Reliability Engineering and System Safety 74 (2001) 239249

www.elsevier.com/locate/ress

Long-term hurricane risk assessment and expected damage

to residential structures
Zhigang Huang a, David V. Rosowsky b,*, Peter R. Sparks c
a
Applied Research Associates, Raleigh, NC, USA
b
Departments of Forest Products and Civil Engineering, Oregon State University, Corvallis, OR 97331-5751, USA
c
Department of Civil Engineering, Clemson University, Clemson, SC 29634-0911, USA
Accepted 11 August 2000

Abstract
This paper presents results from a study to evaluate long-term hurricane risks in the Southeastern United States using event-based
simulation procedures. These risks are dened by (1) the statistical extreme wind climate, and (2) the expected insured losses from damage
to residential structures. A probabilistic hurricane event model developed by the authors is used to evaluate long-term risks. The event model
parameters were derived from a statistical analysis of storms affecting the Southeastern United States and include radius of maximum winds,
central pressure difference, landfall location, storm track, and decay rate. The 50-year mean recurrence interval (MRI) gradient-level and
surface gust wind speeds are evaluated for the region investigated using results from the simulation analysis. When coupled with a damage
model, also developed by the authors, the results from the event-based simulation analysis are used to provide estimates of the expected
losses. The states of North Carolina, South Carolina, and Florida are used to demonstrate the applicability of this procedure for evaluating
expected losses. Implications for setting design wind speeds as well as risk-consistent insurance rates are discussed. q 2001 Elsevier Science
Ltd. All rights reserved.
Keywords: Damage model; Hurricane; Insured losses; Probability; Risk assessment; Simulation; Statistics; Wind speeds

1. Introduction related to the intensity of the storm or in terms of the

economic damage (losses) caused by the storm. In this
Hurricanes are among the greatest natural hazards affect- paper, risk is dened using the second denition as it incor-
ing communities in the United States. In recent years, wind- porates both the extreme wind climate and the vulnerability
storm catastrophes have caused enormous economic losses (susceptibility) of residential buildings (e.g.) to hurricanes.
and placed tremendous burdens on the insurance industry. Therefore, consideration of uncertainty in the risk assess-
Despite signicant improvements in predicting, tracking ment includes both uncertainties associated with the storm
and warning the public about hurricanes, there has been model and the economic loss model. Long-term hurricane
relatively little progress in our ability to estimate expected risk is dened herein through a statistical characterization of
hurricane losses. These losses can be in the form of the extreme wind climate and the associated expected
structural damage, damage to utilities or lifelines, or insurance losses to residential structures.
business interruptions. A number of models have been In order to estimate the hurricane risks in the Southeast-
developed to predict losses due to hurricanes, however, ern United States and provide timely and easily interpreted
these models are largely proprietary and are not available guidance to insurers, emergency managers and the public, a
to the public. As such, they are of little value to individuals GIS-based hurricane hazard assessment system was devel-
or organizations other than those for whom the models were oped to predict expected wind conditions and associated
developed (i.e. insurance companies). damages [1]. This paper reports on the application of the
Short-term hurricane risk can be dened either in terms wind eld model and event-based simulation procedure
developed in that study to the evaluation of design wind
speeds and expected hurricane damage to residential struc-
* Corresponding author. Address: Department of Forest Products, Oregon
State University, 104 Richardson Hall, Corvallis, OR 97331-5751, USA.
tures. Specically, event-based Monte Carlo simulation
Tel.: 11-541-737-8422; fax: 11-541-737-3385. techniques are used to statistically characterize the wind
E-mail address: [email protected] (D.V. Rosowsky). climate and expected insurance losses in the Southeastern
0951-8320/01/$ - see front matter q 2001 Elsevier Science Ltd. All rights reserved.
PII: S 0951-832 0(01)00086-2
240 Z. Huang et al. / Reliability Engineering and System Safety 74 (2001) 239249

United States. The occurrence of a hurricane is modeled as a To overcome these problems, Georgiou's model was modi-
Poisson process while the change in translational speed of a ed by Huang [10] to be able to better predict wind speeds.
hurricane over land is modeled as a Markov process. The Specically, a modication factor was applied to model's
results from the analysis are used to determine 50-year mean estimate of the gradient wind speed. The modication factor
recurrence interval (MRI) gradient-level wind speeds for the is a function of the central pressure difference (DP), the
region being investigated. The surface wind speeds can then radius of maximum wind speed (Rm), and the translation
be determined using appropriate gradient-to-surface conver- wind speed (Vt), as well as the relative distance to the hurri-
sion factors. Using a damage model, also developed by the cane center r=Rm : Surface wind speeds were then obtained
authors, the expected loss, dened as a percentage of the by reducing the gradient wind speed in a manner appropriate
total insured portfolio, can then be estimated. This damage for the terrain. The modied Georgiou's model was used in
model was developed using actual insurance loss (claim) the event-based simulation analysis in this study. Further
data from Hurricanes Hugo and Andrew. When coupled details of this model may be found elsewhere [1,10].
with the damage model, the results from the event-based Once a hurricane makes landfall, the energy balance
simulation analysis are used to provide estimates of the between the heat source and frictional dissipation is
expected losses. These can be presented on a 50-year disturbed due to the reduced availability of heat and
basis or on an annualized basis. While the former increased frictional dissipation. This results in a rise in
corresponds to a typical building design life, the latter atmospheric pressure in the storm's center, and conse-
may be of greater interest to the insurance industry. The quently, the hurricane weakens and the wind speeds
states of North Carolina, South Carolina, and Florida are decrease. The rise in the central pressure is most often used
used to demonstrate the applicability of this procedure for to model the weakening of a landfalling hurricane. Based on an
evaluating expected losses. analysis of historical hurricanes (see Ref. [10]), a decay model
having the following form was assumed

2. Wind eld model and Monte Carlo simulation Dpt Dp0 exp2at 1

A number of models have been used for hurricane wind where Dpt is the difference between the central pressure and
speed simulation and risk analysis studies [29]. The start- the atmopheric pressure at a distance beyond the effect of the
ing point for most of these models has been the gradient hurricane at time t; Dp0 ; the pressure difference when the
level wind eld. Huang [10] evaluated and compared a hurricane crossed the coast; a is the lling constant modeled
number of these models using data from Hurricanes Hugo, as a normally distributed random variable. The mean and
Fran, Bonnie, Earl, and Georges. In general, Batts' model standard deviation of the lling constant a for the three states
was found to overestimate the surface wind speeds, while investigated (North Carolina, South Carolina, and Florida) are
Georgiou's model was found to predict the surface wind shown in Table 1. This decay model is similar to that used by
speeds quite well at inland sites and at sites close to the Vickery and Twisdale [9], however their model generally
ocean when the wind was blowing from the ocean. suggests a more rapid decay rate. In this study, the correlation
However, Georgiou's model overpredicted the wind speeds between the lling rate and the intensity of the storm was not
at sites near the coast and underpredicted the gradient wind found to be statistically signicant. Further information on the
speeds in the region of most intense winds, i.e. the eye wall. decay model may be found elsewhere [10].

Table 1
Statistics of hurricane model parameters (from: Ref. [10]

Parameter Distribution Distribution parameters

North Carolina South Carolina Florida (Atlantic Coast) Florida (Gulf Coast)

Annual occurrence rate, l Poisson l 0:277 l 0:306 l 0:252 l 0:379

Approach angle, u (degrees) Normal m 2:19 m 220:88 m 260:05 m 34:42
s 42:77 s 44:41 s 24:79 s 29:78
Central pressure difference Weibull u 51:120 u 50:094 u 64:831 u 42:751
DP (mb)
k 3:155 k 2:304 p k 3:465 k 3:929
Radius of maximum wind Lognormal l 3:995 l ln260= Dp l 4:045 2 0:0083Dp l 3:984 2 0:012Dp
speed, Rmax (km)
z 0:275 z 0:461 z 0:451 z 0:350
Translation velocity, Vt (m/s) Lognormal l 1:787 l 1:805 l 1:616 l 1:734
z 0:513 z 0:456 z 0:365 z 0:418
Filling constant, a Normal m 0:032 m 0:042 z 0:021 z 0:024
s 0:025 s 0:016 s 0:014 s 0:033
 Z. Huang et al. / Reliability Engineering and System Safety 74 (2001) 239249 241

Probabilistic process models were used in the event-based 3. Damage model

Monte Carlo simulations in this study. Specically, the
occurrence of hurricanes was modeled as a Poisson Process Loss information from Hugo and Andrew was obtained
while a Markov chain was used to describe the evolution of from a large insurer. For uniformity, only data from stan-
the hurricane. Monte Carlo simulation involves the direct dard homeowners policies for single-family dwellings were
sampling from the distributions of all the random variables used. The total insured value of a property was assumed to
in the performance function being evaluated. Since the be 150% of the value of the structure. The insurer provided
generation of random variables is a relatively simple task, information on the number of policies, number of claims,
Monte Carlo techniques can be used efciently to simulate total insured value and the total amount paid in claims in
both discrete and continuous random processes. Simulation each zip (postal) code area. Zip codes with fewer than 20
therefore provides a framework for considering spatial policies were eliminated since the sample size was consid-
uncertainty and temporal uncertainty simultaneously, i.e. ered too small to be reliable. The information from Hurri-
in a time-dependent analysis. Using event-based hurricane cane Hugo covered 81,161 policies in South Carolina with
simulation as an example, the realizations in the time insured values totaling about $10.42 billion. The total
domain are generated rst (i.e. hurricanes are generated number of claims was 44,448 and the total claim amount
according to an arrival model). Then, realizations of the reached $247.4 million (2.4% of the total insured value).
random variables dening the gradient wind eld are gener- Information was collected from 118 zip codes in South
ated in the space domain. Using appropriate gradient-to- Carolina (31% of the state's 381 zip codes). Following
surface conversion factors, the surface wind speeds can Hurricane Andrew, information was collected from 71 zip
also be determined. The hurricane is then moved to the codes in Florida, covering 72,796 policies and with insured
next location and the wind eld is re-generated taking into values totaling about $12.36 billion. The total number of
account spatial changes such as decay. After the hurricane claims was 59,523 and the total claim amount reached
has degraded to the point that wind speeds are no longer $2.64 billion (21.3% of the total insured value).
signicant, the simulation proceeds to the next randomly The damage model was developed by relating the mean
generated hurricane event. These steps are repeated a surface wind speed to the claim ratio and damage ratio in
specied number of times, and the distributions of each zip code. The claim ratio is dened as the total number
maximum wind speeds (e.g.) are determined. These can of claims in a zip code divided by the total number of
then be used in the analysis of extreme wind climate and insurance policies in that zip code. The damage ratio is
hurricane risks. Further discussion about Monte Carlo dened as the amount paid out by the insurer divided by
simulation and its application to hurricane simulation can the total insured value. The reference wind speed was
be found in Ref. [10]. assumed to be the maximum mean surface wind speed,
Five basic variables were used to characterize the wind averaged over 10 min, which would be measured at a height
eld in this study: central pressure difference DP; radius of of 10 m in the middle of an imaginary airport, located at the
maximum winds (Rmax), approach angle (u ), translation geographical centroid of the zip code area. These speeds
velocity (Vt) and annual occurrence rate (l ). These event were determined by multiplying the maximum gradient
model parameters were determined from an analysis of wind speed at the zip-code centroid by the following factors:
historical landfalling hurricanes in the region of interest 0.60 for very exposed coastal islands with wind from the
[1]. Hurricane data covering 112 years (18871998) were sea, 0.50 for the same locations with the wind blowing from
used to determine the distribution and statistical moments the mainland, 0.50 for zip-codes with centroids within
(including possible correlations) of the ve basic variables. 10 km of the coast for winds approaching over the sea,
Three states (North Carolina, South Carolina, and Florida) 0.45 for these locations with winds approaching over the
in the Southeastern United States were investigated, and land, and 0.45 for all other zip code areas. The gradient
site-specic statistical information was determined for wind speeds for these two events (Hugo and Andrew)
each (see Table 1). Since the hurricanes can approach the were obtained using a wind eld model developed by the
Florida peninsula either from the Atlantic Ocean or the Gulf authors [12].
of Mexico, two different sets of statistics were developed to Most insurance losses occurred in suburban and wooded
model the characteristics of hurricanes approaching from areas. However, in a few areas affected by Andrew, the
each of these directions. The transition matrix for the houses were scattered in at, treeless terrain. In these
Markov analysis was developed using historical hurricane areas, the over-land boundary layer would develop very
track information which reported the storm's position every slowly as the wind passed from the sea to the land. Using
six hours [11]. The translational wind speed states in the the fetch factor in Ref. [13], it can be shown that in such
transition matrix therefore corresponded to the ratios of terrain, even 20 km inland, the mean wind speed could still
current states to the translational wind speed at landfall at be 60% of the gradient speed. This value was adopted for all
each six-hour interval. The simulation results were shown to such areas in South Florida. Thus, the effective mean wind
agree well with actual data [10]. Further details can be found speed is dened as the reference wind speed at the notional
elsewhere [1,10]. airport taking into account the local exposure.
242 Z. Huang et al. / Reliability Engineering and System Safety 74 (2001) 239249

Fig. 1. Claim ratio vs. effective mean surface wind speed.

Data from zip codes having similar wind conditions were time the speed reaches 30 m/s. Fig. 2 shows that above
weighted according to the total number of policies in each 35 m/s, the amount of damage increases rapidly, likely the
zip code (i.e. the wind speeds, claim ratios, and damage result of large amounts of rain entering buildings through
ratios were weighted averages). The relationship between breaches in the building envelopes. Considering the loss
the weighted claim ratio and effective surface wind speed is data shown in Figs. 1 and 2, the relationship between
shown in Fig. 1. The relationship between weighted damage claim ratio and effective mean surface wind speed follows
ratio and effective surface wind speed is shown in Fig. 2. It a double exponential form, while the relationship between
is interesting to note that these gures are essentially a the damage ratio and effective mean surface wind speed
quantication of the old Beaufort Scale, which suggested follows an exponential form. Using regression techniques,
that structural damage was likely to begin when the mean the claim ratio and damage ratio corresponding to a given
wind speed reached 20 m/s, and became widespread by surface mean wind speed x can therefore be obtained as [10]
about 30 m/s. Fig. 1 suggests that a few people will le
insurance claims in locations with mean wind speeds less
than 20 m/s, but that nearly everyone will le a claim by the FC x exp2exp20:239x 2 21:21p 100 2

Fig. 2. Damage ratio vs. effective mean surface wind speed.

 Z. Huang et al. / Reliability Engineering and System Safety 74 (2001) 239249 243

and
(
FD x exp0:252x 2 5:823 x # 41:4 m=s
3
FD x 100 x . 41:4 m=s
where FC x is the expected claim ratio in percent and FD x
is the expected damage ratio in percent. Once the maximum
mean surface wind speed at the centroid of each zip code is
known (or has been estimated), the damage ratio and claim
ratio in each zip code are obtained from Eqs. (2) and (3),
respectively.

4. Evaluation of N-year MRI wind speeds

The basic design wind speed for buildings in the United

States is the 3-second gust wind speed having a 2% annual
exceedance probability, i.e. a 50-year MRI (ASCE 7-95).
Assuming independent annual maximum wind speeds, the
probability of exceeding the N-year MRI wind speed in Fig. 3. 50-year MRI gradient wind speed contours (m/s) for the South-
m years is 1 2 1 2 1=Nm : Thus, the knowledge of the eastern United States.
m-year maximum wind speed distribution can be used to
evaluate the design wind speeds. In this paper, a value of wind speed of 60 m/s while the southern part of Florida has
m 50 was assumed since 50 years is the typical design life a 50-year MRI gradient wind speed of 68 m/s. However, in
for a building in the United States. Huang et al. [14] have order to determine design wind speeds at the surface, appro-
also shown that the problem of overestimating the extreme priate conversion factors that properly take into account site
wind climate in hurricane-prone regions by using equivalent characteristics must be applied to the 50-year MRI gradient
annual maximum wind speeds may be overcome by using wind speeds. Note that the 50-year MRI gradient wind speed
50-year maximum wind speeds. It is proposed herein that contours are almost equally spaced with distance from the
the gradient wind speed is an appropriate basis for charac- coast. However, due to the relatively quick transition from
terizing the wind climate in hurricane-prone regions. Since an over-water gust structure to an over-land gust structure
the gradient-level wind speed is unaffected by surface once a hurricane makes landfall [15], the surface gust wind
friction, the uncertainties associated with modeling the speed contours will not be equally spaced.
surface exposure, and thus the gradient-to-surface conver- Surface wind speeds for open terrain at a standard height
sion factors, do not inuence the N-year MRI gradient wind of 10 m can be obtained directly from the gradient wind
speed. The maximum wind speed values from 1000 speeds using appropriate gradient-to-surface conversion
simulated 50-year exposure periods were used as the basis factors. In the present study, gradient-to-gust conversion
for the estimating 50-year MRI gradient wind speeds for the factors of 0.90 and 0.80 are used for sites located directly
Southeastern United States in this study. For each 50-year on the beach and within 10 km of the coast, respectively,
period, hurricanes were simulated using the long-term risk when the wind is blowing from the water to the land. When
model described previously and the maximum gradient and the wind is blowing from the land to the water (with
surface wind speeds at the centroid of each zip code were increased surface roughness), the conversion factors change
recorded. The claim ratios and damage ratios associated to 0.80 and 0.72, respectively. For inland sites, the gust-to-
with each hurricane during the 50-year period were also gradient conversion factor is 0.72 regardless of wind
determined (using the damage model given by Eqs. (2) directions. Similarly, the ratios of mean wind speed to gradi-
and (3)), summed, and divided by 50 at the end of each ent wind speed are 0.65 and 0.50 for sites located directly on
50-year period to obtain the annual claim ratio and annual the beach and open sites, respectively, at initial landfall with
damage ratio for each zip code. The presumption of wind blowing from the water; 0.50 and 0.45, respectively,
independent hurricane events serves as justication for with wind blowing from the land; and 0.45 for open sites
this method. located far inland regardless of wind directions. The land
The 50-year MRI gradient wind speeds for the Southeast- exposure effects on the wind prole are assumed to change
ern United States determined using this procedure are abruptly with change in wind direction, i.e. the gradient-to-
shown in Fig. 3. The gradient wind speeds are seen to surface conversion factor changes immediately once the
drop gradually with distance from the coast, except in wind changes from blowing from the water to blowing
Florida where the gradient wind speeds increase moving from the land, or vise versa. Fig. 4 shows the cumulative
down the peninsula. Most of the coastal areas in North distribution functions for the 50-year maximum gust wind
Carolina and South Carolina have a 50-year MRI gradient speeds for several sites considered in this study. The
244 Z. Huang et al. / Reliability Engineering and System Safety 74 (2001) 239249

Fig. 4. 50-year maximum gust wind speed CDFs.

non-exceedance probability for the 50-year MRI (0.364) is Georgiou [7] reported a value of 37 m/s. Similarly, for the
indicated by the horizontal line. The gure indicates that southern tip of Florida, the 50-year MRI value was
wind speeds initially drop very rapidly with distance from 680:65 44:2 m=s; very close to the value (44 m/s)
the coast. The 50-year MRI gust wind speed is 52.9 m/s at found by Georgiou [7]. For inland areas, a gradient-to-
Folly Beach (located directly on the coast), 47.7 m/s at surface conversion factor of 0.45 was assumed [15]. The
Charleston City (10 km from the coast), and 43.2 m/s 50-year MRI mean wind speed for Columbia, SC, located
at Charleston Airport (25 km from the coast). Thereafter, about 180 km inland, was 500:45 22:5 m=s; which
the decrease is much more gradual, dropping to 41.0 m/s compares well with 22.7 m/s obtained using Georgiou's
at Shaw Air Force Base (150 km inland). A Lognormal model. To be compatible with the design wind speed deni-
distribution was found (rather than an Extreme Type I or tion (open terrain) in ASCE 7-95 [16], a uniform gradient-
Type II distribution as had been found in other studies) to to-surface conversion factor of 0.72 was used to obtain the
provide the best t to the simulated 50-year maximum wind surface gust wind speeds from the gradient wind contour
speeds for all sites considered in the Southeastern United map Fig. 3. The design wind speeds in ASCE 7-95 are
States [10]. considerably higher for coastal areas than those obtained
Wind speeds reported using different averaging times in this study. For example, the value for Charleston in
must be converted to a standard averaging time before ASCE 7-95 is 53 m/s (note that the hurricane importance
they can be compared. However, studies by Batts et al. factor of 1.05 has been taken out), while a value of
and Vickery and Twisdale [5,9] reported fastest-mile wind 600:72 43:2 m=s was found in this study. One possible
speeds and insufcient information was provided about their explanation for this difference is the assumption in ASCE 7-
assumed gust structures and mean wind proles to permit 95 of a very slow exposure transition from over-water to
accurate conversions to 3-second gusts or 10-minute mean over-land after a hurricane makes landfall. The basic design
speeds. Therefore, the results from the present study are wind speed map given in ASCE 7-95 was based on the
compared only with the results by Georgiou [7]. Since the simulation results from Refs. [5,7,9], all of which assumed
reported hourly mean wind speeds at different mileposts in a very gradual transition from over-water to over-land
Georgiou's study are values over water, a gradient-to- exposure. For example, Georgiou assumed the over-water
surface conversion factor of 0.65 [15] was used to estimate to over-land transition zone extended 50 km inland. In fact,
the 50-year MRI mean wind speed at various coastal this transition has been shown to occur relatively quickly in
locations. This resulted in very close agreement with an actual hurricane environment [15]. Moreover, in deriving
Georgiou's results. In the present study, the 50-year MRI the 3-second design gust wind speeds in ASCE 7-95, the
mean wind speed over water near Charleston, SC was gust factors developed by Krayer and Marshall [17] were
600:65 39 m=s; while Georgiou [7] found a value of applied to the simulated wind speeds. However, these gust
36 m/s. The 50-year MRI mean wind speed over water near factors were intended to be representative values over land.
Wilmington, NC was 580:65 37:7 m=s; while Applying the over-land gust factors to over-water mean
 Z. Huang et al. / Reliability Engineering and System Safety 74 (2001) 239249 245

Fig. 5. Expected annual claim ratio for each zip code in North Carolina.

wind speeds results in higher gust wind speeds. Therefore, 1891 to 1998 by Sparks and Huang [15] found the 50-year
the coastal region design wind speeds in ASCE 7-95 are MRI gust wind speed to be 43.9 m/s, very close to the value
higher than would be predicted in this study. Finally, of 580:72 41:8 m=s obtained in the present study.
ASCE 7-95 presents design wind speeds on the coast corre- Using annual maximum wind speed records, the 50-year
sponding to exposure C (open terrain) over land. However, the MRI gust wind speed for the regions not affected by hurri-
wind speeds given by Batts et al., Georgiou and Vickery and canes in South Carolina and North Carolina was found to be
Twisdale [5,7,9] for locations along the coast were represen- approximately 37 m/s and was independent of distance from
tative values over water. This may help to further explain the the coast. The 50-year MRI gust wind speed for Columbia
higher wind speeds in ASCE 7-95 than those obtained herein. based on the hurricane simulation analysis in this study was
Studies of recent hurricanes, including Hugo (1989), found to be 500:72 36 m=s: This suggests that both
Andrew (1992), Bertha (1996), Fran (1996), Bonnie hurricanes and thunderstorms contribute to the extreme
(1998), Earl (1998), and Georges (1998), have shown that wind climate in Columbia. This `transition zone' concept
the transition from an open water exposure to an open is further described in the paper by Huang et al. [14].
country exposure in an actual hurricane environment occurs
much faster than previously assumed [15]. Within 10
20 km after landfall, the gust structure and gradient-to- 5. Prediction of expected insured losses
surface ratios are almost the same as those for hurricanes
well inland. Studies of historical wind records, including The damage model described previously can be used to
sites for which more than 100 years of annual maximum determine the expected loss ratio at each zip code for a
wind speed data are available, found that for a typical single event (see Ref. [12]) or for a period of time, i.e.
coastal location such as Charleston, located about 25 km using the long-term risk model developed herein. The states
inland from the open sea, the 50-year MRI gust speed was of North Carolina, South Carolina, and Florida are used to
approximately 44 m/s [18]. This is similar to the value demonstrate the applicability of the wind eld model and
(43.2 m/s) obtained in the present study. An analysis of the damage model to the expected loss analysis. For each
annual maximum gust wind speeds in Wilmington from 50-year period, hurricanes were simulated and the

Fig. 6. Expected annual damage ratio for each zip code in North Carolina.
246 Z. Huang et al. / Reliability Engineering and System Safety 74 (2001) 239249

Fig. 7. Expected annual claim ratio for each zip code in South Carolina.

maximum gradient and surface wind speeds at the centroid

of each zip code were recorded. The claim ratios and
damage ratios associated with each hurricane during the
50-year period were also calculated, summed, and divided
by 50 at the end of each 50-year period to obtain the annual
claim ratio and annual damage ratio for each zip code. After
a specied number of simulations, the expected values and
distributions for the annual claim ratio and annual damage
ratio were determined. Figs. 510 show the simulated
expected annual claim ratio and damage ratio for each zip
code in North Carolina, South Carolina, and Florida, respec-
tively. These values have also been presented in tabular
form [10]. As seen in Fig. 11, the expected annual claim
ratio and damage ratio drop very quickly with distance from
the coast. For those zip codes on barrier islands in South
Carolina with open terrain (such as Isle of Palms, Sullivans
Island, St. Helena Island, and Hilton Head Island), the
expected annual damage ratio is about 2%. This implies Fig. 9. Expected annual claim ratio for each zip code in Florida.

Fig. 8. Expected annual damage ratio for each zip code in South Carolina. Fig. 10. Expected annual damage ratio for each zip code in Florida.
 Z. Huang et al. / Reliability Engineering and System Safety 74 (2001) 239249 247

account. These non-hurricane losses may be signicant

(from a percentage standpoint) for inland regions in which
these events contribute to the extreme wind climate.
Figs. 3 and 510 also suggest that South Carolina and
North Carolina have similar hurricane risks (considering
both the extreme wind climate and expected annual claim
and damage ratios), while those in Florida are much higher.
The maximum expected annual claim and damage ratios for
coastal areas in Florida are more than twice those in South
Carolina and North Carolina.
Figs. 12 and 13 show the cumulative distribution func-
tions for the annual claim ratio and annual damage ratio for
several sites in South Carolina. The drop in these ratios with
distance from the coast is quite obvious from these gures.
Since signicant landfalling hurricanes are relatively rare
events, variabilities in the annual claim ratio and damage
ratio are relatively large for the sites investigated; coef-
Fig. 11. Expected annual damage ratio vs. distance from coast.
cients of variation (COVs) averaged about 0.40 and 0.80 for
the annual claim ratio and damage ratio (in South Carolina),
that houses on these barrier islands will experience losses respectively. Claim ratio COVs were about 10% lower for
totaling 100% of the insured value, on average, every 50 coastal areas and about 20% higher for areas located well
years. However, only about 20 km inland, the expected inland. Damage ratio COVs were about 10% lower and
annual damage ratio drops to 0.20.3%, or about one about 10% higher for coastal and inland areas, respectively.
tenth of the value for the barrier islands. Further inland, Even though limited information is presently available to
the expected annual damage ratio drops to less than 0.1%. validate the long-term damage assessment procedure, these
The quick drop in the expected annual damage ratio in the results still provide meaningful information for insurance
rst 20 km from the coast is largely due to the fast transition regulators, particularly on a relative risk basis. Such
from over-water winds to over-land winds when a hurricane information could be used, for example, when setting
makes landfall. Beyond that, the effects of hurricane decay appropriate insurance rates.
contribute most to the gradual drop in the expected annual
damage ratio. Similar trends can be seen in North Carolina 6. Conclusions
and Florida. Note that only hurricane risks (losses) were
considered in this research; losses caused by extratropi- Event-based simulation techniques were used to statisti-
cal cyclones and thunderstorms were not taken into cally characterize the long-term hurricane risks in the

Fig. 12. Annual claim ratio CDFs.

248 Z. Huang et al. / Reliability Engineering and System Safety 74 (2001) 239249

Fig. 13. Annual damage ratio CDFs.

Southeastern United States. These risks were dened to References

include both the statistical extreme wind climate and the
expected insured losses for residential structures. The latter [1] Huang Z, Rosowsky DV, Sparks PR. Event-based hurricane simula-
were expressed as expected claim ratios and damage ratios. tion for the evaluation of wind speeds and expected insurance losses,
Proceedings of the 10th International Conference on Wind Engineer-
Historical hurricane events were used to calibrate the ing 1999, Copenhagen, Denmark.
proposed wind eld model and long-term station wind [2] Russell LR. Probability distributions for Texas Gulf Coast hurricanes
speed records were used to validate the simulation results. effects of engineering interest. PhD Dissertation, Stanford University,
The states of North Carolina, South Carolina, and Florida Stanford, CA, 1968.
were used to demonstrate the applicability of the proposed [3] Russell LR. Probability distributions for hurricane effects. ASCE J
Waterways, Harbors, Coastal Engng 1971;97(1):13954.
methodology for expected loss analysis. The major ndings
[4] Tryggvason BV, Surry D, Davenport AG. Predicting wind-induced
from this study were as follows: response in hurricane zones. ASCE J Struct Div 1976;102(12):233350.
[5] Batts ME, Cordes MR, Russell LR, Shaver JR, Simiu E. Hurricane
1. The 50-year MRI surface gust wind speeds over-water wind speeds in the United States, NBS Building Science Series 124.
obtained in the present study agree well with Ref. [7]. Washington, DC: US Department of Commerce, National Bureau of
Standards, 1980.
However, other studies, including those used as the basis [6] Tuleya RE, Bender MA, Kurihara Y. A simulation study of the land-
for design wind speeds in ASCE 7-95, may overestimate fall of tropical cyclones using a movable nested-mesh model. Mon
the 50-year MRI surface wind speeds over land near the Weath Rev 1984;112(1):12436.
coast. [7] Georgiou PN. Design wind speeds in tropical cyclone-prone regions.
2. Due to the rapid reduction in wind speeds, the expected PhD Dissertation. Department of Civil Engineering, University of
Western Ontario, Canada, 1985.
annual losses from hurricanes at sites just 20 km inland
[8] Sanchez-Sesma J, Aguirre J, Sen M. Simple modeling procedure for
are about an order of magnitude less than those on the estimation of cyclonic wind speeds. ASCE J Struct Engng
most highly exposed islands. 1988;114(2):35270.
3. Considering both extreme wind climate and expected [9] Vickery PJ, Twisdale LA. Prediction of hurricane wind speeds in the
insurance losses, Florida has a much greater hurricane United States. ASCE J Struct Engng 1995;121(11):16919.
[10] Huang Z. Stochastic models for hurricane hazard analysis. PhD
risk than North Carolina or South Carolina, which have
Dissertation. Department of Civil Engineering, Clemson University,
similar risks. Clemson, SC, 1999.
[11] Neuman CJ, Jarvinen BR, McAdie CJ, Elms JD. Tropical cyclones of
the North Atlantic Ocean, 18711992. US Department of Commerce-
Acknowledgements NOAA, 1997.
[12] Huang Z, Rosowsky DV, Sparks PR. Hurricane hazard assessment
system for damage to residential structures in South Carolina. Envir
Funding for this work has been provided by the South Geosci 2000;7(1):5765.
Carolina Sea Grant Consortium. The authors gratefully [13] Cook NJ. The Designer's guide to wind loading of building structures,
acknowledge this support. part I. Garston: Building Research Establishment, 1985.
 Z. Huang et al. / Reliability Engineering and System Safety 74 (2001) 239249 249

[14] Huang Z, Rosowsky DV, Sparks PR. Statistical modeling of mixed [16] ASCE. Minimum design loads for buildings and other structures
wind climate using event-based simulation and analysis of wind ASCE 7-95. New York: American Society of Civil Engineers, 1995.
records, Proceedings of the Eighth International Conference on Appli- [17] Krayer WR, Marshall WR. Gust factors applied to hurricane winds.
cations of Statistics and Probability 1999, Sydney, Australia. Bull Am Met Soc 1992;73:6137.
[15] Sparks PR, Huang Z. Wind speed characteristics in tropical cyclones, [18] Sparks PR, Saripalli P. The use of historical records to put modern
Proceedings of the 10th International Conference on Wind Engineer- wind storms into perspective, Proceedings of the Seventh US National
ing 1999, Copenhagen, Denmark. Conference on Wind Engineering 1993, UCLA, Los Angeles, CA.