the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 08 Aug 2025
Targeted Teleconnections and their Application to the Postprocessing of Climate Predictions
Abstract. The demand for skillful climate predictions on subseasonal-to-multidecadal time scales is rising almost by the day, not least because the growing renewable energy sector, but also many other important socio–economic sectors are vulnerable to climate variations. Large scale atmospheric patterns in the North–Atlantic European sector, so-called teleconnections, are well known to have major influence on European climate conditions. For that reason there exists a wide variety of hybrid dynamical–statistical applications, which combine dynamical model output with teleconnections in one way or another to improve the rather modest predictive skill of state-of-the-art dynamical climate forecasts over Europe. The potential improvement generated by these kinds of postprocessing methods is naturally limited by the strength of association between the circulation patterns and the local climate parameters. We propose a statistical technique to retrieve atmospheric patterns—targeted teleconnections—that are maximally predictive for a given climate parameter in a region of choice so as to optimize the potential of statistical postprocessing. The possibility of improvement in forecast skill induced by the implementation of targeted teleconnections is demonstrated in four applications.
- Preprint
(9848 KB) - Metadata XML
-
Supplement
(36504 KB) - BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on egusphere-2025-3664', André Düsterhus, 09 Sep 2025
-
AC1: 'Reply on RC1', Clementine Dalelane, 03 Nov 2025
First of all, we would like to thank Referee 1 for engaging in the reviewing process and for providing his valuable comments, which gave us the opportunity to make our perspective on the mission of statistical climatology explicit.
We will answer the comments, one by one, in detail in the following (original comments in italics).
Review of Dalelane et al.: "Targeted Teleconnections and their Application to the
Postprocessing of Climate Predictions"The manuscript focusses on the application of Partial Least Squares Regression (PLS) in various application of reanalysis evaluation and seasonal climate prediction.
In general, the focus of the manuscript is not clear. It takes for vast sections the form of a statistical paper, especially with the section on related methods. It shows then how it can be applied in various applications, but those seem not to be chosen necessarily to show the strength and weaknesses of the method. As these applications are quite complex for those readers not familiar with it, a proper structure is missing to help a reader to not only understanding what has been done, but what it explains this for the statistics used. Generally, such a paper would not be suitable for the chosen Journal, as it usually addresses physical climate or weather research. Perhaps another journal would therefore be more appropriate (as the about-page of "Weather and Climate Dynamics state", perhaps "Nonlinear Processes in Geophysics is a better fit).
Answer: The scope of WCD is: “the dynamics of extreme weather events (case studies and climatological analyses); weather system dynamics in tropical, midlatitude and polar regions; interactions of atmospheric flows with cloud physics and/or radiation; links between the atmospheric water cycle and weather systems; tropical-extratropical and midlatitude-polar interactions; atmospheric teleconnections and stratosphere-troposphere coupling; boundary-layer dynamics and coupling to land, ocean and ice; atmospheric variability and predictability on time scales from minutes to decades; storm track and Hadley cell dynamics; role of atmospheric dynamics in palaeoclimate and climate change projections; and other aspects of weather and climate dynamics. Theoretical studies, idealized numerical studies, full-physics numerical studies, and diagnostic studies using (re)analysis and/or observational data are welcome.“ It seems to us that a paper that diagnoses atmospheric teleconnections using statistical methods and later on uses these teleconnections to improve climate forecasts does fit into WCD’s scope. But this decision is of course up to the editors. It definitely does not fit into “Nonlinear Processes in Geophysics”, because PLS is a linear method.
Should the target be to publish it in the chosen journal, then a major rewrite would be required. It would need to focus more on the physical arguments, why the statistical results are valid. Up to now the authors try to circumvent these discussions by pointing to a couple of other references. Usually this would be fine, but as the authors try here to discuss not only the typical winter or summer season, but year round, the literature would require a much better backup for these results as currently provided. What are the physical reasons that a reader should trust the statistical result? Also, due to the complexity of the four applications, it lacks the general option to reproduce the results by a reader wishing to do so.
Answer: It is true that the authors do not intend to explore the physical mechanisms behind their findings and are indeed not in a position to do so. It is also true that the literature as of now lacks some of the relevant discussions. But the findings regarding targeted teleconnections are there, they were first achieved by one author (DWD), and then replicated by another group of authors (AEMET). The results for winter and summer do correspond to expectations based on existing literature, so it is not too far-fetched to give the results for less well-explored seasons at least some credence. Whereas the physical mechanisms of the teleconnection are the subject of many studies for winter and at least some for summer, the transition seasons are less well-explored. A detailed analysis of the physical pathways in spring and autumn should therefore be performed in the future to confirm the teleconnections.
The replication of the targeted teleconnections is in any case relatively straight forward, as the data (ERA5) and the method (PLS) are easily accessible. The replication of the seasonal forecast postprocessing is admittedly less easy, but it serves only as a suggestion that might stimulate the reader’s phantasy regarding the usefulness of targeted teleconnections, which is certainly not exhausted by postprocessing of climate forecasts.
In conclusion, as a reviewer I am not of the opinion that the current manuscript is suitable for publication in the current journal. I refrain from outright suggesting its rejection, but would instead suggest, that the authors follow up with a major revision addressing the problems of potential replicability, a better structure and a clearer focus on what the manuscript should be about. Such a complex topic would require a guiding hand to an author and this is up to now not given.
See below further more detailed comments on some major points.
Further comments:
Affiliations: The numbers are not in order, 4 comes before 2.changed
54: "But ML is just another name for statistics" -> ML is a statistical method, but not another name for it. Do the authors mean statistical downscaling?We changed the whole paragraph to: “There has recently been a flurry of downscaling models based on machine learning (ML). ML is a kind of highly nonlinear and very data intensive branch of statistics, and the applications suffer from similar limitations”
But we didn’t mean statistical downscaling. “Machine Learning” is a field of statistics that is only vaguely defined. The term is used with quite different meanings. The greatest common denominator seems to be computer-intensive, often non-linear statistical methods for regression and classification. The term is seemingly meant to create a fictional distance to so-called “traditional statistics”.
134: Direct citation not necessary. Paraphrasing of the content would be sufficient.We understand that paraphrasing makes sense in many contexts. But there is such vast literature on PLS that all possible formulations are virtually exhausted. By direct citation, we at least avoid the allegation of plagiarism.
165: Direct citation not necessary. Paraphrasing of the content would be sufficient.The same as above.
169: Section 3.2: Why such detail on related methods? They either will be applied in the manuscript, or are part of the introduction or discussion as context. How it is solved here is not really understandable.This is to highlight the relationships between the methods, some of which may be known to some readers. It should help to understand the overall similarity of all SVD-related methods, but also their subtle differences making them more or less appropriate for various purposes.
173: Direct citation of such length are inappropriate. I would strongly ask the authors to use their own words.The same as above.
198: The introduction section of this chapter focuses again on things not done in this manuscript and on statistical details. Also such things would belong to introduction or discussion.This discussion was added as a subsection to section 3.
212: I strongly suggest to divide the section 4.1 into subsections, to allow the reader to follow the authors step by step in their arguments. Currently the description of a method, the application of it on a specific application and the result discussion are merged here.The section was divided into two subsections.
223: The authors claim they found no difference, but if this should be a necessary sensitivity test, then it should be properly addressed and statistically quantified. Just visual inspection is not a valid evaluation method.The whole procedure (PLS, statistical estimation of indices, subsampling, and downscaling) was executed in parallel for the two options to detect differences. It was not included in the paper for reasons of space. The paper is admittedly already quite lengthy.
232: "Let us repeat" -> This sentence indicates that the whole section is quite unstructured and complex to follow. I suggest a sketch, which allow a systematic description of the procedure, which is currently not given. Warnings and discussions can then be added at the end of the description or in the discussion.We have divided 4.1 into two subsections. A flow chart for PLS is not really warranted, because it is a standard tool. A flow chart of the whole postprocessing procedure can be found in the supplement.
236: "We decided to regress..." -> Can the number two be justified and quantified by a sensitivity test?
As we wrote “We decided to regress T2M on only two teleconnections, because when dealing with indices estimated from samples (either dynamical climate model output or observations) the trailing indices are less robust. And the trade-off between additional explanatory power and estimation uncertainty quickly tilts to the negative”. Two teleconnections is the lowest number that gives satisfactory results in this application. This number can of course vary with application. Any user has to do their own sensitivity tests.
248: This would need a proper discussion on the physical interpretation of the results of the statistical approach. Claiming that the NAO and EAP are the same and can be interpreted the same is not covered by literature. As those discussions in literature usually focus on DJF and JJA, it would require a detailed analysis at this point. Especially the patterns in MAM and SON need a deeper physical discussion.We reformulate “The DJF-first and second PLS-circulation patterns are reminiscent of NAO and EA, with the southern center of action of the first PLS pattern slightly weaker than in the first EOF pattern. The positive anomalies of the first two PLS patterns are both shifted eastwards, which seems plausible given the location of the target region. The respective T2M winter patterns also correspond approximately to the well-known impact of the EOF modes on European T2M (Simpson et al., 2024). Whereas the physical mechanisms of the teleconnection are the subject of many studies for winter and at least some for summer, the transition seasons are less well-explored. A detailed analysis of the physical pathways in spring and autumn should therefore be performed in the future to confirm the teleconnections.”
264: References are up to here sorted in increasing year of publishing, here it is the other way around.changed
FIG 1: Caption does not state which dataset was used. I assume ERA 5.Added to Fig.1 and Fig.2.
279: Here sub-selection is introduced. While a short section on this approach can be found in the introduction, for a reader unfamiliar with this approach, this section will not be helpful. While it is in a section of what can all be achieved with the statistical approach, it does not provide a structured way of guiding a reader through these complex topics. It confronts the reader with Fig. 3, which neither makes statements of significances, nor improvement by this new approach. This section requires therefore a much better structure to clearly allow the reader to access what was done, how it was done and the opportunity to replicate the results. All is not sufficiently solved in this draft.We reformulate “In DWD, seasonal forecasts of GCFS2.1 with initialization dates from September to December are postprocessed with a subsampling algorithm. That means circulation indices are calculated wrt. given circulation patterns based on the seasonal mean MSLP field of each forecast ensemble member. Subsequently, a given (small) number of forecast members is selected according to their closeness to an empirical first guess (statistical estimation based on observation data) of the future circulation indices, which form the so-called subensemble.
As regards Fig.3: it does indeed show absolute values of MSESS wrt. climatology. We have abstained from marking significances (which we nevertheless did calculate using the Jackknife) for the following reason. In the chosen color scale, the skill difference between the full ensemble and all subensembles is clearly discernible beyond any doubt. (The skill of the full ensemble is itself already significant over the area of Germany, we added a comment to the text.) Furthermore, the improvement from a subensemble based on EOFs to a subensemble based on PLS is also clearly visible for SON, but in contrast not in DJF. Significance tests are a useful tool to discriminate in doubtful cases, but the presented cases are not doubtful at all. We might have needed a significance test to decide whether in DJF the targeted subensemble is more skillful than the untargeted one. But we don’t make such a claim. The actual figures of skill improvement in Fig.3 are not important.
We might add that in such complicated frameworks as climate predictions, significance tests are plagued by a multiplicity of unfulfilled assumptions, both in the case of parametric as well as in the case of resampling tests, stationarity and serial independence being only some of them. For serial dependence, the widely used block bootstrap and auto-correlation corrected t-Test are only weak remedies, because their effective degrees of freedom are much lower than the sample size, so higher sample sizes would be required. And for instationarity, there is no remedy at all expect modelling it explicitly. In total, we should always distrust small improvements as compared to a reference forecast, whereas large improvements, that strike the eye in a suitable color scale, do probably have an element of significance upon themselves.
282: MSESS is introduced without reference and without proper introduction of what is a good or bad score.We added the formula along with some explanation.
292: Exist here a reference?The mentioned estimator has not yet been published. But the essential information lies in rows (b) and (c) anyway.
300: The text talks about improvement in Fig. 4, but as I can evaluate the caption what is shown is absolute values. From absolute values it is not possible to derive information on the difference, so if the aim is to talk about those, I would strongly suggest to the authors showing those.Again, the differences are clearly discernible beyond any doubt.
Fig. 5: Everything is very small in this plot. Readability in this form not given.This plot was cut along with the subsection about multi-model ensemble weighting.
-
AC1: 'Reply on RC1', Clementine Dalelane, 03 Nov 2025
-
RC2: 'Comment on egusphere-2025-3664', Anonymous Referee #2, 26 Sep 2025
Targeted Teleconnections and their Application to the Postprocessing of Climate Predictions by Clementine Dalelane et al.
General comments:
This work proposes the employment of a multivariate regression analysis technique known as Partial Least Squares Regression (PLS) to various applications in climate science. Although there are several other statistical methods proposing ways to obtain targeted-teleconnections in reconstructing surface variables over Europe, this work adds PLS to the list.
The manuscript is well written by organizing the work into different sections. Nevertheless, the overuse of various applications to demonstrate the robustness of the method and the use of complex sentences in the results section have not helped in conveying the intended message to the reader. To improve the readability of the manuscript, I suggest splitting this work into two parts: Part-A detailing the methods through the use of just one elaborative application (for example, to improve seasonal forecasts of temperature and precipitation over Germany); Part-B showcasing all the other applications in more detail. Keeping this in mind, I suggest a major revision of the manuscript before I can recommend it for publication.
Specific comments:
1. Introduction:
- State the novelty of the work explicitly in the introduction.
- This work is missing some important citations. Please find the references below for more details.
2. Methods/Results: The scientific contibution of this work increases significantly if PLS is compared and contrasted with other popular multivariate techniques such as CCA or MCA or RDA. Given that this route takes long, I let the authors make a decision about it.
3. L3-4: Large-scale atmospheric patterns are not teleconnections by themselves. As you have mentioned, large-scale climatic modes of variability can influence surface weather over Europe and elsewhere. It is because of this link they are called teleconnections. Kindly reformulate.
4. L110: Why use MSLP instead of any other upper-level field such as Z500 (mid-troposphere)/Z200 (near tropopause)? I have personally witnessed that Z500 forecasts are more skillful than MSLP forecasts. Nonetheless, the trend prevalent in Z500 fields could make statistical downscaling difficult. I recommend the following investigations:
- Could you compare the skill of MSLP and Z500 forecasts to justify the choice of your predictor?
- Could you compare the trend of Z500 and MSLP?
When comparing the gain in skill using Z500 over MSLP to the complexity in dealing with trend using Z500, you could fairly justify your choice of the predictor.
5. L110-111 and L116-117: Since your validation period (i.e., hindcasts between 1990 and 2020) is already included in training (i.e., in ERA5 between 1951 and 2020), does it not add artificial skill to the statistical forecasts?
6. L117-118: Which method did you use to upscale ERA5 reanalysis onto 1° grid? For information, the S2S4E project conducted a study testing all the available method of regridding on several variables and have made specific recommendations on the choice of optimal methods for different variables (for example, bilinear for T2M and conservative for PR). Kindly state the methods you have employed with justification.
7. L125-126: I suggest discussing the details about detrending here instead of in lines L220-225.
8. L230-231: Since you are relying on (more skillful) MSLP forecasts to reconstruct (less skillful) T2M, did you apply the PLS method lead time by lead time as well? It is important because the forecast skill degrades with lead time.
9. L236-239: The use of just 2 leading components appears arbitrary. Did you notice a drop in the coefficient of determination with the use of additional vectors? Could you plot this figure (either as a response or in the appendix)?
10. Section 4.3.4: The use of “T2M-targeted teleconnections” is not very clear to me in this section. Did you use already improved sub-ensemble of T2M predictions (reconstructed using MSLP) to select specific members of wind speed and solar radiation predictions? Could you please elaborate?
Technical corrections:
1. L121: ‘e’ missing in none.
2. L288: teleconnections
References:
- Michelangeli, P., R. Vautard, and B. Legras, 1995: Weather Regimes: Recurrence and Quasi Stationarity. J. Atmos. Sci., 52, 1237–1256, https://doi.org/10.1175/1520-0469(1995)052<1237:WRRAQS>2.0.CO;2.
- Buizza, Roberto, and Martin Leutbecher. "The forecast skill horizon." Quarterly Journal of the Royal Meteorological Society 141.693 (2015): 3366-3382.
- Lledo, Llorenç, and Francisco J. Doblas-Reyes. "Predicting daily mean wind speed in Europe weeks ahead from MJO status." Monthly Weather Review 148.8 (2020): 3413-3426.
- Büeler, D., Ferranti, L., Magnusson, L., Quinting, J.F. & Grams, C.M.(2021) Year-round sub-seasonal forecast skill for Atlantic–European weather regimes. Q J R Meteorol Soc, 147(741, 4283–4309. Available from: https://doi.org/10.1002/qj.4178.
Citation: https://doi.org/10.5194/egusphere-2025-3664-RC2 -
AC2: 'Reply on RC2', Clementine Dalelane, 03 Nov 2025
First of all, we would like to thank Referee 2 for engaging in the reviewing process and for providing his extraordinarily valuable comments, which will help to improve our manuscript considerably. It was a great pleasure for us to discuss the manuscript with an expert reviewer as competent in the field as this one.
We will answer the comments, one by one, in detail in the following (original comments in italics).
Targeted Teleconnections and their Application to the Postprocessing of Climate Predictions by Clementine Dalelane et al.
General comments:
This work proposes the employment of a multivariate regression analysis technique known as Partial Least Squares Regression (PLS) to various applications in climate science. Although there are several other statistical methods proposing ways to obtain targeted-teleconnections in reconstructing surface variables over Europe, this work adds PLS to the list.
The manuscript is well written by organizing the work into different sections. Nevertheless, the overuse of various applications to demonstrate the robustness of the method and the use of complex sentences in the results section have not helped in conveying the intended message to the reader. To improve the readability of the manuscript, I suggest splitting this work into two parts: Part-A detailing the methods through the use of just one elaborative application (for example, to improve seasonal forecasts of temperature and precipitation over Germany); Part-B showcasing all the other applications in more detail. Keeping this in mind, I suggest a major revision of the manuscript before I can recommend it for publication.
Answer: We agree to the suggestion of splitting the paper in two parts. As a small substitute, we would like to introduce at least some side notes wrt. the applications:
In section 3.3, we added the sentence “Apart from atmospheric teleconnections, another obvious candidate is the ocean-atmosphere teleconnection between SST and SLP or Z500 that has recently gained some attention in \citet{Kolstad2024, Patterson2024, Patrizio2025}.”
In section 4.2.1: “Surprisingly, or perhaps not so much, it turned out that the same T2M-based subensemble shows also favorable skill in some other variables: 100m wind speed, solar radiation and renewable energy production. Another application in the same vein would be an ensemble weighting wrt. targeted teleconnections similar to \citet{Sanchez-Garcia2019}.”
Specific comments:
- Introduction:
- State the novelty of the work explicitly in the introduction.
“the PLS has never been used in this context” has been added to line 91
- This work is missing some important citations. Please find the references below for more details.
Although modes of circulation and weather regimes address the same phenomenon, they do it from different perspectives and the resulting concepts are quite different. While modes of circulation, however constructed, decompose the atmospheric field into a superposition of various components, which are more or less active at the same time, weather regimes divide a set of fields into a number of subsets with only one regime active at the same time. Weather regimes do present a certain potential for predictability, as this potential stems from the atmosphere itself. But this paper is concerned with modes of circulation, so I would prefer not to mingle in weather regimes, and leave references Michelangeli et al.(1995) and Büeler et al. (2021) aside. Strictly speaking, the MJO states in Lledó&Doblas Reyes (2020) are also weather regimes, and although they do have predictive power upon wind speed in Europe, they have existed in their own right before this power was discovered and cannot be regarded a case of targeted construction. As for Buizza&Leutbecher (2015), we thank the referee for this useful reference that fits very well into the discussion.
- Methods/Results: The scientific contribution of this work increases significantly if PLS is compared and contrasted with other popular multivariate techniques such as CCA or MCA or RDA. Given that this route takes long, I let the authors make a decision about it.
We admit that this would be an interesting endeavor. But time limitations are prohibitive.
- L3-4: Large-scale atmospheric patterns are not teleconnections by themselves. As you have mentioned, large-scale climatic modes of variability can influence surface weather over Europe and elsewhere. It is because of this link they are called teleconnections. Kindly reformulate.
Reformulated “Large scale atmospheric patterns in the North--Atlantic European sector are well known to have major influence on European climate conditions, these links being called teleconnections.”
- L110: Why use MSLP instead of any other upper-level field such as Z500 (mid-troposphere)/Z200 (near tropopause)? I have personally witnessed that Z500 forecasts are more skillful than MSLP forecasts. Nonetheless, the trend prevalent in Z500 fields could make statistical downscaling difficult. I recommend the following investigations:
- Could you compare the skill of MSLP and Z500 forecasts to justify the choice of your predictor?
- Could you compare the trend of Z500 and MSLP?
When comparing the gain in skill using Z500 over MSLP to the complexity in dealing with trend using Z500, you could fairly justify your choice of the predictor.
Again, these are very interesting suggestions, which certainly deserve thorough investigation, but resources are limited. The choice of MSLP over Z500 is a historic one going back ultimately to Dobrynin et al. (2018).
- L110-111 and L116-117: Since your validation period (i.e., hindcasts between 1990 and 2020) is already included in training (i.e., in ERA5 between 1951 and 2020), does it not add artificial skill to the statistical forecasts?
This is the ubiquitous dilemma of statistical climatology. To alleviate it, the statistical forecasts used to subselect the hindcasts are calculated on the basis of observation previous to the forecast date only, again going back to Dobrynin et al. (2018).
- L117-118: Which method did you use to upscale ERA5 reanalysis onto 1° grid? For information, the S2S4E project conducted a study testing all the available method of regridding on several variables and have made specific recommendations on the choice of optimal methods for different variables (for example, bilinear for T2M and conservative for PR). Kindly state the methods you have employed with justification.
I am not able to locate the results on regridding on the S2S4E website. I would be grateful if you could provide a link. As to the method used, it was conservative for all variables. Although the findings of S2S4E are surely authoritative, the differences for variables like T2M and SLP are minor.
- L125-126: I suggest discussing the details about detrending here instead of in lines L220-225.
Discussion relocated to section 2.1
- L230-231: Since you are relying on (more skillful) MSLP forecasts to reconstruct (less skillful) T2M, did you apply the PLS method lead time by lead time as well? It is important because the forecast skill degrades with lead time.
The PLS is executed on ERA5 data, not on hindcast data. The relationships between SLP and T2M are simultaneous depending only on the 3-month period. In contrast, the construction of the statistical estimate of future circulation indices based on observed SST (not subject of this paper) is indeed done separately for each lead time and start date.
- L236-239: The use of just 2 leading components appears arbitrary. Did you notice a drop in the coefficient of determination with the use of additional vectors? Could you plot this figure (either as a response or in the appendix)?
We experimented extensively with this parameter. The coefficients of determination do of course increase beyond the second component. With an increasing number of components, the two coefficients of determination (EOF and PLS) eventually even converge. The trouble is that the indices have to be estimated for operational implementation and the estimation error grows with the order of the component. The choice of just two components is due to the fact that the overall skill of the whole procedure barely increases, when three components are used. Furthermore, this also depends on the dynamical model relationship SLP->T2M. We noticed surprising differences between two consecutive versions of GCFS.
- Section 4.3.4: The use of “T2M-targeted teleconnections” is not very clear to me in this section. Did you use already improved sub-ensemble of T2M predictions (reconstructed using MSLP) to select specific members of wind speed and solar radiation predictions? Could you please elaborate?
No, we used just the same subsensemble that was selected on the basis of T2M-targeted teleconnections. The improvement is incidental, but should occur whenever similar physical processes are at work in the atmosphere AND the relationship is realistically represented in the model. For wind and radiation this seems to be the case, for pr it is definitely not.
Technical corrections:
- L121: ‘e’ missing in none. corrected
- L288: teleconnections corrected
References:
- Michelangeli, P., R. Vautard, and B. Legras, 1995: Weather Regimes: Recurrence and Quasi Stationarity. Atmos. Sci., 52, 1237–1256, https://doi.org/10.1175/1520-0469(1995)052<1237:WRRAQS>2.0.CO;2.
- Buizza, Roberto, and Martin Leutbecher. "The forecast skill horizon." Quarterly Journal of the Royal Meteorological Society 141.693 (2015): 3366-3382.
- Lledo, Llorenç, and Francisco J. Doblas-Reyes. "Predicting daily mean wind speed in Europe weeks ahead from MJO status." Monthly Weather Review 148.8 (2020): 3413-3426.
- Büeler, D., Ferranti, L., Magnusson, L., Quinting, J.F. & Grams, C.M.(2021) Year-round sub-seasonal forecast skill for Atlantic–European weather regimes. Q J R Meteorol Soc, 147(741, 4283–4309. Available from: https://doi.org/10.1002/qj.4178.
Citation: https://doi.org/10.5194/egusphere-2025-3664-AC2
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 827 | 98 | 20 | 945 | 43 | 46 | 41 |
- HTML: 827
- PDF: 98
- XML: 20
- Total: 945
- Supplement: 43
- BibTeX: 46
- EndNote: 41
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
Review of Dalelane et al.: "Targeted Teleconnections and their Application to the
Postprocessing of Climate Predictions"
The manuscript focusses on the application of Partial Least Squares Regression (PLS) in various application of reanalysis evaluation and seasonal climate prediction.
In general, the focus of the manuscript is not clear. It takes for vast sections the form of a statistical paper, especially with the section on related methods. It shows then how it can be applied in various applications, but those seem not to be chosen necessarily to show the strength and weaknesses of the method. As these applications are quite complex for those readers not familiar with it, a proper structure is missing to help a reader to not only understanding what has been done, but what it explains this for the statistics used. Generally, such a paper would not be suitable for the chosen Journal, as it usually addresses physical climate or weather research. Perhaps another journal would therefore be more appropriate (as the about-page of "Weather and Climate Dynamics state", perhaps "Nonlinear Processes in Geophysics is a better fit).
Should the target be to publish it in the chosen journal, then a major rewrite would be required. It would need to focus more on the physical arguments, why the statistical results are valid. Up to now the authors try to circumvent these discussions by pointing to a couple of other references. Usually this would be fine, but as the authors try here to discuss not only the typical winter or summer season, but year round, the literature would require a much better backup for these results as currently provided. What are the physical reasons that a reader should trust the statistical result? Also, due to the complexity of the four applications, it lacks the general option to reproduce the results by a reader wishing to do so.
In conclusion, as a reviewer I am not of the opinion that the current manuscript is suitable for publication in the current journal. I refrain from outright suggesting its rejection, but would instead suggest, that the authors follow up with a major revision addressing the problems of potential replicability, a better structure and a clearer focus on what the manuscript should be about. Such a complex topic would require a guiding hand to an author and this is up to now not given.
See below further more detailed comments on some major points.
Further comments:
Affiliations: The numbers are not in order, 4 comes before 2.
54: "But ML is just another name for statistics" -> ML is a statistical method, but not another name for it. Do the authors mean statistical downscaling?
134: Direct citation not necessary. Paraphrasing of the content would be sufficient.
165: Direct citation not necessary. Paraphrasing of the content would be sufficient.
169: Section 3.2: Why such detail on related methods? They either will be applied in the manuscript, or are part of the introduction or discussion as context. How it is solved here is not really understandable.
173: Direct citation of such length are inappropriate. I would strongly ask the authors to use their own words.
198: The introduction section of this chapter focuses again on things not done in this manuscript and on statistical details. Also such things would belong to introduction or discussion.
212: I strongly suggest to divide the section 4.1 into subsections, to allow the reader to follow the authors step by step in their arguments. Currently the description of a method, the application of it on a specific application and the result discussion are merged here.
223: The authors claim they found no difference, but if this should be a necessary sensitivity test, then it should be properly addressed and statistically quantified. Just visual inspection is not a valid evaluation method.
232: "Let us repeat" -> This sentence indicates that the whole section is quite unstructured and complex to follow. I suggest a sketch, which allow a systematic description of the procedure, which is currently not given. Warnings and discussions can then be added at the end of the description or in the discussion.
236: "We decided to regress..." -> Can the number two be justified and quantified by a sensitivity test?
248: This would need a proper discussion on the physical interpretation of the results of the statistical approach. Claiming that the NAO and EAP are the same and can be interpreted the same is not covered by literature. As those discussions in literature usually focus on DJF and JJA, it would require a detailed analysis at this point. Especially the patterns in MAM and SON need a deeper phyiscal discussion.
264: References are up to here sorted in increasing year of publishing, here it is the other way round.
FIG 1: Caption does not state which dataset was used. I assume ERA 5.
279: Here sub-selection is introduced. While a short section on this approach can be found in the introduction, for a reader unfamiliar with this approach, this section will not be helpful. While it is in a section of what can all be achieved with the statistical approach, it does not provide a structured way of guiding a reader through these complex topics. It confronts the reader with Fig. 3, which neither makes statements of significances, nor improvement by this new approach. This section requires therefore a much better structure to clearly allow the reader to access what was done, how it was done and the opportunity to replicate the results. All is not sufficiently solved in this draft.
282: MSESS is introduced without reference and without proper introduction of what is a good or bad score.
292: Exist here a reference?
300: The text talks about improvement in Fig. 4, but as I can evaluate the caption what is shown is absolute values. From absolute values it is not possible to derive information on the difference, so if the aim is to talk about those, I would strongly suggest to the authors showing those.
Fig. 5: Everything is very small in this plot. Readability in this form not given.