Comparing an insurer ’ s perspective on building damages with 1 modelled damages from pan-European winter windstorm event 2 sets : a case study from Zurich , Switzerland 3

9 With access to claims, insurers have a long tradition of being knowledge leaders on damages caused by e.g. 10 windstorms. However, new opportunities have arisen to better assess the risks of winter windstorms in Europe 11 through the availability of historic footprints provided by the Windstorm Information Service (Copernicus 12 WISC). In this study, we compare how modelling of building damages complements claims-based risk 13 assessment. We describe and use two windstorm risk models: an insurer’s proprietary model and the open 14 source CLIMADA platform. Both use the historic WISC dataset and a purposefully-built, probabilistic hazard 15 event set of winter windstorms across Europe to model building damages in the canton of Zurich, Switzerland. 16 These approaches project a considerably lower estimate for the annual average damage (CHF 1.4 million), 17 compared to claims (CHF 2.3 million), which originates mainly from a different assessment of the return period 18 of the most damaging historic event Lothar/Martin. Additionally, the probabilistic modelling approach allows 19 assessing rare events, such as a 250-year return period windstorm causing CHF 75 million damages, including 20 an evaluation of the uncertainties. Our study emphasises the importance of complementing a claims-based 21 perspective with a probabilistic risk modelling approach to better understand windstorm risks. The presented 22 open source model provides a straightforward entry point for small insurance companies. 23


2.2.1
Historic windstorm hazard event set

147
The historic windstorm hazard event set − denoted "WISC historic" − contains wind gust footprints for around

173
The wind gust speeds from "WISC historic" are considered to be realistic compared to observations for areas at These newly created "probabilistic" footprints can be viewed as scenarios of plausible windstorms as they only differ slightly from historic events, retaining both the spatial extent and general structure. In countries close to For using the scenarios in a qualitative risk assessment framework, the probabilistic windstorm footprints can be used as they are, but for a quantitative risk assessment the frequencies of the windstorm footprints need to be estimated. In an effort to assign reasonable frequency estimates to the probabilistic windstorm footprints, we

229
"WISC probabilistic extension" includes footprints for 4'118 probabilistic windstorm events, along with the 230 142 original windstorm events in "WISC historic" (Table 1), and provides a basis of an event-based risk 231 assessment for winter windstorms with return periods of around 250 years, a scenario relevant for regulatory 232 requirements in the insurance sector. It is important to note that this method incorporates a lot of uncertainty,

233
including but not limited to the sampling uncertainty of rare events in a relatively short time range (i.e., 75 years 234 in case of "WISC historic").

235
Encouragingly, the hazard event set "WISC probabilistic extension" shows considerably higher wind gust 236 speeds in the canton of Zurich as compared with "WISC synthetic" (Fig. 1). Nonetheless, the maximum wind 237 gust speeds of the most extreme event in "WISC probabilistic extension" are not considerably higher than those 238 of Lothar/Martin, the most extreme event in both "WISC historic" and the insurance claims data.

259
In both damage models, the extent of damage results from the intensity of the windstorm event (i.e., hazard), the

275
The initial step is a simple spatial overlay of the gridded maximum wind gust speeds during the respective

293
In the next step of the damage model, the probability of buildings affected is calculated with a stochastic 294 approach. The respective windstorm event was automatically categorised according to its severity (here,

315
As the total value of the exposure is different between the GVZ exposure, the CLIMADA exposure, and the

332
There are several risk assessment metrics that can be calculated with a set of event damages, which are the main 333 result from the damage modelling described above.

374
However, for certain applications in the insurance industry the tail view of "WISC probabilistic extension" is an 375 important feature of the dataset. The sampling uncertainty of "WISC historic" is too large to provide, for 376 instance, a comparison criterion between two different exceedance frequency curves from different models.
Therefore, we propose to illustrate the probabilistic content of "WISC probabilistic extension" by using 378 bootstrapping of all probabilistic damage events. In this way, a "probabilistic envelope" around the best-guess

442
The red ribbon in Fig. 2 shows a possibility to illustrate the probabilistic envelope for the modelled damages

529
Insurance companies undertake their business under a strict regulatory environment, and having enough capital 530 to cover rare events is one of the regulatory requirements. The damage amount reached on average every 531 250 years is an often-mentioned indicator for such a rare event. However, the insured damages and also the 532 modelled damages based on "WISC historic" do not span a long enough period by far to make an empirical 533 prediction of a damage amount with a return period of 250 years. All methods of extrapolation from these 534 datasets suffer from the sampling uncertainty (shown as confidence intervals in Fig. 2). The hazard event set information. The fact that the probabilistic envelope for the modelled damages based on "WISC probabilistic extension" (red ribbon in Fig. 2) does not cover the full range of the sampling uncertainty for the modelled 540 damages based on "WISC historic" (yellow ribbon) shows two things: on the one hand, it shows the tail view,

570
Furthermore, it is not considered that buildings are partially adapted to local wind conditions (e.g., multi-storey 571 buildings or exposed buildings located on mountain tops).

572
Not every building is equally affected during a windstorm event. To take that into account, in the GVZ damage 573 model a random resampling of affected buildings was applied according to an assumed degree of impact (red windstorm events in the canton of Zurich. With every further windstorm, these assumptions will however 578 become more reliable in the future. In contrast, the deterministic PAA values (Schwierz et al. 2010), as used in 579 the CLIMADA impact model, are much smoother and thus allow a smooth damage modelling (Fig. 2).

580
However, these values are not specific for windstorms in the canton of Zurich and they do not allow a stochastic 581 sampling as in GVZ's damage modelling approach.

582
The rapid estimate of the damage potential in the event of extreme, unprecedented windstorm events shown in

820
The urban areas of the two main cities Zurich (left) and Winterthur (right) are marked in blue.