General comments

The paper analyzes the relevant question for risk assessment of the temporal clustering of events. The study is based on insurance loss data covering south-west Germany over the period 1986-2023. The hazard types associated to the losses are flood, storm and hail and they are associated to meteorological data. The hazards flood and storm are further separated into different phenomena since they can derive from different weather conditions: fluvial, pluvial or mixed floods on one hand and large-scale storms and convective gusts on the other hand, which makes sense. Then the methodology is described, and the results analyzed and discussed.

Specific comments

While seasonality is handled both in the loss data and in some hazards’ characterization, it is not considered when identifying the major loss events. In my opinion, because seasonality is clearly identified in the loss distribution, it should be considered in the characterization, otherwise the percentile is computed in mixing different types of losses, which can be misleading. Therefore, I would suggest that the discussion devoted to the loss distribution analysis in section 4 is moved before the major loss events identification in section 2, justifying that the identification should be made on a seasonal basis. Then the same methodologies can be applied for each main season of occurrence and further summarized at the annual scale if necessary.

The identified clustered hazards are physically relevant, which is reassuring, but one may wonder whether such an analysis was really necessary to derive the results. An interesting question regarding these events is the role of decadal variability, which is hard to infer with less than 40 years of observations. The identified clustering may be explained both by the fact that the clustered hazards derive from the same weather situation and by the fact that those weather situations occur more frequently during certain decades compared to others. This should be considered when analyzing trends too.

Technical corrections

Line 18: “Damage by those hazard” : hazards

Line 98: the closing bracket should be removed after 36 000 km²

Line 110: “onlyf” is written instead of “only”

Line 126: “Given the the different environmental conditions” 2 instances of “the”, one should be removed

line 472: “(see Fig 8” the closing bracket is missing

line 570: “It should furthermore not be neglected is that there is a stochastic element”: “is” should be removed

The study examines how clusters of extreme weather events in southwestern Germany lead to higher losses than random occurrences. They analyzed insurance loss data from 1986 to 2023, adjusted for inflation and contract changes, using algorithms and clustering metrics to assess the significance of clustering for different hazards and their combinations. The research shows that clusters, particularly involving hail, floods, and storms, mainly happen in the summer and are linked to increased losses. The study highlights the growing number of extreme weather clusters over 38 years and their impact on risk assessment and insurance. The authors advocate for a holistic approach to hazard and risk analysis, considering the amplified risks of multiple combined hazards, especially in the context of climate change.

The work is of great relevance for the multi-hazard community and makes a step forward towards the understanding of complex risk dynamics. I would recommend the publication of the manuscript if the authors are able to further clarify or justify the following points:

1. I understand the data cannot be made publicly available but the authors should at least share the code so that a reader can follow the detailed steps of the analysis and perhaps reproduce it with different datasets.

2. The clustering for both single and multi hazard only takes into account the temporal dimension. Wouldn't it be better to do a spatio-temporal clustering especially considering that the weather data comes at 1km2 resolution? How the results would change?

3. Can we assume that the percentage of insured assests is constant across the considered region and over time? How would this influence the results of the analysis?

4. It would be interesting to add the vulnerability dimension in the study since the loss declared would depend also on that and not just the hazard number.

5. Why not trying a loss independent clustering too? e.g. based on hazard intensity or other parameters.

6. I would also discuss more why daily insured damage. At a first glance it seems too short as a temporal window and prone to errors, double counting and so on.

7. How is exposure taken into account? Is it assumed to be uniform over the whole region? Somewhat related to quesiton 3. & 4.

I would like the authors to discuss further these points in the main text or at least provide an explanation behind their choices. Having said that, I acknowledge the relevance as well as the technical robustness of the work done by the authors.

The study “Impact-based temporal clustering of multiple meteorological hazard types in southwestern Germany” is a detailed analysis of multi-hazard occurrence and related insured losses in Baden-Württemberg. I find the study well-written and methodologically sound. However, the main issue with the study is its relevance. As noted by the first reviewer, the analysed hazards are mostly clustered because they are physically connected to the same causes. This can be easily derived already from Figure 3 that the hazards are either caused by winter extra-tropical cyclones or summer convective storms. As the events are largely confined to two short seasons, analysing clustering up to 60 days will naturally show strong clustering. One question is how relevant are the losses of clustered events compared to the loss of the “main” event. Were the losses in the past 40 years really clustered, or single major events were strongly dominant and the co-occurring losses (of the same or different type) were not really important? I think the authors should more strongly highlight what they think is their contribution with this study, and especially why clustering is of any actual relevance in the study area in terms of societal and economic impacts, or at least for the insurance sector.

Other potentially major issue I see is indicated at the beginning of section 2: “This study is based on loss data (Sect. 2.1.1) from a building insurance company”. It is understandable that insurance data can’t be shared publicly. However, at minimum, the insurance company needs to be identified by name, and in the data availability section, information needs to be provided how other researchers could apply to that company to also have access to that information. Making any verifiability of the study impossible is against the editorial policies: https://www.natural-hazards-and-earth-system-sciences.net/policies/data_policy.html

Finally, in L532: “Furthermore, this increasing trend is also influenced by non-meteorological factors, which could not be factored in.” It’s not true that it can’t factored in. Exposure growth is major driver. Between 1991 and 2023, the value of fixed assets in Germany, in price-adjusted terms, increased by 68% for buildings and similar amount for other types of assets (as per Destatis database). Many studies on exposure-adjusted losses for different hazards are available. Also, strong exposure growth over the study period could affect event detection, creating an artificial upward trend in number of events.

Minor comments:

L20: the statistic refers to Germany, not Europe. For losses in Europe, I would suggest referring to: https://doi.org/10.2908/CLI_IAD_LOSS

Section 2.2: It would be beneficial to know what is the magnitude of minimum and maximum losses of filtered events. Also, Figure 2 doesn’t have any scale in the axes.

Section 4.1: This section could be much shortened by moving the analysis why certain events are more damaging to the discussion (the last two paragraphs), because it is not the main topic of the paper.

Section 5.2: It would be good to explain to readers unfamiliar with it, that the 21-day window is also taken from insurance practice. Note that you explain this in context of specific 3- and 7-day windows.

Figure 11: Why do you suddenly switch to a 14-day window? Also, the labels shouldn’t obscure the graph.