I would like to thank the authors for the substantial effort to revise the paper and to reply to my comments. Unfortunately, I find that several of the points I raised were not adequately addressed. I would like to insist on the following:
- My overall impression is that the authors should make substantial revision of scientific language and make an effort for clearer organisation of the paper.
- Regarding the methods, I still do not fully understand the whole approach (see comments below).
- Finally, I have two general but critical comments:
A. The paper demonstrates results about the costliest storms, including famous cases, but we gain almost no knowledge about the storm characteristics that produce the damages. For instance, we gain no insights into claims’ distance from storms’ center, storms’ intensity and wind speed at claims’ location, seasonality of the clusters and storms, what kind of cyclones form the clusters, whether these clusters are multi-centered cyclones, or distinct storms (like in the case of Lothar). Figure 11 seems to summarize all new knowledge, but it only generally quantifies impacts because of clusters. Still, apart from the obvious (clusters are important for impacts), there is no information that someone could use in a subsequent study (see comment below on Fig. 11).
B. The paper insists too much on the 96h claims window and Generali’s terms. This is fine because it connects to the claims dataset, which seems to be very important in linking cyclones with economic impacts. However, this dataset is not open access, and the 96h clause is Generali-specific (maybe I am wrong). I am aware that the method is adjustable to other temporal claim windows, but this would make the paper only attractive to other insurances with their own contract specificities and, of course, their own proprietary datasets.
Combining these two critical comments together, I believe that the broader impact of the paper is rather weak, especially for the scientific community. I am thus inclined to recommend the rejection of the paper, except if the authors are willing to provide a more detailed association of claims to storms’ physical characteristics so that results are useful to bridge the scientific community with the industry.
Specific comments
All line numbering refers to the manuscript with tracked changes.
1. Presentation
I am afraid I have to repeat the first comment of my original review. While the authors adequately addressed every single comment that I had in my first review, in my view the manuscript suffers from use of plain language and lack of focus on the narrative. The whole manuscript demands important editing rather than addressing single phrases. I provide several examples below.
Line 23: automated scheme -> cyclone tracking method. Please be consistent throughout the manuscript on this terminology.
Line 30: I appreciate this discussion on cyclone tracking schemes, but it seems a bit disconnected from your motivation and inadequate for the beginning of the article. You could move it to the methods (?).
Line 34: I am not sure what makes SSI a “robust framework”. If I am not mistaken, SSI is a proxy for high-impact windstorms, not a framework. But even so, I am not sure what makes it “robust” here. Please elaborate, rephrase, or put into context.
Line 36: “restricted vision” is informal language. Anyway, SSI is anyway irrelevant to the existance or not of a storm.
Line 37: A jet stream can be an atmospheric feature or a large-scale atmospheric condition. “Background dynamical environment” seems a bit uncommon, if not odd.
Lines 40–41: I think that the severity of impacts only depends—from the hazard perspective—on the wind speed intensity. This is irrelevant to the underlying physical conditions. Please rephrase. Actually, the next and last phrase of the paragraph seems to be adequate.
Lines 58–59: Is the reference to Mailler et al. suggesting that “This metric was introduced based on the observation that storm occurrences do not follow a Poisson distribution with constant intensity”? If not, please provide a reference, or else move the citation to the end of this phrase.
Line 66: The article has many phrases like this one which linguistically seem to make sense but, at least to me, I feel that I do not gain much insight into the issue. First of all, “threat” is informal language; second, we do not understand if extreme cyclones are clustered or are part of clustering; and third, “pronounced” is not really giving much qualitative or quantitative information.
Lines 66–70: Awkward use of parentheses.
Lines 75–77: These two phrases seem paradoxical. The first one reads: “claims falling outside an event definition are either not reimbursed..” and the next phrase reads: “..event definitions are typically based on claim date”. If an event is defined by the claim, then how can a claim fall outside of an event?
Lines 72–85: My impression is that in crucial parts of the text, the authors use plain language; therefore, I miss some important aspects of the rationale and motivation of this study. This new paragraph has many examples of inaccurate text. For instance, in the first phrase of the paragraph we read that insurances are interested in the “representation of hazard”. I guess “hazard” is some kind of a measure for wind intensity. Then, to what does the term “event”, which is introduced right after, refer? Does it refer to the date when high wind speeds caused some damages, or to a cyclone itself? If it is the former (and this is why an “event” is defined by the claim), then why do they “lack physical coherence”? What does it mean that events are “adjusted to maximise.. reimbursable claims”? Then it is mentioned “consistent hazard-based definition”: consistent with what? And so on.
Line 81: This phrase seems to suggest that without linking high wind speed to physical drivers (the storms, in our case), it is not possible to anticipate potential impacts in the future. I guess that a 200-year return period of impacts due to windstorms can be estimated by multi-year simulations of wind speed alone.
Lines 83–85: This phrase is cryptic to me. “Hazard representation” in what? In atmospheric modelling? I guess that “accuracy” refers to the minimization of high wind speed bias (provided that the hazard is high wind speed). If this is the case, then how does bias correction strengthen the “physical consistency of event definition”? What do “efficiency”, “reliability”, and “processes” refer to here? Why is wind speed accuracy relevant to this study?
Lines 86–97: This whole paragraph comes too late. It gives the motivation and reasoning of this study. All the above paragraphs make sense to me after I read this one.
Line 90: I am not sure why non-inclusion of storms’ physical characteristics in risk calculation is presented as a limitation. I think that it is simply irrelevant to the risk since hazard is only defined by wind speed.
Line 101: I am not really sure that C3S was designed primarily for reinsurance purposes.
Line 107: What kind of biases? What is the limitation that is covered by this study?
Line 110: “Meteorology” is not a sector.
Lines 111–114: Most of the content of this phrase is not relevant to a “meteorological standpoint” but to risk calculations (i.e., “consequences”, “disentangling impacts”, “conditions contributing to damages”).
2. Methods
Line 137: Please change the reference according to the dataset used. C3S is the service hosting them.
Line 140: Sect. C -> Appendix C
Line 142: If I got it right, the time of dstorm is not necessarily related to impacts, so “impacts date” is not a good way to name the date that a storm reaches 7.5°W. Please revise.
Line 140: What is meant by “well-developed”? Please revise the whole manuscript to avoid plain language and give physical content to the terms you use.
Line 152: I am not sure I understand this ±12h window around cyclones’ center. Does this mean that the gust footprint is defined as the maximum gusts within 1300 km of each track point and within ±12h? That would mean that for some storms there will be an overlap sharing the same maximum values, right? For instance, in Fig. 1 the track points of the two storms are really close to each other (temporally and spatially) and probably they share the same maximum gusts. How often does it happen for at least two storms to share the same wind-gust maximum?
Line 157: Relaxed relative to what? Apologies if I missed it, but I did not find the criteria used to retain or reject a track.
Line 178: are -> area?
Lines 175–180: I am not sure I understand the rationale here. If we consider that a gust footprint is calculated within 1300 km, this is because we believe that important impacts can take place within this radius. However, line 176 suggests that within 700 km one might get more intense winds. Is there a relationship between wind speed and radius? If so, please elaborate. Thereafter, lines 176–179 seem to suggest that this radius is chosen to tune the ratio between individual storms and clusters, while the 4-day window is chosen according to Generali’s contracts. Is there an actual physical reason for the 4-day window and the 700 km radius? If yes, please elaborate. If not, I strongly suggest changing your approach so that clustering is based on physical criteria. Then, you can check how many storm clusters might match the claims.
I appreciate the inclusion of Fig. 2, but I still find it hard to understand. First of all, Fig. 2 refers to an example which comes much later in the text (Section 4.2). Then, line 229 starts the description of the figure from the second histogram. Trying to figure out the first histogram, I deduce from it that there is a considerable number of claims. However, on the right-hand side it reads “1 Claim”. Actually, I am not sure I understand the length of the red lines. In the caption it reads “storm events and their associated number of claims”. Does this mean that the length of a red line corresponds to the number of claims from a single storm? I actually thought that the example here is to associate claims per storm. If so, then the red vertical lines already show this information (?!). Also, I do not understand how red lines are placed between the grey bars. I guess that a red line is placed according to the “dstorm” date? If so, this is not clear. Also, if this is the case, then I do not understand the comment about “good alignment” in line 231. If the claims are declared with a lag-time difference with respect to dstorm, then a “good alignment” might be just a matter of coincidence, right? Finally, I guess that the claims shown in the histogram correspond to the whole of France(?). If so, is it correct to assume that part of the histogram is due to claims in eastern France while the dstorm has to do with the storm being located over the Atlantic Ocean? Actually, if a storm moves, then instead of a red line there should be a “red area”, since while the storm moves towards the east, it captures within its radius more and more regions, right?
I do not understand the last phrase of the caption.
Schematic -> Example
Plots -> histograms?
Line 232: Storms are mentioned here as “large-scale events” but previously it is argued that the methods are tuned to retain small-scale systems.
Line 234: Is there a reference or a basis for the assumption that a storm will result in at least 50 claims?
Line 241: storm date -> dstorm
Line 243: “The performances of the association are evaluated with three metrics, comparing the identified storms to the claim’s local maxima. The local maxima are identified by peaks over the time series of claim count gathering at least 10 claims.” This is an example of phrasing that gives me a hard time understanding the methods. How is the storm defined so that it is compared to a local maximum of claims? What is meant by peaks in the time series of claims? Are time series of claims for all France? I guess only for claims overlapping with the 1300 km radius. If different, please revise.
Lines 244–245: I do not understand these phrases. What is meant by major physical event? Is the physical event the storm? If so, this phrase suggests that there are also secondary or minor storms, and the tuning should capture the “major” one. Why is this assumed and not shown?
The paragraph continues with the definition of the metrics, but actually none is really clear to me (especially the last one in line 251).
Bottom line: how and why is the date that a storm is close to the reference longitude of 7.5°W important to define impacts? It seems completely arbitrary. Why is there no spatial criterion explicitly mentioned in this section? Do claim attributions with spatial criteria with respect to the storms’ area of influence? In fact, I am not really sure that the authors replied to my relevant comment in the previous review that reads “I actually failed to understand several concepts in section 3.1... better understand their necessity and use.”
I also do not understand the rationale of the reply to my question in the previous review: “In your methodological approach,... Could you please clarify.” So regarding the claims, a person might be mistaken by ±1 day on the date of impacts. So would it be reasonable (and much simpler) to simply consider that if an area within 1300 km from a storm center overlaps with a claim’s location within ±1 day, it could be attributed to that storm? In fact, the metrics used in the paper are not adequately explained, the conceptual diagram is quite difficult to understand, and I still lack examples or a good explanation about how a storm being 3 days before reaching the longitude of 7.5°W can be related to claims in France.
Section 3.3, and especially lines 305–310: I understand that the 40 costliest storms are consistent with the strongest ones in a Meteo-France open public dissemination (Meteo-France, 2023), but this only validates that major storms are part of your dataset. This is certainly good, but your method should be expected to include the costliest storms in your datasets anyway (else it would be worrisome). I think the interest here is to see all the weaker storms and the role of clusters in enhancing impacts. Unfortunately, so far we gain no particular insight into the representativity of the captured storms from the whole methodology, or into the weaknesses and advantages of your methods regarding clustered storms. We also gain no insight into the role of smaller storms in producing damages. I also have a bit of a hard time understanding Fig. 6. Actually, “global damage statistics of storms”, “distribution of the total cost per storm, the mean cost per claim and the number of claims per storm”, and “aggregating policy-based costs at the storm level” are not explained. So far, the methods were done to associate Generali’s claims to storms, but apparently things get a bit perplexed with two additional monetary metrics (“policy-based costs” and “cost per claim”).
Section 4
Basically, Sections 4.1 and 4.2 focus on known cases that have been addressed thoroughly in the past. Using known cases is good to support the methodological approach, as it is done here. Thereafter, however, there is just a small section (4.3) to provide new insights into storms’ relationship to impacts, which is rather weak. In fact, results in this section are summarized by Figure 11. This is a rather interesting figure, but it is hard to make use of it for any purpose other than quantifying the importance of clusters. As of now, we do not know the representativity of clusters in claims with respect to individual storms. We have no specific knowledge of the physical characteristics of the storms that produce the damages. We do not know, for instance, whether a loss rank 3 or 4 has a small percentage just because a very intense storms are only marginally affecting France. We also do not have knowledge about the lag dates between the events, the wind intensities that produce the damages, and whether, for instance, claims are collocated in clusters (at least this would provide some insight into recovery).
Line 405: Awkward phrasing. Pretty much all extratropical storms are formed due to wave activity driven by the jet. Please rephrase or remove.
Lines 421–422: This phrase is not really clear. I am sorry for the naive question, but if I got it right, any person will claim insurance within 4 days from the damages. Then, methodologically, a specific storm can be associated to a claim if the claim takes place within +4 days from the first time that the claim’s location is within 1300 km radius of this specific storm. Probably this comment connects to the methodological approach, but I think that this is a much simpler and accurate approach than the one in Section 2.
Line 638: There are no experiments in this paper. Please revise. |
Broadly I like this paper. The findings may be a bit specific to this work and I wonder about the scope for broader applicability but its a nice piece of work which solves a clear problem in working with loss data.
My more specific comments are therefore mostly minor:
17. This is a global figure? Seems very high for Europe, even if Economic rather than Insured loss.
25-29. I Don't entirely understand what this bit is saying.
45-49. Seems well justified.
Figure 1 caption - Should 'different by' be 'defined by'?
120-130 - Interesting discussion of clustering metrics and comments on how a clear definition has not been fixed, or at least not used within the insurance industry. From a loss perspective is it useful to split two clusters affecting France at the same time into separate clusters if they do not overlap?
155-159 Clearly a good dataset with locations recorded. Postcode level is surely fine for ERA5 resolution hazard.
164 - Good observation and an important problem to address.
174 - I can believe this but can you give a bit more explanation/justification. It reads slightly like you argue that the loss data over-represents big storms then remove the little storms yourself. From figure 2a I interpret that lowering N leads to more claims dates than storm dates which is a different problem.
203 - 'Enables the capture'
Section 2 - Overall I find this interesting. Loss data is often focused around extreme events for vulnerability development and this study is an opportunity to work with a sufficiently complete loss history that the impact of smaller events is observed.
246 - How typical Generali's portfolio is of the French market is crucial here. I expect the spread of risks cannot be shown but some commentary could be made as to whether the missing events were very focused on a particular region etc. Some commentary on the exposure comes in 271.
Fig 4 - Is there a relationship between total storm count and cluster counts? The grey bars corresponding to less clustered years appear to group in the 2000s which is also a period of low overall activity based on the red dots.
Section 4.1 seems like the most crucial results, using the dataset and methods discussed in much of the rest of the paper to draw conclusions.
Fig 5. I would expect b) cost per claim to be similar, as observed. But do a) and c) depend on the definition of a storm? If your catalogue contains more small storms the total cost and total claims would be lower, as is observed.
Fig 6. Could we see the spread of losses within clusters for different cluster lengths? For example in a 3 storm cluster is the second ranked loss still 26%?
Section 4.2 shows some value to the method. Often losses from these storms are indistinguishable.
Fig 7/8 what is the resolution of the loss contour maps? The scales go to 6k and 5k respectively but within what area are 5k claims observed? Also there are some white dots that look like they hit Paris that don't seem to be explained.
365-368 Losses are often due to debris, hence consecutive storms may impact losses. This is very difficult for the insurance industry to study and datasets such as this may enable that.
Section 4.3 is also interesting. As stated in 384 Klaus is often treated as a single event.