The impact of hurricanes on a property portfolio: an empirical study based on loss data in Portugal

Abstract

Purpose

The simulated scenarios can be used to compute risk premiums per risk class in the portfolio. These can then be used to adjust the policy premiums by accounting for storm risk.

Design/methodology/approach

A complete model to analyse and characterise future losses of the property portfolio of an insurance company due to hurricanes is proposed. The model is calibrated by using the loss data of the Fidelidade insurance company property portfolio resulting from Hurricane Leslie, which hit the centre of continental Portugal in October, 2018.

Findings

Several scenarios are simulated and risk maps are constructed. The risk map of the company depends on its portfolio, especially its exposure, and provides a Hurricane risk management tool for the insurance company.

Originality/value

A statistical model is considered, in which weather data is not required. The authors reconstruct the behaviour of storms through the registered claims and respective losses.

Full text

Translate

Turn on search term navigation

1. Introduction

In this work, we use a new approach to study the impact of hurricanes on the property portfolio of an insurance company based on insurance data only. Firstly, we propose a linear model for longitude and latitude, based on the company's loss data, to define the Hurricane trajectory on the mainland. In this approach, we fix the landfall point of the Hurricane as the initial point. Secondly, we use a regression tree model to construct geographical risk classes based on the distance from the affected councils' main cities to the Hurricane trajectory or the Hurricane landfall point. Finally, we add these variables in the (existing) regression models as predictors for claims and loss severity. The final model provides an analysis tool for risk management, namely in assessing losses under different scenarios, according to the company's exposure.

The economic consequences of extreme weather events, such as hurricanes, have become more and more relevant. Table 1 shows that the average annual costs of weather and climate events in the US have significantly increased in the last decades. In Europe, the average costs also increased, mainly from the period of the 1980 to 1990s, and then remained stable during the following decades.

Part of the economic losses from climate and extreme weather events are covered by insurance companies. According to Munich Re [1], in 2021, natural disasters were so destructive that insurance companies worldwide were obliged to pay out more than 120 billion dollars (predicted amount) in compensations, which is the highest value in the last two years. In Europe, the total cost with climate-related and extreme weather events from 1980 to 2020 is estimated to be 487 billion euros (2020 values), 109 billion euros (2020 values) of which were covered by the insurance sector [2].

In Portugal, insurance companies have paid out approximately 478 million euros (2020 values) in compensations from 1980 to 2020. Some of the weather and climate-related events that provoked the highest losses to insurance companies in Portugal in the last years are described in Table 2. It can easily be seen that wildfires in the centre of Portugal in October 2017 were the most costly event. The second most costly event was Hurricane Leslie, for which insurance companies paid out approximately 101 million euros. Tropical Hurricane Leslie was the most devastating event regarding the number of claims.

Indeed, although mainland Portugal is not on the usual path of tropical cyclones, it has been hit by some extratropical depressions, which were originally tropical cyclones. In the period from 1995 to 2021, Portugal was influenced by ten extratropical depressions, of which five in the last six years. Others have become a part of the general circulation between the Azores and mainland Portugal.

In 2018, 18 tropical cyclones were observed in the North Atlantic. One of them, Leslie, at the end of the cyclone season, was formed on September 23rd and entered Portugal on October 13th through the region of Figueira da Foz. When it hit Portugal, it became an extratropical depression, but still with a lot of wind and rain activity. Figure 1 shows the maximum instantaneous wind field (gust) for values greater than 90 km/h, based on observations made by the network of surface meteorological stations of the Instituto Português do Mar e da Atmosfera (IPMA) and of the Intermunicipal Communities (CIM) of Coimbra and Viseu. The distribution of the wind field shows the passage of Leslie over the territory of mainland Portugal. This subject is further explored in (Pasch and Roberts, 2019; Baatsen et al., 2015; Befort et al., 2019; Bentley et al., 2016; Chan, 2005; Dacre and Gray, 2009; Oliveira et al., 2020).

According to Munich Re, tropical storms, such as hurricanes, typhoons and cyclones, are among the most costly natural hazards [3]. For instance, Hurricane Katrina was the most costly natural disaster of all time for the insurance sector, with losses exceeding 60 billion dollars. Hurricane Lorenzo, which hit the Azores in October 2019, caused total economic losses of 330 million euros [4]. Additionally, an increase in the frequency of tropical hurricanes in Western Europe during early autumn (Aug–Oct) is expected in the future (Haarsma et al., 2013). Accordingly, in this study, we present a methodology to assess the risk for insurance companies regarding this type of events in Continental Portugal.

2. Literature review

Climate and weather change and climate-related extreme events, have been identified as important challenges to insurance companies (Mills, 2005; Dlugolecki, 2000). On the one hand, insurers need to develop new insurance products that help mitigate these risks, particularly in developing countries, where the effects of climate change are expected to be more severe (Mills, 2005; LE Franzke, 2017; Botzen, 2013). On the other hand, insurers must find innovative methodologies to measure their portfolios' physical risk, considering the natural uncertainty of projections and the difficulty in diversifying the risk due to geographical correlation (Charpentier, 2008). Natural hazards' economic and social impact has also been extensively studied over the last few years (see, for instance (Donat et al., 2011; Jagger et al., 2008; Kim et al., 2016; Pinelli et al., 2004; Prahl et al., 2016; Prahl et al., 2015), and the references therein).

We contribute to the literature by proposing a model to analyse future losses in the property portfolio of an insurance company caused by hurricanes in Continental Portugal. Weather data is not required in this model, but rather we reconstruct storm behaviour through the claims and losses registered. We calibrate the model using the loss data of the insurance company Fidelidade for Hurricane Leslie. This analysis enables us to conclude that single-family houses are especially vulnerable to hurricanes and that the losses would be much more significant if a storm similar to Leslie were to strike Continental Portugal in other regions where the company has a higher exposure.

The closest paper to ours is (Kim et al., 2016), although the authors include climate data in their setup. They present a statistical model that predicts losses based on the variables of wind speed and buildings' age, floor area and real estate evaluation. To validate the model, the authors use the data set of Texas Windstorm Insurance Association records for claims payouts for commercial buildings after Hurricane Ike. Similarly, in the “Florida Hurricane Public Loss Model” project from Florida International University, a model based on meteorological data was developed to assess Hurricane risk for the insurance industry [5]. Finally, several authors have employed damage functions to assess the expected costs of these events (Donat et al., 2011; Prahl et al., 2016; Prahl et al., 2015). In (Jagger et al., 2008), the authors present a probabilistic model, predicting aggregated US losses caused by tropical hurricanes.

3. Data

It is natural to use climate data when modelling the damage caused by extreme meteorologic events (Cf. (Donat et al., 2011; Prahl et al., 2016; Prahl et al., 2015)). However, the climate data of extreme weather events is extremely variable in terms of space and time, especially when the wind is the main variable of concern. This makes the use of climate data very challenging, as it is common that the available values are averaged or extrapolated, resulting in the extreme values not being well represented in some datasets. Data regarding Hurricane Leslie is an example of the lack of reliability and accessibility of spatial climate data. According to the Portuguese Institute for Sea and Atmosphere (IPMA), a 176 km/h wind gust was recorded in Figueira da Foz [6]. However, looking at the dataset ERA-5 reanalysis from the European Centre for Medium-Range Weather Forecasts (ECMWF), the maximum wind gust value recorded for that site and day is 88.00 km/h.

Although the records of IPMA appear to represent extreme events better, such meteorological observations are not easily available and can only be obtained on request. For these reasons, we only consider data from the insurance company in this paper. Namely, we consider information regarding the portfolio of the company, including losses due to the meteorological event Leslie. To perform this analysis, we consider the property insurance portfolio of Fidelidade, which is the insurance company that has the biggest market share in Portugal. The dataset consists of a portfolio of 1,303,984 policies in 2018 and comprises all 278 councils of Continental Portugal. More than 99% of the policies cover losses due to storms, meaning that the damage provoked by floods and wind is covered. About 900,917 insured properties are located in councils with at least one claim due to Leslie, of those, approximately 1% reported claims. Table A1, in Appendix 1, displays those variables where information is available for each policy, as well as the possible values they can take. We define the risk class as the set of all those policies that have the same characteristics of the variables provided by the insurance company in Table A1.

4. Model for the storm path

In this section, we present an approach to model, the storm path, based on the insurance company's data. This approach differs from most works in the literature where climate data is also considered and is essential to build the damage functions that are later used to estimate losses, as in (Donat et al., 2011; Prahl et al., 2015, 2016). The construction of models that only use loss data has obvious disadvantages. However, they have the great advantage of relying on data from the company that, unlike climate data, is easier to measure and obtain.

We first infer the trajectory of the storm after its landfall from the loss data. By trajectory of the storm, we mean the imaginary line around which the observed claims are distributed, which is the line that passes through the affected councils, i.e. councils with claims. The trajectory is obtained by using the least-squares method. Once the trajectory is obtained, we define geographical areas of risk by introducing two new variables, which we call distance 1, dist1 and distance 2, dist2. This information is then included in the regression models for the number of claims and their costs.

In this work, we consider loss data aggregated by each council, although other granularities can be used. We start by defining the cost ratio and the ratio of affected buildings for each council i, as introduced in (Heneka et al., 2006), which we denote respectively by CR_i and RAB_i: ${C R}_{i} = \frac{l o s s i n i}{t o t . a m o u n t i n s u r e d i n i} {R A B}_{i} = \frac{n u m b . o f c l a i m s i n i}{t o t . n u m b . o f p r o p e r t i e s i n s u r e d i n i}$

These quantities provide a measure of the impact of the storm, which is relative to the exposure of the company in each council, allowing for comparisons in different regions and epochs according to the variability of the portfolio of the insurer.

Figure 2, left, displays the cost ratio caused by Hurricane Leslie, by council, from which we can infer the path of the Hurricane from its landfall in Figueira da Foz and its subsequent weakening along its path towards the northeast, which is reflected by the lower values of the cost ratio as it moved inland.

To estimate the trajectory followed by the Hurricane, we consider a least square problem to find the line that best adjusts to the geographical points of the main city of each affected council, subject to the restriction that the line must pass through the landfall point, which, in this case, was Figueira da Foz. We consider latitude and longitude measures. Hence, the least square problem with constraint to be solved is given by (1):(1) $\min_{α, β > 0} \sum_{i = 1}^{n} {({l a t}_{i} - (α + β {l o n}_{i}))}^{2}, s . t . α = {l a t}_{k} - β {l o n}_{k}$ where lat and lon denote latitude and longitude, and k = ” Figueira da Foz”. Figure 2, right, represents the geographical representation of the affected councils' main cities and the estimated line, i.e. the solution to the constraint least square problem Equation. (1).

The part of the trajectory line going from Figueira da Foz until the furthest council in the northeast of the country that reported claims, which is Bragança, has a length of approximately 260 km, meaning that the Hurricane caused losses to the company for at least 260 km travelling inland. There is no information available regarding the damages caused beyond the borders of Continental Portugal, but for reasons of prudence, we assume that the event could produce damages over 300 km before dissipating.

Based on the estimated trajectory, we define two new variables: dist1, representing the distance of the insured object to the storm's landfall point and dist2, representing the perpendicular distance of the insured object to the trajectory line. The use of variable dist1 is justified by the decrease of the observed cost ratio along the trajectory line, starting from the landfall point, while the use of variable dist2 is justified by the fact that the affected councils are closely distributed around the trajectory line (see Figure 2). Variables dist1 and dist2 are based on the insurer loss data and express an indirect measure of the damage. To understand how these variables relate to the observed cost ratio, CR and the ratio of affected buildings, RAB, we develop a regression tree model, representing CR and RAB for different values of dist1 and dist2 (see, for instance (Loh, 2011; Friedman, 2017), for tree-based and regression tree models). We use a regression tree based on the ANOVA method (Loh, 2011). The smallest number of observations allowed in a terminal node is 20, which is used as the stopping criterion for the size of the tree.

Figure 3 shows the two regressions obtained for CR and RAB of those councils affected by Hurricane Leslie, considering variables dist1 and dist2. In this case, the regression tree algorithm has divided the space into five regions for both CR and RAB, and the values of the response variable in each region are also reported in Figure 3, in the blue squares of the terminal nodes.

The regression trees obtained provide a mathematical description of the effect of variables dist1 and dist2, related to the storm's trajectory, on CR and RAB. We use this regression tree to build two categorical variables, denoted intensity1 and intensity2, one for CR and the other for RAB, respectively. These two categorical variables have levels defined by the splits in the regression tree, as described in Table 3. Accordingly, the variables intensity1 and intensity2 capture the combined effect of dist1 and dist2 on the different values of RAB and CR, respectively. These are the variables that are used in the regression models for predicting claim frequency and claim amounts in light of the geographical exposure of the Hurricane.

5. The model for predicting the losses

In this section, we present a model to estimate the expected losses from the storm in a given region based only on actuarial data. In particular, we consider property damage loss data from Hurricane Leslie to calibrate the regression model. In our approach, the loss data is seen as an indirect measure of the climate variables. The final model will be used afterwards to simulate losses caused by a similar weather event in different geographical regions and, ultimately, to build a storm risk map.

We first devise a model based on the trajectory of the storm to classify which councils were affected by the storm event, i.e. those councils with at least one claim. Afterwards, we estimate the claim frequency and amount for each affected council.

For the sake of computational simplicity, we assume that the coordinates of the policies coincide with the coordinates of the respective council's main city. The distances dist1 and dist2 are thus used to predict a binary outcome of 1 or 0, if the council is affected or not, respectively. This is carried out by employing a random forest model for classification based on Breiman's random forest algorithm (see (Breiman, 2001)).

Since the size of this data set is small, the predictive ability of the model is evaluated by means of a 10-fold cross-validation (see, for instance (James et al., 2013), for an explanation of the k-fold cross-validation technique). The performance of the estimate is assessed through the mean and variance of the sensitivity (or positive rate) and specificity (or negative rate) of the cross-validation tests: $sens . = \frac{num . true positives}{num . true pos . + num . false neg .} s p e c . = \frac{num . true negatives}{num . true neg . + num . false pos .}$

We now aim to model the average number of claims for those councils that are affected by the storm. To this purpose, we only consider the affected councils, as using the whole portfolio would lead to biased estimates. In order to obtain a functional relation between the predicted probabilities of claims and the characteristics of the policies, a logistic regression (cf. (Gelman and Hill, 2006)) is used to model the probability of the binary event as to whether a given policy registers a claim or not. To account for the storm, we consider various characteristics of the policy as explanatory variables, together with the variable intensity1. The linear regression is represented in Equation (2). For a complete description of all the variables employed, refer to Table A1, where:(2) $\log (\frac{p}{1 - p}) = β_{0} + β_{1} Type of Property + β_{2} Year of Construction + β_{3} F r a m i n g o f t h e H o u s i n g + β_{4} T y p e o f t h e H o u s i n g + β_{5} A l t i t u d e + β_{6} T y p e o f F l o o r + β_{7} F o r e s t A r e a + β_{8} B u s h A r e a + ε$

All the explanatory variables in our logistic regression (2) are categorical, and thus, each one is represented by a dummy variable (artificial variables that adopt the values of 0 or 1), with each one representing a different level of the explanatory variable (see Table A1). The estimation results are presented in Table A2, in Appendix 2.

In the logistic regression (2), p represents the probability that the policy has a claim, where p $\in$ (0,1). We require a model to decide the cut-off value of p above, for which we consider that there is a claim. A randomised procedure is adopted, which is based on the sampling outcomes 0 or 1 for each record from a Bernoulli distribution using the probabilities estimated through the logistic regression.

We use a ten-fold cross-validation to validate the adjustment of the logistic regression. The quality of the model is then assessed based on its ability to predict the average number of claims on a given risk class in the test data set rather than for a single policy.

Next, we attempt to model the average cost of a claim when it occurs. From our data set for Hurricane Leslie, the number of policies that reported a claim and thus represented a cost for the company was approximately 8,500, which represents approximately 1% of the total number of policies of the affected councils. It should be remembered that, for modelling purposes, only those councils classified as having at least one claim due to the storm are considered. If we consider the cost distribution relative to the whole portfolio for the affected councils, the distribution is highly right-skewed, with most of its mass concentrated in 0. Should we instead consider the cost distribution relative to those policies which reported at least one claim, then the distribution is also highly right-skewed. However, there is no probability mass concentrated in 0. In the latter case, we were able to log-transform the cost distribution and observed that the log-cost distribution was well approximated by a normal distribution. This enables us to use a multiple linear regression (MLR) model (Gelman and Hill, 2006) to predict the average log-cost. We consider the MLR model of Equation (3), where, once again, the explanatory variables are the characteristics of the insured property, together with a variable that accounts for the storm, which in this case is intensity2.(3) $\log (Cost | Cost > 0) = β_{0} + β_{1} Type of Property + β_{2} Capital Insured + β_{3} T y p e o f H o u s i n g + β_{4} U r b a n A r e a + β_{5} I n t e n s i t y 2 + ε$

The estimated values of model (3) are presented in Table A3 in Appendix 2, where the prediction ability of the model is assessed with a ten-fold cross-validation.

6. Case scenarios in Portugal

Given the impact of these phenomena on the insurer's portfolio, it is crucial to quantify the losses that could be caused by such events in those regions that have a higher exposure for the insurance company. Accordingly, in this section, we simulate the impact that a storm such as Leslie should make if it lands in a different part of Continental Portugal. For the purpose of analysing different scenarios, we need to simulate the trajectory of the storm and then choose the landfall point. Next, we follow the methodologies described in Section 5 for each simulated trajectory. To simulate the trajectory, we chose two elements that can be fixed or simulated, namely: the storm's length and its entrance angle. As we assume that Hurricane Leslie caused damages inland along 300 km, the length of the trajectory after its landfall is fixed at 300 km.

Once again, for the sake of computational simplicity, we assume that both the exact point where the Hurricane made land and the coordinates of the policies coincide with the coordinates of the respective council's main city. This implies that each policy belonging to a certain council has the same values of dist1 and dist2. Once the trajectory is drawn, the variables dist1 and dist2 are obtained for each policy, as well as, subsequently, intensity1 and intensity2.

In Figure 4, we present three maps of Continental Portugal which highlight the cost ratio in the different councils, assuming that a Hurricane such as Leslie reaches either Cascais, Porto or Faro in Portugal, with an entrance angle of 45°, 330° and 60°, respectively. The trajectory of the Hurricane is also represented in the maps.

In Table 4, for each scenario, we present the number of claims, the cost, and the mean cost per claim (MCC) in relation to Hurricane Leslie. The MCC is computed as: $M C C = \frac{T o t a l c o s t s}{T o t a l n u m b e r o f c l a i m s}$

According to our simulations, should a Hurricane such as Leslie reach Portugal in Cascais or in Porto, then the insurance company can expect a total number of claims that are approximately 3.22 and 2.61 times higher than the number of claims that were made in Figueira da Foz due to Leslie. On the other hand, should the landfall point be Faro and then the insurance company would expect approximately half of the claims observed in Figueira da Foz. These results are to be expected, as Cascais and Porto belong to the Lisbon Metropolitan Area and Porto Metropolitan Area, respectively, which are the two regions with the largest exposure for the company. Further detailed discussion on these three scenarios, namely regarding the claim costs, can be found in Appendix 3.

We now estimate the expected cost under many different scenarios in Continental Portugal by repeating the same simulations as those above over a large number of times. These scenarios are constructed by assigning different probability distributions to the (1) landfall council, (2) the entrance angle and (3) the trajectory length. The following two scenarios are considered:

Scenario A: In this scenario, we assume that the councils on the coast of Portugal that are south of Setubal are less likely to be hit by a Hurricane, with a total probability of 2/6 of that of Setubal, with the northern councils having a total probability of 4/6. We assign equal probability to the councils in each of the two regions, and the trajectory is assumed to have a fixed length of 300 km. For the West Coast councils, the entrance angle is simulated according to a triangle distribution with support between −90° and 90°, with a mode of 45°. For the South Coast councils, the entrance angle is simulated according to a triangle distribution of between 0 and 90°, with a mode of 45°.
Scenario B: We now assume that the coastal councils can be hit by a Hurricane, according to a uniform distribution, which has the same probability. The length of the trajectory is assumed to follow a continuous uniform distribution between 200 and 400 km. The entrance angle also follows a continuous uniform distribution in the interval of −90°–90° in the West councils and in the interval of 0°–90° in the South councils.

We present the distribution of the CR in Figure 5 and Table 5, and also the relative total cost, respectively, in continental Portugal for scenarios A and B. Despite the similarities between the two risk maps, we can easily see from Table 5 that the distribution of the total cost could be significantly different in different scenarios. Further analysis of the risk maps is presented in Appendix 3.

7. Conclusions

In this work, we simulate the impact of a storm such as Leslie for the entire portfolio of the insurance company Fidelidade, using data on the policies of the company's property portfolio. The claims are considered to be indirect observations of the intensity of the Hurricane. Thus, we model the storm path through these insurance data. The storm path is then included in the claim frequency and severity regression models.

Several scenarios are simulated and used as the basis for building risk maps. The results of our analysis show that there is a high probability that a future event with the same intensity as Hurricane Leslie will cause larger losses than Leslie. The simulated scenarios also enable us to compute the risk premium per risk class in the portfolio, which could be included in the policy premium calculation to account for this type of climatic event. It was possible to observe from the simulations that less-populated regions can lead to higher losses than more urbanised areas. However, this depends on the exposure of the company with regards to the type of property insured. This model thus provides a Hurricane risk management tool for the insurance company.

Due to the nonavailability of further significant data, the scenarios simulated in this analysis were just based on Hurricane Leslie and were extrapolated for various regions in Portugal. Nevertheless, the extrapolation carried out and its generalisation to other storms should be performed with caution. In addition, in our work, we have only considered property portfolio data, and it is most likely that the Hurricane could have affected other lines of business. Either way, it is extremely important that the insurer measures the risk associated with these events, even though the simulated scenarios are based on the scarce data available.

The authors thank to two anonymous referees and to the editor for the valuable comments and suggestions that helped improve the paper. Alexandra Moura and Carlos Oliveira were partially funded by FCT – Fundação para a Ciência e Tecnologia (Portugal), with national funding through research grants CEMAPRE/REM UIDB/05069/2020 and EXPL/EGE-ECO/0886/2021. The authors thank IPMA for kindly providing Figure 1.

Notes

1.Information available at https://www.munichre.com/en/company/media-relations/media-information-and-corporate-news/media-information/2022/natural-disaster-losses-2021.html (site visited on February 07 2023).

2.Information from the European Environment Agency available at https://www.eea.europa.eu/data-and-maps/daviz/economic-damage-caused-by-weather#tab-chart_2 (site visited on February 07 2023).

3.Information available at the website of Munich Re: https://www.munichre.com/en/risks/natural-disasters-losses-are-trending-upwards/hurricanes-typhoons-cyclones.html (site visited on February 07 2023).

4.Information available at: https://www.publico.pt/2019/10/14/sociedade/noticia/furacao-lorenzo-provocou-prejuizos-330-milhoes-euros-1889978 (site visited on February 07 2023).

5.This information is available at (visited on September 04 2023) https://www.ihrc.fiu.edu/research/projects/florida-public-hurricane-loss-model/

6.This information is available at (visited on February 07 2023) https://www.ipma.pt/pt/media/noticias/news.detail.jsp?f=%2fpt%2fmedia%2fnoticias%2farquivo%2f2018%2fleslie-3.html&msclkid=28dcdfafb9b311ec84dc5343cca45285

7.By weighted correlation, the authors mean that the correlation between the predicted and observed values weighted by the proportion of policies in each risk class.

Figure 1

Maximum instantaneous wind (gusts) with values greater than 90 km/h

[Figure omitted. See PDF]

Figure 2

Left: distribution of the observed cost ratios, by council, caused by Hurricane Leslie over Continental Portugal. The blue dot refers to Figueira da Foz, which is the landfall point of the Hurricane. Right: coordinates of the main cities of the councils that reported at least one claim due to Hurricane Leslie and the estimated trajectory of the Hurricane

[Figure omitted. See PDF]

Figure 3

Partition of the space defined by variables dist1 and dist2 performed by the regression tree method applied to the observed RAB (left) and the observed CR (right) of the councils with reported claims due to Hurricane Leslie. The leaves provide information on the predicted RAB (left) and CR (right), as well as the proportion of policies, in each node

[Figure omitted. See PDF]

Figure 4

Cost ratio map of Continental Portugal obtained for the following scenarios. Left panel – landfall point: Cascais, 45° entrance angle; middle panel: landfall point – Porto, 330° entrance angle; right panel – landfall point: Faro, 60° entrance angle

[Figure omitted. See PDF]

Figure 5

Distribution of the CR by council in continental Portugal due to a Hurricane such as Leslie in scenarios A (left) and B (right)

[Figure omitted. See PDF]

Table 1

Average cost per year of weather and climate-related disasters in the US and Europe. Costs for the US are in billion dollars (2022 values) and costs for Europe are in billion euros (2020 values)

Time period	US (billion $)	Europe (billion €)
1980–1989	20.5	7.3
1990–1999	31.4	11.8
2000–2009	58.7	13.4
2010–2019	93.6	14.9

Note(s): Information available at https://www.eea.europa.eu/data-and-maps/daviz/natural-disasters-events-5#tab-googlechartid_googlechartid_chart_111 and https://www.ncei.noaa.gov/access/monitoring/billions/(sites visited on February 07 2023) (Smith, 2023)

Source(s): Data from National Oceanic and Atmospheric Administration and the European Environment Agency. Table was authors’ own creation

Table 2

Weather and climate-related disasters in Portugal

Date	Type of event	Region of Portugal	Indemnity (million €)	Claims
Jan 2014	Storm	Mainland	11.5	5,544
Sept 2014	Floods	West region	1.5	489
Nov 2015	Floods	Algarve	15.5	1,762
Aug 2016	Fires	Madeira	19.7	328
Jun 2017	Fires	Pedrógão	22.4	493
Oct 2017	Fires	Centre	235.4	4,177
Oct 2018	Leslie	Centre and North	101.0	38,000
Dec 2019	Elsa and Fabian	Centre and North	42.0	22,700

Note(s): This information can be found on the following websites: https://www.apseguradores.pt/Portals/0/doc/publicacoes/Revista%20APS%2001_PT%20-%20FINAL.pdf?ver=2019-07-05-101014-453 https://www.apseguradores.pt/pt/comunica%C3%A7%C3%A3o/not%C3%ADcias/2019/tempestadeleslie-sinistros (site visited on February 07 2023) https://www.apseguradores.pt/pt/comunica%C3%A7%C3%A3o/not%C3%ADcias/2020/articleid/142/tempestade-elsa-e-fabien-%E2%80%93-balan%C3%A7o-final-de-dados-do-setor-segurador-22-7-mil-sinistrosparticipados-com-um-custo-estimado-de-42-milh%C3%B5es-de-euros (site visited on February 07 2023)

Source(s): Data collected from the Portuguese Association of Insurers. Table was authors’ own creation

Table 3

Definition of variables intensity1 and intensity2, capturing the effect of variables dist1 and dist2 on RAB and CR, respectively

	Intensity 1	Intensity 2
level 1	dist1 < 54 km	dist1 < 54 km
level 2	54 km ≤ dist1 < 78 km	54 km ≤ dist1 < 78 km
level 3	78 km ≤ dist1 < 107 km	78 km ≤ dist1 < 107 km
level 4	dist1 ≥ 107 km and dist2 < 44 km	dist1 ≥ 107 km and dist2 < 27 km
level 5	dist1 ≥ 107 km and dist2 ≥ 44 km	dist1 ≥ 107 km and dist2 ≥ 27 km