, South Korea
3Climate Services and Research Department, APEC Climate Center, Busan, South Korea
4Center for Ocean-Atmosphere Prediction Studies, Florida State University, Florida, USA
Corresponding Author: Baek-Min Kim, Department of Environmental Atmospheric Sciences, Pukyong National University, Yongsoro 45, Namgu, Busan 48513, Korea.
e-mail : baekmin@pknu.ac.kr
Abstract
Seasonal prediction of Eurasian winter surface air temperature (SAT) is fundamentally limited by systematic biases in representing Arctic Oscillation (AO) teleconnections. Across the APEC Climate Center (APCC) Multi-Model Ensemble (MME), most component models misrepresent the impact of the AO on SAT, leading to degraded skill in Eurasian forecasts. To address this deficiency, we applied an AO correction method that replaces the distorted AO–SAT relationship simulated by models with one constrained by observations. The correction systematically enhanced winter SAT prediction skill across all APCC MME models, demonstrating that the improvement is not model-specific but generalizable across the ensemble. Importantly, the magnitude of the SAT ACC improvement showed a strong inverse relationship with intrinsic AO skill: models with the weakest ability to capture AO variability experienced the greatest gains (correlation r = − 0.65). This finding highlights that AO correction functions not as a uniform adjustment, but as a targeted remedy for a structural deficiency common to current generation seasonal forecast models. By demonstrating that a targeted correction of the AO yields systematic skill improvements—particularly in models with the poorest representation of Arctic variability—this study establishes a generalizable and physically-motivated post-processing framework for enhancing seasonal climate prediction.
Key words:
Arctic Oscillation correction
Seasonal prediction skill
APCC multi-model ensemble
A
1. Introduction
The prediction of winter surface air temperature (SAT) over Eurasia carries high socio-economic stakes, influencing everything from energy demand management and agricultural planning to the mitigation of risks from extreme cold events (Cohen et al., 2014; Sillmann et al., 2017). A primary driver of this wintertime variability is the Arctic Oscillation (AO), a large-scale seesaw of atmospheric pressure between the Arctic and the mid-latitudes that governs the southward flow of cold polar air (Thompson and Wallace 1998; Wang et al., 2005; Kim et al., 2021b). However, a grand challenge in seasonal forecasting is that even the most sophisticated, state-of-the-art coupled general circulation models (CGCMs) systematically struggle to capture the AO's influence. This deficiency, which often manifests as a misplacement of the AO's centers of action or an underestimation of its amplitude, results in a weakened and unrealistic teleconnection strength over Eurasia and North America (Kang et al., 2014; Ren and Nie, 2021). This creates a critical gap in forecast skill precisely where it is needed most, particularly over central Eurasia where skill remains stubbornly low (Jung et al., 2018; Yhang et al., 2025).
Recent advances suggest that statistical post-processing, when guided by physical understanding, can offer a promising path to mitigating these systematic biases. Our prior work, for instance, provided initial evidence that a targeted correction of the AO–SAT relationship could effectively improve prediction skill in a single model (Sim et al., 2024). However, this left critical questions unanswered. While this approach has shown promise, can it serve as a generalizable strategy across a diverse ensemble of state-of-the-art forecast systems? And, crucially, how does the effectiveness of such a correction relate to the intrinsic skill of the models themselves? This study addresses these gaps by: (1) systematically evaluating the AO correction methodology across all 14 models of the APCC MME; (2) quantifying the relationship between a model's intrinsic AO skill and the improvement gained from the correction; and (3) assessing the added value of a hybrid approach that combines physical model improvements with statistical post-processing. By focusing on AO correction as a targeted remedy for structural deficiencies in current-generation models, this study highlights a physically-motivated and generalizable post-processing strategy for advancing seasonal climate prediction over Eurasia.
The remainder of this paper is organized as follows: Section 2 describes the data, models, and AO correction methodology applied in this study. Section 3 presents the results, focusing on the misrepresentation of AO–SAT teleconnections in the APCC MME models, the improvements achieved through AO correction, and the contribution of PKNU models to the ensemble. Section 4 summarizes the key findings, discusses their broader implications for seasonal prediction, and outlines directions for future research.
2. Data and methods
2.1 Development of the PKNU Model
PKNU CGCMv1.0 (hereafter PKNU) is included in the APCC MME group, while PKNU_I is currently under development (Table 1). The version of PKNU used in this study is based on the Community Earth System Model version 2 (CESM2) coupled climate system model, which includes atmosphere (CAM6), ocean (POP2; Smith et al., 2010; Danabasoglu et al., 2012), land (CLM5; Lawrence et al., 2019), and sea ice (CICE5; Hunke et al., 2015) components (Sim et al., 2024). The initialization method follows the CMIP6-endorsed protocols (Griffies et al., 2016; Tsujino et al., 2020), employing ECMWF reanalysis version 5 (ERA5; Hersbach et al. 2020) and Japanese 55-year Reanalysis (JRA-55; Kobayashi et al. 2015) products for the atmosphere, land, ocean, and sea ice. Ensemble forecasts, consisting of 31 members, are generated using the random field perturbation (RFP; Magnusson et al. 2009) method. For further details regarding the model description and experimental design, readers are referred to Sim et al. (2024).
PKNU_I is developed as an upgraded version of PKNU by incorporating sea-ice observational data into the initialization process while maintaining the model’s dynamical balance. This enhancement is physically motivated, as sea-ice anomalies strongly regulate surface heat fluxes and atmospheric circulation, thereby exerting a first-order influence on winter SAT. This modification improves the representation of sea-ice initial conditions, thereby enhancing the model’s prediction skill for winter SAT over Eurasia (Sim et al., 2025).
PKNU_IAO represents a further extension of PKNU_I, in which the AO correction method is applied to systematically replace the biased AO–SAT teleconnections simulated by the model with observation-constrained relationships. In contrast to PKNU_I, which primarily improves initialization, PKNU_IAO directly addresses structural teleconnection errors, enabling a more comprehensive correction framework. Detailed descriptions of the AO correction procedure are provided in Section 2.3.
PKNU version | OCN & ICE Initial conditions | AO correction |
|---|---|---|
PKNU | Only ATM forcing | X |
PKNU_I | ATM forcing & ICE nudging | X |
PKNU_IAO | ATM forcing & ICE nudging | O |
2.2 Configuration of the APCC MME Models
The APCC MME used in this study consists of 14 models, including the PKNU model (Table 2). Throughout this paper, the term "APCC MME" refers to this set of models unless otherwise specified. The PKNU_I model, an upgraded version not officially part of the MME, is analyzed separately, and its inclusion in any analysis is explicitly stated. To ensure a fair comparison, the hindcast period is fixed to span from 1993 to 2016. For models with shorter hindcast records—BCC (1993–2015) and HMC (1993–2015)—the analysis was conducted over their available hindcast periods. All analyses are performed using the ensemble mean of each model. For verification, we used JRA-55 for the winter seasons (December–January–February, DJF-mean) from 1993/1994 to 2016/2017. The prediction skill of winter SAT is evaluated using the anomaly correlation coefficient (ACC), which measures the temporal correlation between the model hindcasts and reanalysis anomalies relative to the respective climatology. Prior to analysis, long-term trends were removed from both the hindcast and reanalysis data.
Organization | System name | Periods | Ensemble size | Reference |
|---|---|---|---|---|
(1) PKNU | CGCMv1.0 | 1993–2016 | 31 | Sim et al. (2024) |
(2) APCC | SCOPS | 1993–2016 | 10 | Ham et al. (2019) |
(3) BCC | CSM1.1M | 1993–2015 | 24 | Wu et al. (2010) |
(4) BOM | ACCESS-S2 | 1993–2016 | 27 | Wedd et al. (2022) |
(5) CMCC | SPS3.5 | 1993–2016 | 40 | Gualdi et al. (2020) |
(6) CWA | TCWA1Tv1.1 | 1993–2016 | 30 | Juang et al. (2024) |
(7) ECCC | CANSIPSv3 | 1993–2016 | 40 | Lin et al. (2020) |
(8) HMC | SL-AV | 1993–2015 | 11 | Fadeev et al. (2019) |
(9) KMA | GLOSEA6GC3.2 | 1993–2016 | 28 | Kim et al., (2021a) |
(10) METFR | SYS8 | 1993–2016 | 25 | Penabad and Dorel (2023) |
(11) MGO | MGOAM2.4 | 1993–2016 | 10 | Meleshko et al. (2014) |
(12) NASA | GEOS-S2S-2.1 | 1993–2016 | 4 | Nakada et al. (2018) |
(13) PNU-RDA | CGCMv2.0 | 1993–2016 | 35 | Ahn et al. (2018) |
(14) UKMO | GLOSEA6 | 1993–2016 | 28 | Williams et al. (2018) |
2.3 AO Correction Method
The AO correction post-processing method applied in this study follows the approach proposed by Sim et al. (2024). This method was designed to address the limited prediction skill in central Eurasia, where model biases in simulating the AO have been identified as a major source of error. The procedure can be formally expressed as:
In Eq. (1),
denotes the corrected SAT forecast. Subscripts f, h, and o represent forecast, hindcast, and observation, respectively; Reg denotes the linear regression operator. The equation consists of three terms: the first term
represents the dynamical model-predicted SAT. The second term
removes the AO–SAT teleconnection effect as simulated by the dynamical model. The third term
, restores the observed AO–SAT teleconnection using the AO index estimated by a multiple linear regression (MLR) model. To prevent artificial skill inflation from data leakage, all statistical relationships, including the observational regression and the MLR model, were calculated using a leave-one-out cross-validation (LOOCV) scheme. In this procedure, the data for the year being predicted is excluded from the training of the statistical models.
Two complementary strategies were employed:
1.
AO Pattern Correction The spatial loading pattern of the AO simulated by each model was replaced with the observed AO pattern derived from linearly detrended reanalysis data. This correction is justified by the fact that many models exhibit systematic errors in the location of the AO’s centers of action (Kang et al., 2014), which in turn generates spurious teleconnection patterns across the Northern Hemisphere. To avoid degrading prediction skill in regions where observed and modeled regressions diverge, this study introduced an additional constraint not included in Sim et al. (2024): regression-based AO impacts were removed only in regions where both the observed and model-simulated AO–SAT regressions were statistically significant at the 95% confidence level, thereby ensuring that meaningful signals were preserved. This constraint serves as a conservative measure to ensure that the correction is only applied where there is robust evidence of an AO teleconnection in both the model and observations. It prevents the erroneous removal of skillful model signals in regions where the model simulates a statistically significant local effect that is unrelated to the observed AO pattern.
2.
AO Index Correction
Even with the correct spatial pattern, models still face challenges in predicting the temporal variability of the AO index (Riddle et al., 2013; Nie 2021). To address this, this study employs an empirical prediction model for the AO index based on the multiple linear regression (MLR) approach detailed in Sim et al. (2024). This model is constructed using three physical precursors, chosen for their known influence on the AO and their low inter-correlation, which minimizes multicollinearity issues. The predictors are summer (June-August) sea surface temperature (SST) averaged over a highly correlated North Atlantic region (45°W–25°W, 30°N–45°N); September sea-ice concentration (SIC) from the Barents-Kara and Bering Seas; and October snow cover extent (SCE) from a North American region (110°–85°W, 35°–50°N) that exhibits a negative correlation with the AO index. This statistical prediction of the AO index effectively reduces systematic biases in amplitude and phase found in the dynamical model's forecast, thereby improving its temporal predictability. The statistical correction thus complements the pattern correction, jointly addressing both spatial and temporal deficiencies. Further technical details and validation are provided in Sim et al. (2024).
3. Results
3.1 Winter surface air temperature prediction skill
A primary challenge in seasonal forecasting is the significant disparity in prediction skill between oceanic and terrestrial regions, with land areas consistently underperforming. This is evident in our comparison of winter SAT prediction skill, evaluated using the ACC, over the Northern Hemisphere extratropics (north of 20°N) across the APCC MME and PKNU models (Fig. 1). The ECCC model, for instance, exhibited the highest overall performance with a relatively balanced land-ocean skill gap of only 0.08. In stark contrast, the NASA model showed the largest discrepancy of 0.23, with markedly deficient skill over land (ACC = 0.13) despite reasonable performance over the ocean (ACC = 0.36).
This systematic weakness over land is generally attributed to complex terrestrial interactions and feedback mechanisms that are challenging for models to simulate accurately. Furthermore, prediction over major landmasses like Eurasia is fundamentally limited by the misrepresentation of large-scale atmospheric teleconnections, which will be the central focus of subsequent sections. The practical significance of these improvements is highlighted by the PKNU model's upgrade; for example, the upgrade from PKNU to PKNU_I reduced the land-ocean skill gap from 0.17 to 0.13, raising its rank from ninth to fifth. This represents a meaningful enhancement in the model's ability to correctly forecast the sign and pattern of winter temperature anomalies. Therefore, these results underscore a critical imperative: advancing overall seasonal prediction capabilities requires a focused effort on improving skill over land, the domain where current models consistently show their most significant deficiencies.
Fig. 1
Comparison of winter SAT prediction skill over the Northern Hemisphere extratropics (north of 20°N) across APCC MME participating models and PKNU_I. Gray bars represent the total domain, green bars denote land areas, and blue bars indicate ocean areas, respectively, as defined by a land-sea mask. The values above each bar correspond to the anomaly correlation coefficient (ACC) averaged over each domain. The models are arranged in ascending order of prediction skill based on the total-domain ACC.
3.2 AO prediction characteristics of APCC MME participating models
One of the most important factors controlling winter land SAT in the Northern Hemisphere is the AO (Thompson and Wallace, 1998). The AO exerts a pronounced influence on Eurasia and North America (Fig. 2a). However, most models fail to represent this impact realistically, particularly by underestimating its strength over Eurasia. Figure 2b shows the regression pattern averaged across the APCC MME participating models. Models systematically underestimate the influence of the AO over Eurasia and North America, while simultaneously producing spurious and statistically significant relationships over the Pacific Ocean, where no such effect is expected based on observations. This highlights the inability of current models to correctly capture the AO–SAT teleconnection structure, which is a key source of uncertainty (or prediction error) over land regions.
Fig. 2
Comparison of the Arctic Oscillation (AO) influence on winter surface air temperature (SAT). (a) Regression pattern of the AO index from JRA-55 against SAT anomalies, representing the observed AO–SAT relationship. (b) Multi-model average of regression patterns between the AO index and SAT from APCC MME participating models. Green hatches indicate statistical significance at the 95% confidence level.
Beyond the spatial pattern, the models also struggle to predict the temporal variability of the AO index, further limiting their usefulness for seasonal prediction. Most models show low skill, with correlation coefficients below 0.3 (Fig. 3a). The PKNU and PKNU_I models likewise exhibit very limited predictive capability, with correlation coefficients of only 0.11 and 0.04, respectively, neither of which is statistically significant.
Fig. 3
The inverse relationship between intrinsic AO prediction skill and the improvement gained from the AO correction for the APCC MME participating models and the PKNU_I model. (a) Intrinsic AO prediction skill, measured as the correlation coefficient between the predicted and observed winter (1993/94–2016/17) AO index. (b) The corresponding improvement in winter SAT prediction skill (ΔACC) after applying the AO correction. Skill changes are shown for the Northern Hemisphere extratropical domain (north of 20°N) (gray), land areas only (green), and ocean areas only (blue). Models in both panels are ordered by their increasing AO skill as shown in panel (a).
3.3 Application of AO Correction to the APCC MME
To overcome the misrepresentation of AO teleconnection patterns and the generally low predictive skill for the AO index, we applied the AO correction method to each model (Fig. 3b). The results reveal that improvements were predominantly concentrated over land regions, whereas changes over the ocean were minimal, reflecting the greater importance of correcting AO-related biases for land-based climates. Specifically, the land-domain ACC showed substantial gains across most models, with the CWA and NASA models exhibiting the largest improvements of + 0.052. In contrast, improvements over the ocean were negligible and, in some cases, slightly negative (e.g., -0.005 for the BOM model), confirming that the correction primarily benefits land-based forecasts. A key finding emerges when analyzing these improvements: the magnitude of the ACC improvement shown in Fig. 3b has a strong inverse relationship with each model’s intrinsic AO index prediction skill, as presented in Fig. 3a. This relationship is quantified by a significant negative correlation coefficient (r = − 0.65).
This finding indicates that the AO correction is most effective for models with the weakest AO prediction skill. The compensatory mechanism operates by directly substituting the model's flawed AO representation with observation-based information. Specifically, the correction first removes the model's distorted AO–SAT teleconnection pattern and replaces it with the observed spatial pattern. It then substitutes the model's low-skill AO index with a more accurate index predicted by an empirical model. Therefore, for models weak in simulating the AO, the correction bypasses these flawed internal dynamics and imposes a more realistic AO influence, leading to targeted improvements in AO-sensitive regions like Eurasia.
This inverse relationship is clearly illustrated by comparing two models that represent the extremes in forecast improvement after the correction: the CWA model, which gained the most skill, and the ECCC model, which gained the least (Fig. 4). These two cases perfectly exemplify the relationship with intrinsic AO skill: CWA, with its large ACC improvement, has very poor AO prediction skill, while ECCC, with its minimal improvement, has one of the strongest AO prediction skills in the ensemble (Fig. 3a). In the CWA model, the poor AO skill is reflected in spurious negative correlations over Eurasia in the raw forecast (Fig. 4a). The AO correction dramatically reverses this, leading to substantial skill improvements across the continent (Fig. 4b, c). Conversely, the ECCC model, with its high AO prediction skill, already shows strong raw performance over Eurasia that is only marginally enhanced by the correction (Fig. 4d-f).
Fig. 4
Comparison of the effect of AO correction for two representative models from APCC MME. Panels (a–c) show results from the CWA model, and panels (d–f) show results from the ECCC model. (a, d) Raw SAT anomaly correlation coefficients (ACCs), (b, e) SAT ACCs after applying AO correction, and (c, f) the differences between AO-corrected and raw fields during boreal winter (December–January–February, DJF) for the period 1993/94–2016/17. Green hatches indicate regions where the results are statistically significant at 95% confidence level.
We further evaluated the prediction skill of the APCC MME after applying AO correction to each individual model (Fig. 5). The results indicate that statistically significant improvements were achieved across the AO-sensitive regions, particularly over large parts of Eurasia, including both northern and central sectors where original model performance was weakest.
Fig. 5
Effect of AO correction applied to the APCC MME. SAT anomaly correlation coefficients (ACC) for (a) the raw APCC MME, (b) the APCC MME with AO correction, and (c) the difference between (b) and (a) during boreal winter (December–January–February, DJF) for the period 1993/94–2016/17. Green hatches indicate regions where the results are statistically significant at 95% confidence level.
3.4 A Case Study in Hybrid Improvement: The Complementary Roles of Physical Model Development and Statistical Correction
The analysis of the PKNU model versions provides a powerful case study on the synergistic benefits of combining physical model development with targeted statistical post-processing. The three versions represent a controlled experiment demonstrating a two-pronged strategy for improving seasonal forecasts. The baseline PKNU model exhibits the characteristic low predictive skill over central Eurasia, a direct consequence of its poor representation of the AO teleconnection (Fig. 6a). The first step, a physical improvement, is represented by PKNU_I. By incorporating observed sea-ice initialization, this version improves the simulation of regional phenomena like the Barents Oscillation (Sim et al., 2025), resulting in targeted skill gains over northern and southern Eurasia (Fig. 6b). However, this physical enhancement alone fails to solve the larger-scale problem over central Eurasia, which remains dominated by the large-scale AO bias. The final step, PKNU_IAO, demonstrates the power of the statistical correction. By directly addressing the large-scale AO-related bias, this post-processing step yields the broad-scale improvements across the continent that the physics-based changes alone could not achieve (Fig. 6c).
Fig. 6
Winter SAT prediction skill of the PKNU CGCM versions. SAT anomaly correlation coefficient (ACC) for (a) PKNU, (b) PKNU_I, and (c) PKNU_IAO during boreal winter (December–January–February, DJF) for 1993/94–2016/17. Green hatches indicate regions where the results are statistically significant at the 95% confidence level.
We then examined the contribution of each version to the full MME by sequentially including them into a baseline MME that excluded any PKNU model (Fig. 7). While the baseline PKNU offered no meaningful improvement, PKNU_I provided modest gains. The most significant advancement came from PKNU_IAO, which substantially enhanced the MME's predictive skill across Eurasia and parts of North America. This case study makes a clear point: physical improvements and statistical post-processing are not competing approaches but powerful, synergistic complements. The former can enhance skill by improving local processes, while the latter can address the large-scale, systematic biases common across many models, with a hybrid approach yielding the greatest overall gains.
Fig. 7
Contribution of different PKNU model versions to the APCC MME. SAT anomaly correlation coefficients (ACC) during boreal winter (December–January–February, DJF) for 1993/94–2016/17: (a) MME without PKNU, (b) MME with PKNU, (c) MME with PKNU_I, and (d) MME with PKNU_IAO. Panels (e–g) display the differences relative to (a). Green hatches indicate regions where the results are statistically significant at the 95% confidence level.
4. Summary and conclusions
This study demonstrated that a targeted statistical correction of AO-related biases systematically improves winter SAT prediction skill across the 14 models of the APCC MME. Crucially, the magnitude of this improvement was inversely proportional to the models' intrinsic AO prediction skill (r = − 0.65), indicating that the correction provides the greatest benefit to the models with the most deficient representation of Arctic variability. Application to the PKNU model, evaluated as a case study, further underscored the value of a hybrid approach, where combining physical model improvements (via sea-ice initialization) with teleconnection-based post-processing yielded the most substantial gains in forecast skill.
The strong negative correlation between intrinsic AO skill and corrective improvement has profound implications for the seasonal prediction community. It suggests that despite increasing model complexity and resolution, a foundational bias in representing Arctic-midlatitude dynamics persists across a wide range of state-of-the-art models. This finding challenges the notion that simply improving model physics will be sufficient and highlights the immediate value of physically-motivated post-processing as a necessary component of operational forecast systems. This paradox reveals that progress in seasonal forecasting may be stalled not by a lack of model complexity, but by a failure to correctly represent a single, crucial physical mechanism.
Despite these promising results, several limitations must be acknowledged. The effectiveness of the AO correction is ultimately constrained by the predictive skill of the statistical MLR model used to estimate the AO index; if the regression-based reconstruction is itself limited, the improvement will be modest. Furthermore, while this post-processing technique effectively mitigates a key source of forecast error, it does not address the underlying dynamical shortcomings within the models themselves, which may arise from errors in stratosphere–troposphere coupling or Arctic sea-ice feedbacks.
Future work should proceed along two complementary paths. First, post-processing methods must be refined to ensure robustness and multivariate physical consistency, potentially by extending this framework to other key teleconnection patterns such as the El Niño–Southern Oscillation. Second, and more fundamentally, the climate modeling community must prioritize addressing the root causes of the poor AO representation in dynamical models. A dual approach that combines more intelligent post-processing with improved model physics offers the most promising path toward bridging the gap between statistical and dynamical approaches and advancing the field of seasonal climate prediction.
A
A
Conflict of Interest StatementThe authors have no conflicts of interest.
Acknowledgments:
This work was funded by the Korea Meteorological Administration Research and Development Program (grant: RS-2025-02313090). The main calculations were performed using the supercomputing resource of the Korea Meteorological Administration (National Center for Meteorological Supercomputer). We also acknowledge that the APCC Multi-Model Ensemble (MME) Producing Centers for making their hindcast/forecast data available for analysis and the APEC Climate Center for collecting and archiving them and for organizing APCC MME prediction.
A
Data Availability
All data that we used are freely and publicly available as follow links: Japanese 55-year Reanalysis (https://jra.kishou.go.jp/JRA-55/index_en.html#download).
Code Availability
The source code in this study is based on the National Center for Atmospheric Research/University Corporation for Atmospheric Research (NCAR/UCAR) Community Earth System Model (CESM) version 2.1.3 whose code can be acquired from the CESM2 repository (https://github.com/escomp/cesm.git).
REFERENCES
Ahn JB, Lee J, Jo S (2018) Evaluation of PNU CGCM ensemble forecast system for boreal winter temperature over South Korea. Atmosphere 28(4):509–520
Cohen J, Screen JA, Furtado JC, Barlow M, Whittleston D, Coumou D, Jones J (2014) Recent Arctic amplification and extreme mid-latitude weather. Nat Geosci 7(9):627–637
Fadeev RY, Tolstykh MA, Volodin EM (2019) Climate version of the SL-AV global atmospheric model: development and preliminary results. Russ Meteorol Hydrol 44(1):13–22
Gualdi S, Borrelli A, Cantelli A, Davoli G (2020) Mar Chavesmontero, M., Masina, S., … Tibaldi, S. The new CMCC operational seasonal prediction system
Ham S, Lim AY, Kang S, Jeong H, Jeong Y (2019) A newly developed APCC SCoPS and its prediction of East Asia seasonal climate variability. Clim Dyn 52(11):6391–6410
Hersbach H, Bell B, Berrisford P, Hirahara S, Horányi A, Muñoz-Sabater J, Thépaut JN (2020) The ERA5 global reanalysis. Q J R Meteorol Soc 146(730):1999–2049
Juang HMH, Wu TY, Liu PYB, Lin HY, Lee CT, Kueh MT, Lin PL (2024) The First Operational Version of Taiwan Central Weather Bureau’s One-Tier Global Atmosphere–Ocean Coupled Forecast System for Seasonal Prediction. Weather Forecast 39(5):745–763
Kang D, Lee MI, Im J, Kim D, Kim HM, Kang HS, MacLachlan C (2014) Prediction of the Arctic Oscillation in boreal winter by dynamical seasonal forecasting systems. Geophys Res Lett 41(10):3577–3585
Kim H, Lee J, Hyun YK, Hwang SO (2021a) The KMA Global Seasonal forecasting system (GloSea6)-Part 1: operational system and improvements. Atmosphere 31(3):341–359
Kim G, Yang S, Lee WS (2021b) Combined effect of ENSO and AO on winter temperatures of the Korean peninsula on subseasonal time scales. J Clim 34(18):7493–7506
Kobayashi S, Ota Y, Harada Y, Ebita A, Moriya M, Onoda H, Takahashi K (2015) The JRA-55 reanalysis: General specifications and basic characteristics. J Meteorological Soc Japan Ser II 93(1):5–48
Lin H, Merryfield WJ, Muncaster R, Smith GC, Markovic M, Dupont F, Erfani A (2020) The Canadian seasonal to interannual prediction system version 2 (CanSIPSv2). Weather Forecast 35(4):1317–1343
Meleshko VP, Matyugin VA, Sporyshev PV, Pavlova T, Govorkova V, Shkolnik I, Baidin A (2014) MGO general circulation model (version MGO-03 T63L25). Proc. of Voeikov Main Geop. Obs, 571, 5–87
Nakada K, Kovach RM, Marshak J, Molod A (2018) GEOS S2S-2_1 File Specification: GMAO Seasonal and Sub-Seasonal Forecast Output (No. GSFC-E-DAA-TN55924)
Penabad E, Dorel L (2023) Description of System8-v20210101 C3S contribution (available at: https://confluence.ecmwf.int/display/CKB/Description+of+System8-v20210101+C3S+contribution)
Ren HL, Nie Y (2021) Skillful prediction of winter Arctic Oscillation from previous summer in a linear empirical model. Sci China Earth Sci 64(1):27–36
A
Saha S, Moorthi S, Wu X, Wang J, Nadiga S, Tripp P, Becker E (2014) The NCEP climate forecast system version 2. J Clim 27(6):2185–2208Screen JA, Deser C, Smith DM, Zhang X, Blackport R, Kushner PJ, Sun L (2018) Consistency and discrepancy in the atmospheric response to Arctic sea-ice loss across climate models. Nat Geosci 11(3):155–163
Sillmann J, Thorarinsdottir T, Keenlyside N, Schaller N, Alexander LV, Hegerl G, Zwiers FW (2017) Understanding, modeling and predicting weather and climate extremes: Challenges and opportunities. Weather Clim extremes 18:65–74
Sim JH, Kim BM, Lee JG, Lim YK, Kim JH, Kim JH (2025) Sea Ice Initialization and Its Impact on Winter Seasonal Prediction Skill over the Northern Hemisphere in Coupled Forecast System. Journal of Climate
Sim JH, Kwon M, Jang YS, Kim HR, Kim JH, Yang GH, Kim BM (2024) Improving the seasonal forecast by utilizing the observed relationship between the Arctic Oscillation and Northern Hemisphere surface air temperature. Environ Res Lett 19(7):074039
Wedd R, Alves O, de Burgh-Day C, Down C, Griffiths M, Hendon HH, Zhou X (2022) ACCESS-S2: the upgraded Bureau of Meteorology multi-week to seasonal prediction system. J South Hemisphere Earth Syst Sci 72(3):218–242
Williams KD, Copsey D, Blockley EW, Bodas-Salcedo A, Calvert D, Comer R, Xavier PK (2018) The Met Office global coupled model 3.0 and 3.1 (GC3.0 and GC3.1) configurations. J Adv Model Earth Syst 10(2):357–380
Wu T, Yu R, Zhang F, Wang Z, Dong M, Wang L, Li L (2010) The Beijing Climate Center atmospheric general circulation model: description and its performance for the present-day climate. Clim Dyn 34(1):123–147
Yhang YB, Lim CM, Jeong D (2025) APEC climate center multi-model ensemble dataset for seasonal climate prediction. Sci Data 12(1):303