General Comments:

This manuscripts investigates the predictability of seasonal sea ice in the Pacific-Arctic sector. The authors develop and use a regional linear Markov model with a suite of atmosphere-ocean variables. They find that prediction skill of this model is improved relative to a similar previous pan-Arctic model and that there are key sources of predictability for sea ice in different seasons of the Pacific-Arctic sector. The paper is well-written, the structure makes sense, and the figures are clear. I find the use of this regional statistical model interesting and the results are clear and novel, but I think there are a few flaws in the development of this model that may have an impact on the key results.

The main concern I have is if the authors are missing a key source of predictability for summer Arctic sea ice: sea ice thickness. Numerous studies going back to the seminal work by Blanchard-Wrigglesworth et al., (2011) have shown that sea ice thickness is a key source of predictability for summer Arctic sea ice. In this Markov model, it is excluded, with no clear reason why. I think the authors should add this variable or motivate why it is excluded in a clear fashion. For instance, rather than stating in words why it is not included, it would be nice to see supporting figures that show what happens when it is included. Or, the main figures of the paper should be revised/updated with this variable included. Similarly, the authors exclude subsurface ocean heat content as a variable. This too has been shown to be a key source of predictability for wintertime Arctic sea ice. In the detailed comments below I have suggested how the authors can improve this by including SIT (PIOMAS) or OHC (ORAS5).

Specific Comments:

Data and Methodology:

The Model:

L147-148: It seems this spatial domain (40°N to 84°N, 120°E to 240°E) captures much (if not all) of the Pacific-Arctic sector. But, it feels a bit arbitrary where exactly the longitudinal domains end. For instance, if the same analysis is performed from 90°E to 270°E, are the results similar? Given that the authors base some of this domain choice on regions that exhibit different sea ice variability (i.e., Figure 1), if the authors instead calculated regional sea ice area domains and computed the standard deviation of each region, then grouped the regions by similar magnitudes of variability, will the results be similar? To me this seems like a more precise way of quantifying the Pacific-Sector as regions like the Canadian Archipelago or Laptev Sea could be included or excluded. Basically, I’d like for the authors to conduct this sensitivity analysis and either provide a short explanation of why the results are insensitive to the exact domain or add some supporting information to justify this choice. 

L148-149: Similarly, can the authors perform the same analysis on sea ice concentration from 0 to 100% instead of 15% to 95%? It is unclear why the authors chose this limit. I can understand the 15% threshold (eliminating area and extent), but why 95%? It would be nice to see the analysis performed without these cutoffs and if they differ, an explanation as to why.

L159-161: I can understand why the authors chose this combination of atmospheric and oceanic variables. But, it appears one key variable is missing: sea ice thickness. Why have the authors omitted this variable? Over the past few years, numerous studies have shown that sea ice thickness is a key source of predictability of Arctic sea ice, particularly for the summer. This, to me, is my biggest concern with this manuscript. I’d like the authors to include sea ice thickness (SIT; from PIOMAS most likely) in their analyses, as I think the results will be different. Perhaps this model may even be more skillful.

This also pertains to the ocean domain. What about including sub-surface ocean heat content (OHC)? This has also been shown to be an important source of predictability for Arctic sea ice, particularly on the Pacific side in the wintertime. The authors can include this variability using Ocean Reanalysis System 5 (ORAS5). I think including this variable will improve the weaknesses of the model (as stated in Lines 590-594).

Finally, it is shown that SIC contributes to the trends and gives prediction skill. But, why even include this at all? It would be more interesting if the authors could develop this regional Markov model without SIC so that it is predicting SIC/SIA. Perhaps I am missing the motivation to include this. If that is the case, please add a justification in this section.

L166-186: The introduction of the Markov model could be slightly more clear. In particular, the bolded variables are not present in Eqs. (2)-(4). Does that make them not matrices? As it stands, I think they should be matrices. 

Model construction and assessments:

Construct an optimal model for each season:

L280-283: I think this cross-validation experiment will be greatly impacted by the addition of SIT and OHC (and/or removal of SIC). This section may need to be reworked with these variables included.

Conclusions:

L521: I think this paper would benefit from a clear “Discussion” section and then the “Conclusion” section can be streamlined and the results would be clearer.

Technical Comments:

Introduction:

L17: Add hyphen to Pacfic-Arctic? Also, consider keeping it uniform across the paper: Pacific-Arctic instead of switching to Arctic-Pacific

L41: Consider changing “shrinking” to “decreasing” and adding extent, so that it states “The decreasing Arctic sea ice extent contributes…”

L48: Consider changing to “The rapid retreat of summer Arctic sea ice extent has also created…”

L60: Change “which” to “and”.

L88: This sentence does not make sense. Can the authors reframe this? I am not sure what “is initialized in the spring”. Is this in reference to the spring predictability barrier?

L522: Add hyphen (check hyphens throughout the manuscript)

General Comments:

The authors present a regional sea ice forecasting method using a linear Markov model.  The methodology is variation of an established pan-Arctic version of this model.  The novelty is a focus on the Bering Sea and the Sea of Okhotsk and using varying predictors by season.  The method shows good prediction skill compared to the pan-Arctic version and to an anomaly persistence model.

Overall, this is a good study and is worthy of publication, but some additions would make this paper more compelling.  I would like to see more discussion on the choice of predictors.  There have been many studies on sea ice forecasting, only a couple of which are referenced here.  A more in-depth literature review could be a good place to start when discussing the choice of predictors.  The study area is also a bit confusing.  It is stated that this method forecasts in the Bering Sea and the Sea of Okhotsk, but Figures 1, 2, 6, & 7 show forecasting in most of (if not all) of the Chukchi Sea as well.  Maybe redefine your study area or clarify your delineation of sea ice regions more clearly.

Specific Comments:

Literature Review: You’ve listed a couple previous studies of sea ice forecasting methods, but I think you are missing out on a lot of work that has been done, even quite recent publications.  I think lines 57-87 could benefit greatly from a more in-depth discussion of previous forecasting methods.  Section 2.1 would also benefit from this when discussing your choice of predictors.  See references below:

Andersson, Tom R., J. Scott Hosking, María Pérez-Ortiz, Brooks Paige, Andrew Elliott, Chris Russell, Stephen Law, et al. “Seasonal Arctic Sea Ice Forecasting with Probabilistic Deep Learning.” Nature Communications 12, no. 1 (August 26, 2021): 5124. https://doi.org/10.1038/s41467-021-25257-4.

Chi, Junhwa, and Hyun-choel Kim. “Prediction of Arctic Sea Ice Concentration Using a Fully Data Driven Deep Neural Network.” Remote Sensing 9, no. 12 (December 2017): 1305. https://doi.org/10.3390/rs9121305.

Horvath, Sean, Julienne Stroeve, Balaji Rajagopalan, and William Kleiber. “A Bayesian Logistic Regression for Probabilistic Forecasts of the Minimum September Arctic Sea Ice Cover.” Earth and Space Science 7, no. 10 (2020): e2020EA001176. https://doi.org/10.1029/2020EA001176.

Horvath, Sean, Julienne Stroeve, and Balaji Rajagopalan. “A Linear Mixed Effects Model for Seasonal Forecasts of Arctic Sea Ice Retreat.” Polar Geography 0, no. 0 (October 15, 2021): 1–18. https://doi.org/10.1080/1088937X.2021.1987999.

Wang, Lei, Xiaojun Yuan, Mingfang Ting, and Cuihua Li. “Predicting Summer Arctic Sea Ice Concentration Intraseasonal Variability Using a Vector Autoregressive Model.” Journal of Climate 29, no. 4 (December 8, 2015): 1529–43. https://doi.org/10.1175/JCLI-D-15-0313.1.

For instance, in lines 116-117, is there a reason 850 hPa geopotential height and winds were chosen over other pressure levels?

Methodology:

Lines 94-99: Here you differentiate the Bering Sea and the Chukchi Sea, but as mentioned above it looks like you are indeed forecasting in both areas.  More clarification here is needed.

How are you detrending the model in Section 3.4?  Removing the trend from the data prior to running the model? Detrending the predictions with the observed trend? With the predicted trend? It is difficult to assess how well the detrended model is actually doing without knowing this.

Other:

Lines 324-328: Similar results were found in Horvath et al., 2020, namely that forecasts made in March showed worse skill than those made in January and February.  Perhaps this warrants further analysis.

Technical Comments:

LN 303-305: The fact that more modes are needed during the cold season is repeated.  I suggest combining these sentences to clarify.

LN 505: remove ‘the’.  “In other words,…”