Storms are physically characterised by a storm track that follows the trajectory of the ETC. While various tracking algorithms effectively capture the general patterns of storm tracks, the choice of methodology significantly influences the number and characteristics of detected ETCs . Each tracking algorithm is developed based on a specific conceptualisation of what best defines a cyclone, such as mean sea level pressure (MSLP) or relative vorticity at 850 hPa (ζ850). Additional criteria on the total length of the track, duration or intensity of the storms are also applied to constrain the set of tracks detected. Tracking algorithms typically yield consistent results for intense cyclones and during phases of intense development within their life cycle. However, larger discrepancies arise during the genesis and lysis stages of ETC . Differences in the representation of the life cycle are particularly pronounced for weaker, slower-moving, and short-lived ETCs . Additionally, absolute cyclone counts are highly sensitive to the chosen tracking methodology, which can lead to differences across studies . We use the object-oriented TRACK algorithm to obtain tracks over the Northern Hemisphere from ERA5 data . The ERA5 dataset has a horizontal resolution of 0.25°×0.25°, corresponding to 18 km at 50° N, and we use data covering the 1979–2024 period from October to March.

To track ETCs, the input data is smoothed over a T42 resolution (approximately 2.8°) to filter out noise. ETCs are then identified based on 6-hourly values of ζ850 connected with nearest neighbour search. No conditions are set on the total length of the trajectories or on the minimum value of the maximum vorticity along the track, ζ850max. We apply a spatial constraint to retain only storms that could have an impact on France. This is done by selecting the storms for which the track has at least one tracked point within 1300 km from the territory of metropolitan France. The radius of 1300 km ensures that the largest depressions, which can have an impact, distant from the centre of maximal vorticity, are not discarded. The choice of the footprint radius was motivated by previous studies , who used a radius of 10° around the centre of minimal pressure. Similarly, defined their footprints with a 1000 km radius. Although this radius may appear large for weaker depressions, we conducted several sensitivity tests using smaller radii and a variable radius conditioned on the minimum pressure observed throughout the ETC life cycle. Similar results were found with these additional tests (cf. Appendix. ). Additionally, only depressions for which the tracks lasted more than 24 h are kept. This leads to a set of 4439 storms over the period ranging from March 1979 (included) until March 2024 (excluded).

For each identified track, we define the storm occurrence date (dstorm) as the time when the storm centre, defined by the location of maximum ζ850, is at its minimum longitudinal distance from the reference meridian at 7.5° W. This corresponds to the time step at which the Euclidean distance between the storm centre and the 7.5° W longitude line is minimal. This reference point along the track is used to assign a unique identifier to each storm. Storms are named using dstorm together with the latitude and longitude of the storm centre at this time. For example, Storm Lothar reached its minimum distance to the 7.5° W meridian on 26 December 1999 at 12:00 UTC, when its centre was located at 4.2° W and 51.5° N. It is therefore labelled as “1999-12-26 12h [-4.2; 51.5]”. This naming convention uniquely identifies storms, even when multiple storms crossed the 7.5° W longitude on the same date. An example of two storms, their occurrence area and their naming with respect to the 7.5° W longitude line, is illustrated in Fig. . Storm footprints are defined using hourly 10 m wind gust with a circular spatial mask of 1300 km radius around the centre of maximal vorticity and a temporal window of ±12 h around each tracked point. Alternatively other footprint areas can be used, such as for the definition of storm clusters. We will note the rkm-footprint areas, the footprint area of a given storm with a radius of r km moving with the centre of maximal vorticity.

The TRACK scheme has been shown to effectively represent ETCs and is well-suited for impact analyses e.g.. However, we applied less restrictive thresholds on cyclone duration and intensity compared to other studies e.g. which required a minimal cyclone duration of 48 h, a minimum track length of 1000 km and a minimal ζ850max threshold of 10-5 s-1. This was done to ensure the inclusion of more ETCs, particularly fast-moving and weaker storms, which can be responsible for significant damage locally. also showed the importance of secondary cyclogenesis in serial clustering, which are often small-scale storms. Capturing storms of all scales is thus key to understanding damage during clustered events.

The analysis is performed using a single storm tracking algorithm, which can present some limitations, as different tracking schemes can produce differences in cyclone detection, track structure, and life-cycle representation . Restricting the set of storm tracks to only intense systems would introduce a selection bias in the association procedure, as claims could then only be linked to intense storms while impacts from weaker or shorter-lived ETCs would be excluded. Including depressions of all intensities allows for a more exhaustive representation of the ETCs that may contribute to observed damage.

Definitions of storm clusters vary across the literature and depend on both temporal and spatial criteria. Previous studies have employed a range of approaches, including absolute frequency metrics , relative frequency metrics , and different temporal windows . In this study, we adopt an absolute frequency-based approach, complemented by a spatial criterion. Specifically, two or more storms are considered part of the same cluster if their 700 km-footprint areas intersect, and their impact dates are separated by no more than 96 h. The choice of a 700 km radius is based on the 6° angular distance around the centre of maximum vorticity, which has previously been used to define the region of strongest wind impacts . The radius is intentionally smaller than the one used to define the storm footprint in Sect. , as the goal here is to capture regions exposed to the most intense wind gusts from multiple ETCs. Figure  illustrates an example of a cluster of storms with the tracks of the two ETCs, their own footprints (defined in Sect. ) and the intersection of their 700 km footprint areas. The 96 h temporal window is consistent with storm definitions used in Generali France's reinsurance contracts. To focus on events relevant to the French territory, we further require that the footprint area intersection occur over France. In cases where one identified cluster is entirely contained within another, the smaller (subset) cluster is discarded. We conducted a sensitivity analysis (Appendix ) on both the temporal and spatial criteria to assess how the clustering definition influences the number of detected events, as well as the share of claims and losses. Although we adopted a 96 h window specific to Generali France, rather than the 72 h window commonly used in reinsurance practice, this choice mainly affects the number of storms grouped within clusters. Importantly, the main conclusions regarding the proportion of losses and claims remain robust to those parameters.

Under this definition, a single storm can be part of several clusters. Among the 4439 storms affecting France over the 1979–2024 period, 1283 are part of at least one cluster. While one storm is found to participate in as many as 4 distinct clusters, the average is 1.18 clusters per storm. In total, 517 storm clusters are identified, with an average of 2.7 storms per cluster and a maximum of 13 storms in a single cluster.

Generali France's (P&C) claims portfolio, spanning from 1998 to 2024, is used as the impact dataset. All impact analyses involving these insurance losses are restricted to the 1998–2024 period. From the most recent analysis of building insurance in France, the market share of Generali can be estimated at 3 % in France , which corresponds to ≈1×106 contracts. In France, residential property insurance is mandatory, and coverage reaches nearly 99 %. Consequently, Generali France's P&C claims portfolio provides a meaningful representation of storm-related residential property insurance impact. The analysis focuses on windstorm claims recorded during the extended winter season (September–April). Each claim is characterised by the geographical location of the damaged property, the date of declaration and the date of damage. In this study, we refer to the estimated date of damage as the claim date (dclaim), which can differ from the actual date of damage. Damage intensity is quantified by the net insured loss, over which deductibles and coverage limits are truncated. As a result, some values may be zero or negative; these entries are excluded from the analysis. In addition, claims exceeding EUR 150 000 in net insured loss at the date of damage, which correspond to “severe damage”, are treated through a separate process and are also excluded. To ensure consistency across years, all monetary losses are detrended and converted to constant EUR 2015, written “EURcst2015”, using inflation indices provided by the Institut National de la Statistique et des Etudes Economiques . Claims can be categorised as “closed”, “open” or “out of order”. Closed claims have been paid based on the reported loss. Open claims are still under assessment, with losses subject to revision. Claims labelled as “out of order”, meaning not attributable to the relevant hazard, are excluded from the dataset. The full dataset was extracted on 10 December 2025 and contains approximately 2 % of opened claims, primarily corresponding to events in March and April 2024, at the end of the 2023–2024 winter season. After applying the filtering mentioned above, the cleaned dataset contains about 200 000 entries.

The geographical locations of insured properties are derived from textual addresses and converted into latitude and longitude using the , the national authoritative reference for French addresses. The geocoding tool from the Environmental Systems Research Institute is used to match addresses to this reference. Among the reported claims, 75 % have a point level location match, meaning that the exact building is identified. For 18 % of cases, only the street or hamlet name is identified. Lastly, ≈7% have locations only determined at the postcode level.

To characterise the financial impact of individual storm events, we define a set of metrics commonly used in the insurance sector. Let Ni,j,s denote the number of claims and Ci,j,s the total cost at location (i,j) attributed to a storm s. From these quantities, we compute the total cost, Cs=∑i,jCi,j,s; the total number of claims, Nsclaim=∑i,jNi,j,s; and the mean cost per claim, Csclaim=∑i,jCi,j,s∑i,jNi,j,s for a given storm s. Together, these metrics describe not only the overall financial impact of each storm but also the average severity of individual claims. They are widely used in insurance practice to interpret recorded losses .

Claim dates filled in the damage dataset are inherently uncertain as they are reported by policyholders based on their perception and interpretation of the hazard. This reporting bias, often referred to as historical perception, leads the raw claim dataset to over-represent intense windstorms and under-represent weaker events . As a result, claim dates (dclaim) may differ from the actual dates on which damage occurred. To mitigate this issue, a common approach is to aggregate the damage over several days e.g.. However, such aggregation can mask the contribution of smaller-scale storms when multiple events occur in close succession, and it implicitly assumes that a single storm is responsible for all damage within the aggregation window. This limitation is illustrated by Fig. a, which shows the daily number of declared claims. Claims appear nearly every day, a pattern that is inconsistent with the actual frequency of damaging storms and instead reflects reporting dispersion rather than true event occurrence. To address this data quality issue while still capturing the impact of all ETC events, we propose an association method that maps the claims and the storms. This approach reassigns claims to specific storm events, thereby correcting claim dates and improving their consistency with the timing of the underlying storm hazard.

Step 1 links each claim to all storms for which the storm date (dstorm) is close to the claim date dclaim. This closeness is determined by a temporal window defined by several days before (Xb) and after (Xa)dstorm. Concretely, storms are selected so that dstorm verifies dstorm-Xb≤dclaim≤dstorm+Xa (in days).

Since the ultimate goal is to associate each claim with a single storm, the storm most likely responsible for the damage is selected following a method based on wind gust value. In Step 2, each claim is assigned to the storm producing the highest wind gust at the claim location. Figure b illustrates the resulting storm-claim associations after Step2, showing 8 storms and the number of claims attributed to each of them. For the events on the 23 and 24 January 2009, dstorm of the identified storms corresponds well to the days on which the maximal number of claims are reported. However, during periods with fewer reported claims per day, between the 17 and 21 January 2009, up to four candidate storms are identified, each associated with at most 50 claims, and in some cases fewer than 10. Given that storm events are associated with substantial impacts, and that Generali France's exposure is broadly representative of the French market, it is unlikely that a storm occurrence would result in so few claims. To account for this, we introduce a minimum-claim threshold, denoted nclaims, which is later optimised to better match the observed impact. In Step 3, any storm associated with fewer claims than this threshold has its claims reassigned to the temporally closest storm. This procedure is applied iteratively, increasing the threshold by increments of 10 claims at each time, until stable results are obtained. Figure c shows the storm events identified after the last step of this association strategy.

The proposed association depends on three parameters Xb,Xa and nclaims, respectively corresponding to the number of days before dstorm, the number of days after dstorm and the minimal number of claims associated with a storm. These parameters help fine-tune the association between claims and storms. The performance of the association is evaluated with three metrics that compare the characteristics of the identified storms (date dstorm, number of storm events, number of associated claims) with the raw temporal distribution of claims. Claim maxima are identified by maxima in the time series of claim counts (Fig. a), computed by aggregating claims over the entire French territory. Tuning the association parameters relies on the assumption that the identified storms adequately capture both the timing and magnitude of the highest daily claim intensities, despite potential temporal shifts between storm occurrence and reported damages. The local maxima identified using the distribution of the number of claims as a function of time represent the target of the association. The constructed metrics evaluate whether the number of storms identified by the association methods corresponds to the number of claims maxima and whether the dates of the claim maxima are not too far from the dstorm dates. The precision metric (Mdays), in number of days, is computed as the maximal difference between dstorm and its closest local maximum. The smaller this difference, the better the association. The frequency metric (Δf), in number of events, is defined as the difference between the number of storms associated with claims and the number of local maxima. Lastly, the completeness metric (Pclaims), in percentage, corresponds to the percentage of claims associated with a storm event. These metrics collectively provide an overall assessment of how effectively the association strategy links claims to storm events. No spatial performance is evaluated at this stage. It is assumed that the association with the storm producing the highest wind gust should already capture the spatial claim distribution.

In mathematical terms, let {S1,…,Sn} be the dates of the n storms identified and {L1,…,Lm} the dates of the m claims local maxima. Then the frequency difference is defined as Δf=n-m. The precision metric can be written as Mdays=max⁡j∈[1,n]min⁡i∈[1,m]|Sj-Li|. The difference between the dates |Sj-Li| is expressed in days, as the local maxima computed over the claim dates are at the day level. The sensitivity and the optimisation of these metrics are detailed in the Supplement. The retained optimal values are Xb=3,Xa=3 d and Nmin_claims=50. These values will be kept for the remainder of the study. The resulting aggregation window of 7 d exceeds the 5 d used by and the 3 d considered by . We expect that such large windows better account for the potential postponed impact of a windstorm, but also for the error made by the insured person when declaring the damage date. Additionally, the linking strategy based on the maximal wind-gust value should ensure the association with the correct storm driver.

Assessing the representativeness of the resulting catalogues remains a methodological challenge due to the variety of available approaches. In a comparison between academic and insurance-based catalogues, highlights significant discrepancies in storm frequencies, major event identification, and estimated losses. These differences stem from the primary objectives of each catalogue but may have important implications for risk assessment. The proposed approach, specifically tailored to Generali France's exposure, produces a more robust event catalogue in which individual claims are systematically linked to ETCs along with their associated characteristics.

A key consideration is the catalogue's ability to accurately capture the most severe historical events. When the association method described above is applied, 322 storms are associated with an impact for Generali France over the 1998–2024 period. We call these events “impactful storms”. Among these, the 40 costliest events are presented in Fig. . Storms Kurt, Lothar, Martin, Klaus, Xynthia, Eunice, and Ciaran were the costliest for Generali, with losses exceeding EUR 10 million per event. Figure  illustrates that our method successfully captures high-impact storms documented in global datasets . Further comparison with Météo France's catalogue, produced using the Storm Severity Index (SSI), reveals a strong agreement over the period 1998–2024 . Only the storms of 16–17 December 2019 (storm Andrea and Calvann) rank among the most intense in Météo France's analysis, but do not appear in the top 40 costliest events from Generali France. This discrepancy would require further investigation, with one plausible explanation being the influence of Generali France's portfolio.

The resulting catalogue exhibits pronounced interannual variability, both in the number of storms associated with damage and in the corresponding loss amounts. On average, 6.7 damaging storms are identified per year (standard deviation 3.9), while annual losses have a median value of EUR 4 million and a much larger standard deviation of EUR 35 million, reflecting the strong year-to-year variability of impacts. Figure  underlines that the costliest winter seasons do not always correspond to the year with the most numerous storms or clusters of storms. This is the case for the winter 1999/2000 with only 7 storms associated with damage, but corresponding to the costliest winter. The winter 2013/14 is also notable with 14 storms identified with damage, but a total loss for the season close to the average. This high number of storms associated with moderate losses can come from the diversity of our storm tracks. As small and fast-moving depressions were kept in the set of storm tracks, it is more likely to encounter small storms associated with little losses. The variation of losses can be explained by both the intensity of the storms and the vulnerability of the exposed areas. This underscores the complex interplay between meteorological variability and socioeconomic vulnerability.

The pronounced interannual variability in the number of ETCs reflects year-to-year changes in storm intensity and track density . Beyond storm occurrence alone, and show that variability in storm-related losses is further amplified by the interaction between storm frequency and the spatial distribution of exposed assets. Consequently, years with comparable numbers of storms can produce very different loss outcomes depending on storm intensity and the regions affected by their tracks. The patterns of interannual variability in both storm counts and associated losses shown in Fig.  underscore the combined role of meteorological variability and exposure in shaping windstorm losses.

The quality of our catalogue of claims and storm events can be assessed using the insurance-based metrics defined in Sect. . For the set of impacting storms, the mean and median costs per storm are EUR 1 million and EUR 0.27 million, respectively. The marked difference between these values reflects a highly skewed distribution of losses across events: most storms generate relatively moderate losses, while a small number of damaging storms contribute disproportionately to the total loss. This concentration of losses in a few extreme events corresponds to a heavy upper tail in the distribution of costs per storm. A comparable skewness is observed in the distribution of the number of claims per storm. Although the mean number of claims reaches 591, the median is only 102, again reflecting the dominant influence of a limited number of very costly events. Such heavy-tailed behaviour is a well-documented feature of windstorm insurance data and arises from the combined effect of storm intensity, storm footprint, and the spatial distribution and vulnerability of exposed assets. As a consequence, summary statistics such as means are strongly driven by a few extreme storms, making the statistical characterisation of storm-related losses particularly challenging. Similar findings have been reported in previous studies for both claim counts and loss amounts , which consistently show that cumulative losses are typically dominated by a small subset of high-impact events.

Regarding the mean cost per claim, report values ranging from EUR1530 to 2335 in 2023 (in constant EUR) for residential properties, corresponding to EUR 1285 and 1962 in 2015 constant EUR. These estimates are lower than the mean cost per claim of approximately EUR 2700 obtained in this study from Generali France's insured losses. also showed, based on an in-depth analysis of the 2018/19 winter season, that the average cost per claim strongly depends on the type of infrastructure affected. The higher mean losses found in our dataset can therefore be partly explained by Generali France's specific exposure and portfolio composition, as well as by the longer historical record considered here. Generali France's portfolio is concentrated in densely urbanised areas such as Paris, Lyon, Lille, and Bordeaux, as well as in the south-east and northern regions of France, where higher asset values likely contribute to larger average claim costs.

To further assess the implications of the storm–claim association, we analyse the differences between reported claim dates and the actual dates of the associated storm events. Linking claims to storms allows us to compare the declared claim date (dclaim) with the corresponding storm date (dstorm). Figure  shows the probability distribution of absolute difference between these two dates. We find that 77 % of claims were reported within 24 h of the storm occurrence date, while 96 % were reported within a 7 d window.

Comparing this distribution with the 96 h temporal window used to define storm clusters highlights the limitations of relying solely on fixed aggregation periods. While most claims are reported close to the actual storm date, delays exceeding 24 h already introduce the possibility that losses from different successive storms become temporally mixed in the claims record. In periods of enhanced storm activity, where multiple events affect the same region within a few days, such reporting lags can blur the boundaries between individual storms. When delays accumulate over several temporally compounded events, the effective spread of reported claim dates may exceed the 96 h clustering window. Without an explicit association between each claim and its corresponding physical storm, losses may therefore be attributed to the wrong event or artificially distributed across multiple storms. This temporal misalignment complicates the attribution of damage to individual storm and makes it more difficult to disentangle true meteorological clustering from variability introduced purely by reporting practices. By contrast, the method developed in this study, which links each claim directly to identified storm tracks, helps restore a physically consistent timeline of impacts. This approach improves the separation between successive events and enables a clearer interpretation of loss patterns during clustered storm periods.

The winter of 1999 represented a significant challenge for the insurance sector in Europe, particularly in France, due to the successive storms Kurt, Lothar and Martin on 24, 26, and 27 December 1999. Storm Kurt had a northern trajectory, causing severe damage across Scandinavia . Lothar, meanwhile, brought widespread destruction to France, Switzerland , Germany , and Belgium. Storm Martin followed a path similar to that of Storm Lothar, leading to an exacerbated impact in the regions mentioned previously .

The strength and positioning of the upper-level zonal jet provided the primary large-scale forcing that triggered the explosive deepening of storm Lothar . With only 24 h separating Storm Lothar from Storm Martin, their impacts are difficult to distinguish in aggregated data. At the time, the reinsurance hour clause was loosened for recovery purposes and allowed for separating some losses . Nonetheless, this distinction was made solely for economic constraints and was not based on the actual damage caused by each storm. As a result, most studies treat Lothar and Martin as single events when evaluating their collective impacts .

Studies have underlined difficulties in disentangling the individual impacts of these storms. Figure  illustrates the spatial distribution of claims around the dates of both storm events. We underline that claims were declared for each of the days and over the whole of France during the subset period. For cities such as Paris, claims were identified each day. Without more information about the intensity and location of the storms, it is thus impossible to differentiate between the successive events.

Storms Kurt, Lothar and Martin respectively had an impact on 24, 26, and 27 December 1999. However, no claims were declared on the 24, and the impact of Lothar and Martin is likely mixed with the claims declared between the 26 and the 28. Restricting the analysis to claims reported strictly on the date of each storm would risk misattributing impacts and linking damages to incorrect meteorological conditions. Furthermore, attributing claims to the wrong event and date may affect the proper aggregation of losses under the 96 h reinsurance clause.

We applied our association method using the maximum wind gust speed of the three storms and the insurance claims for the period between 20 and 30 December 1999 (Xa=Xb=3 d). 1514 claims (out of 73 605) cannot be attributed to any of the three storms. Figure  shows the footprints of storms Kurt, Lothar and Martin along with their associated claims. The distinction between their estimated impacts is particularly evident over central France and corresponds well to the shift of the highest wind gust. Our association procedure leads to a better focus on the impact regions of successive events, allowing us to distinguish their effects even when storms share a common impact area. Hence, the claim patterns in Fig.  can be refined to those in Fig. .

We underline that the three storms have resulted in some damage in Paris. Although Storm Lothar had the highest windgust in the Paris area, the claim date played a key role in the distinction. This means that the claims observed in Paris and attributed to storm Kurt must have had a claim date earlier than 23 December 1999 (Xb=3 d before the impact date of storm Lothar). Similarly, the claims observed in Paris and attributed to storm Martin must have been declared later than Xa=3 d after storm Lothar, so later than 29 December 1999.

We highlight that damage is separated solely by the most intense wind gust. Our approach does not account for the persistence of strong wind gusts or the possibility that damage may result from the last storm, even if it did not produce the highest wind gust. This represents the primary limitation of the method. An alternative approach could involve using a different association method, such as linking events based on the closest track in terms of distance. However, for clustered storms affecting the same region, this may not be ideal. As illustrated in Fig. , weaker wind gusts can be observed near the centre of maximum vorticity. This suggests that some damage could be incorrectly attributed to areas with relatively low local wind gusts.

Although reinsurance practice typically aggregates windstorm losses within fixed temporal windows of 72 or 96 h, major events such as storms Lothar and Martin illustrate that a finer temporal separation may be required to distinguish the impacts of successive storms. The claim–storm association method enables this separation by attributing losses to individual storm tracks rather than to a predefined time window. At the same time, it successfully captures delayed damage, with claims that are reported several days after the physical event still being correctly assigned to the responsible storm.

The storm–claim association also underlines the relative contribution of smaller storms, which are usually discarded. As discussed in Sect. , it is crucial to account for all storms, as fast-moving systems or smaller depressions can generate strong surface winds. If these storms are not captured, all the damage may be attributed to the strongest storm, which might not be the one responsible for the strong winds at the given location. The example of Storm Klaus, represented in Fig.  serves to illustrate the importance of including small depressions and accurately distinguishing their impacts.

Storm Klaus affected Southern France and Northern Iberia between 23–24 January 2009. It was characterised by an explosive deepening of 37 hPa within 24 h, which is relatively uncommon for this latitude . This rapid intensification was driven by an extended and intense polar jet at upper levels. At the surface, strong wind gusts caused significant damage to infrastructure and forests .

Storm Klaus is often treated as a standalone event by many insurers . Claims filed within a 72 or 96 h window around 24 January 2009, are typically attributed solely to Klaus. However, our association results reveal that Storm Klaus was part of a broader storm cluster. The tracking algorithm identified a preceding depression, which we here name Storm A, crossing Northern France on 23 January 2009, which produced significant wind gusts exceeding 20 m s⁻¹ in central and northern France. Additionally, a secondary low-pressure system developed on 25 January 2009, that we name Storm B, was potentially responsible for claims in the South-West of France. As illustrated in Fig. , wind gusts generated by the 23 January and 25 January ETCs exceeded those generated by Klaus in several regions. This suggests that damages in northern and central France may not be attributable to Klaus alone. These findings demonstrate the added value of our storm-claim association method and highlight the importance of including smaller depressions in storm databases.

This example demonstrates the importance of the accurate storm–claim association, especially in cases of clustering. Attributing all impacts to Storm Klaus neglects the influence of smaller depressions. The relative contribution of these small systems within a 96 h cumulative sum is essential for understanding the role of serial storm clustering. This study highlights potential pitfalls in the interpretation of clustered events, as aggregated losses, claims, or impacts over a period as large as 72 or 96 h often encompass damage from multiple storm events. It remains unclear whether Klaus alone would have caused similar levels of damage had it not been preceded by another low-pressure system. This consideration is critical for understanding compound and cascading impacts.

The clustering method described in Sect.  can be applied to the subset of storms that produced insured losses. This allows us to identify high-impact clusters, defined as clusters for which at least two storms are associated with damage, such as Lothar and Martin in 1999. Among the 322 storms associated with impacts for Generali, 161 belong to 72 high-impact clusters. This represents nearly 50 % of damaging storms, compared with only 29 % when clustering is applied to the full set of tracked storms. It means that the impactful storms are more frequently part of clusters.

Although clustered storms are not the most numerous, they account for the majority of losses. Figure  already shows that 82 % of the 40 costliest storms for Generali occurred within clusters. Extending this analysis to the full dataset reveals that clustered storms are responsible for 85 % of total windstorm losses and claims. This share far exceeds their physical occurrence frequency. Therefore, storms in clusters appear not only more likely to cause impacts, but also more likely to be associated with the most damaging events.

Section  highlighted the strong interannual variability in both storm frequency and losses. A similar variability is found in storm clustering. The period from 2000 to 2005, for example, is characterised by few impactful storms for Generali (less than 10 storms) and correspondingly moderate losses. In contrast, the decade 2013–2023 is associated with important storm losses and a prominent storm clustering. Interestingly, the number of clustered storms in a season is not, by itself, the main driver of seasonal losses. Winter 2021/22 illustrates this: although only three storms are classified as clustered, they accounted for most of the seasonal losses. As discussed in Sect. , this indicates that the method also captures the contribution of smaller depressions embedded within active periods, whose cumulative impacts can be substantial.

For the costliest winters on record for Generali France, including 1999/2000, 2008/09, 2009/10, 2019/20, and 2023/24, nearly all losses are associated with clustered storms. These seasons were dominated by well-known major events, such as Storms Kurt, Lothar, and Martin in 1999/2000; Storm Klaus in 2008/09; Storm Xynthia in 2009/10; Storms Amélie, Fabien, Ciara, Ines, and Karine in 2019/20; and Storms Ciarán, and Domingos in 2023/24. This confirms that the most damaging seasons tend to be characterised not only by intense individual storms, but also by the close temporal succession of multiple impactful systems.

Consistent with Sect. , we use the same insurance-based metrics (defined in Sect. ) derived from the claim–storm association, but here we distinguish between individual storms and storms occurring within clusters (Fig. ). Although both samples contain a similar number of events, clustered storms show a much wider distribution of total cost per storm (Cs). They include the most damaging events, resulting in higher mean and median losses, but also a larger share of low-cost storms, reflected in a lower first quartile. This broad spread highlights the heterogeneous nature of clustered periods, which combine one or more high-impact storms with several weaker yet still damaging systems. A similar pattern emerges for the number of claims per storm (Fig. c). Clustered storms generate more claims on average, but with greater variability, spanning both very active and relatively minor events. As a consequence, clustered periods tend to be more costly overall, primarily because the most damaging storms, both in terms of total losses and claim counts, are more often embedded within clusters. The distribution of impacts among successive storms within clusters is examined in more detail in Sect. . In contrast, the mean cost per claim (Csclaim; Fig. b) is similar for clustered and individual storms (EUR 2786 vs. EUR 2687). This indicates that clustering influences total losses mainly through the number of claims and their allocation across successive storms, rather than through substantial differences in the average severity of individual claims.

Individual and clustered impactful storms also differ in intensity. Figure  presents the distributions of minimum MSLP, maximum ζ850, and storm duration. Storms associated with damage exhibit lower minimum MSLP and higher maximum vorticity than the full set of potentially impacting storms, with these contrasts being even more pronounced for storms embedded in clusters. Consistent with , the most intense storms affecting Generali France are therefore more likely to occur as part of storm sequences. In terms of duration, individual and clustered storms show broadly similar distributions. However, impactful storms, whether isolated or clustered, tend to persist longer than the complete ETC sample, suggesting that longer-lived systems have a greater potential to generate substantial damage. The spatial distribution of storm tracks, including all storms as well as individual and clustered impactful storms, is shown in Appendix .

These physical differences explain the patterns observed in the insurance metrics (Fig. ). Clustered storms concentrate the most intense systems, corresponding to the highest costs and largest numbers of claims, while also including less severe events whose losses are distributed across the cluster. The greater depth, higher vorticity, and longer duration of clustered storms drive their overall amplified financial impact.

Figures  and together emphasise the importance of an event-based analysis, showing that losses are driven not only by the clustering phenomenon but also by the characteristics of individual storms. Section  shows that clustered storms are the most intense and are associated with amplified financial impacts, but the total loss within a clustered event is not necessarily concentrated over a single storm. To better understand how damage is distributed within clusters, we examine the contribution of each storm member to the overall cluster loss. For each storm within a cluster, we compute its share of the loss as the total cost of that storm divided by the total cost of the cluster. This share is then compared to the storm's loss and occurrence rank, where a loss rank of 1 corresponds to the costliest storm and an occurrence rank of 1 corresponds to the earliest storm in the cluster.

From the 75 identified cluster events, Fig.  shows the share of the loss as a function of these ranks. Figure a highlights a wide spread of the loss within the several clusters, the costliest storm can account for between 41 %–99 % of the total cluster losses. On average, we found that the costliest storm of the cluster represents 70 % of the cluster's total cost, indicating that a single storm often dominates the overall loss. Nonetheless, the remaining share of losses still reflects a significant contribution from additional storms within the same cluster, underlining the need for insurance of grouping claims around clusters. This further demonstrates that the total observed damage often arises from the temporal clustering of multiple storms.

Figure b shows that the order of arrival is not a systematic driver of loss. On average, the first storm contributes 41 % of the cluster loss, but this value varies widely, from nearly zero to nearly the total impact of the cluster. Similar variability is observed for the second storm, while the third and fourth storms generally contribute less. Few clusters contain more than two storms, and for clusters with three or four members, an equally split loss corresponds to roughly 30 % and 25 % of the total cluster loss, respectively. This can explain the loss values observed for clusters with more storm members.

Finally, the near-equality of the medians and means in the box plots suggests that the distribution of cluster losses, by both loss and occurrence rank, is approximately Gaussian. Unlike the event-based analysis shown in Fig. , the cluster-based perspective indicates that the most extreme storms do not systematically account for the entirety of cluster losses. If they did, the distribution of losses at the cluster scale would be as strongly skewed as the distributions of cost and number of claims per storm. Instead, losses are more evenly distributed among the storms composing each cluster.

A similar analysis can be applied to the claim frequency (Appendix. ), represented by the number of claims per event. The results are comparable: on average, the most impactful storm accounts for 68 % of the total claims, with substantial variability in the share of claims across different claim ranks. Figure  also indicates that the order in which storms occur within a cluster does not appear to influence how claims are distributed among the storms.