Você está na página 1de 9

Ocean Engineering 38 (2011) 1338–1346

Contents lists available at ScienceDirect

Ocean Engineering
journal homepage: www.elsevier.com/locate/oceaneng

A comparison of estimators for the generalised Pareto distribution


Edward B.L. Mackay a,n, Peter G. Challenor b, AbuBakr S. Bahaj c
a
GL Garrad Hassan, St Vincent’s Works, Silverthorne Lane, Bristol BS2 0QD, UK
b
Ocean Observing and Climate Group, National Oceanography Centre, Southampton SO14 3 ZH, UK
c
Sustainable Energy Research Group, University of Southampton, Southampton SO17 1BJ, UK

a r t i c l e i n f o a b s t r a c t

Article history: The generalised Pareto distribution (GPD) is often used to model the distribution of storm peak wave
Received 27 April 2010 heights exceeding a high threshold, from which return values can be calculated. There are large
Accepted 15 June 2011 differences in the performance of various parameter and quantile estimators for the GPD. Commonly
Editor-in-Chief: A.I. Incecik
used estimation methods such as maximum likelihood or probability weighted moments are not
optimal, especially for smaller sample sizes. The performance of several estimators for the GPD is
Keywords: compared by the Monte Carlo simulation and the implications for estimating return values of
Extremes significant wave height are discussed. Of the estimators compared, the likelihood-moment (LM)
Peaks-over-threshold estimator has close to the lowest bias and variance over a wide range of sample sizes and GPD shape
Generalised Pareto distribution
parameters. The LM estimator always exists, is simple to compute and has a low sensitivity to choice of
Estimator
threshold. It is recommended that the LM estimator is used for calculating return values of significant
wave height when the sample size is less than 500. For sample sizes above 500 the NEW estimator of
Zhang and Stephens (2009) can give accurate results for low computational cost.
& 2011 Elsevier Ltd. All rights reserved.

1. Introduction various parameter and quantile estimators can vary greatly in


terms of their bias, variance and sensitivity to threshold choice,
The peaks-over-threshold (POT) method is commonly used to and consequently affect the accuracy of the estimated return
estimate return values of significant wave height (e.g. Mathiesen values.
et al.,1994; Ferreira and Guedes Soares, 1998; Naess, 1998; Caires The most commonly used methods for the GPD used for
and Sterl, 2005). In the POT method a distribution is fitted to the estimating return values of significant wave height are maximum
exceedances of a variable above a high threshold. Results from likelihood (ML), the method of moments (MOM) and probability
extreme value theory show that for any random variable from a weighted moments (PWM). Each of these methods has disadvan-
distribution, which satisfies certain regularity conditions, the tages, which although well known in statistical literature, have
distribution of exceedances of a high enough threshold will tend not been discussed in the offshore engineering literature. For
to a member of the generalised Pareto family (see e.g. Coles, 2001; example, the ML estimators of the GPD parameters are asympto-
Beirlant et al., 2004). This result has led to the generalised Pareto tically efficient (as the sample size tends to infinity the ML
distribution (GPD) being used in conjunction with the POT estimators achieve the Cramer–Rao lower bound for the variance
method to estimate return values of environmental variables. of an unbiased estimator). However, Hosking and Wallis (1987)
There are numerous factors, which affect the accuracy of showed that the ML estimates do not display this asymptotic
estimates of the estimated return values, such as the length and property for sample sizes up to 500. They also noted that some-
accuracy of data available, the criteria used to indentify indepen- times the solutions to the ML equations do not exist and that at
dent storm peaks, the choice of threshold and covariate effects. other times when the solutions do exist there can be convergence
In this paper we focus on the effect that the method used to problems with the algorithm they used to find them. Hosking and
estimate the parameters of the GPD has on the accuracy of the Wallis (1987) introduced the PWM estimators for the GPD and
estimated return values. This topic has received a great deal of compared them to the MOM and ML estimators. They showed
attention in the statistical literature, but relatively little attention that the MOM and PWM estimators have lower bias and variance
by the offshore engineering community. The performance of than ML estimators for sample sizes less than 500. However, both
the MOM and PWM estimators are sensitive to threshold choice
and sometimes result in non-feasible estimates, in that supports
n
Corresponding author. inferred from the estimates do not contain all the observations
E-mail address: ed.mackay@gl-garradhassan.com (E.B.L. Mackay). (Dupuis, 1996).

0029-8018/$ - see front matter & 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.oceaneng.2011.06.005
E.B.L. Mackay et al. / Ocean Engineering 38 (2011) 1338–1346 1339

Since the study of Hosking and Wallis there have been 3. Methods for estimating GPD parameters
numerous new estimators proposed for the parameters of the
GPD. An extensive discussion of the various methods that have Several estimators have been proposed for the threshold, u, at
been proposed has been presented by de Zea Bermudez and Kotz which the GPD can be considered a valid model for the data (e.g.
(2010). Although de Zea Bermudez and Kotz (2010) provide some Moharram et al., 1993; Sing and Guo, 1995; Dupuis, 1998). These
guidance on the relative merits of the various methods, they do methods remove the subjectivity in the choice of threshold, but
not provide a quantitative comparison of the performances using do not illustrate whether the GPD can be considered a valid model
a simulation study. Many simulation studies have been presented for the data. Confirming the validity of the model for the data is of
before (for recent examples see Ashkar and Nwentsa Tatsambon, vital importance if it is to be used to extrapolate outside the range
2007; Deidda and Puliga, 2009). However, these studies have of observations. Therefore the method preferred by most practi-
either compared only a few estimators or compared more tioners is to estimate the scale and shape parameters for a range
estimators but for a limited range of sample sizes or GPD shape of thresholds and examine the variation of certain statistics with
parameters. The performance of an estimator can vary consider- the threshold (see e.g. Coles, 2001). In the following we will
ably with both the sample size and the value of the GPD shape consider estimators for the shape and scale parameters only. Note
parameter. Therefore estimators that perform well in some that after pre-selection of the threshold u, the three-parameter
situations may perform poorly in others. The purpose of this GPD is reduced to a two-parameter distribution.
paper is to make a systematic comparison of the estimators that To keep paper to a reasonable length, we have only included
have been proposed for the GPD over a wide range of sample sizes the mathematical details where they are necessary for the
and shape parameters and make practical recommendations for discussion. For information on the numerical implementation of
the most suitable methods for estimation of return values of the methods the reader is referred either to the referenced papers
significant wave height, Hs (defined as 4 times the standard or to de Zea Bermudez and Kotz (2010) where the majority of the
deviation of the surface elevation). methods are discussed in detail. Throughout this section we will
The paper is arranged as follows: the generalised Pareto denote a sample of n observations from a common population as
distribution is defined in Section 2 and an overview of parameter x1, y , xn and denote the order statistics as x(1) rx(2) r y rx(n).
estimation methods is presented in Section 3. In Section 4, the
performance of the estimators is compared using the Monte Carlo
simulation. The results are discussed in Section 5 and an example 3.1. Moment based methods
is presented to illustrate the differences in practical situations.
Finally, conclusions are presented in Section 6. There are several approaches, which obtain estimators for GPD
parameters using various types of moments of the distribution.
The simplest approach, known as the method of moments (MOM)
obtains estimators by equating the sample mean and variance
2. The generalised Pareto distribution
with the population mean and variance (Eqs. (2) and (3)) and
rearranging. The use of the MOM estimators implicitly assumes
Let Y be a random variable and X ¼Y  u be the exceedances of
that x o1/2, since for x Z1/2 the variance is not defined.
Y over some high threshold u. Then X has a generalised Pareto
Following a similar idea to the method of moments, the
distribution if and only if the distribution function of X, condi-
sample probability weighted moments (PWMs) can be equated
tional on Y exceeding u, is
with the theoretical PWMs to obtain estimators for the para-
(
1ð1 þ xx=sÞ1=x for x a 0 meters. The PWMs of a variable X with distribution function F are
FðxÞ ¼ PrfX r x9Y 4 ug ¼ ð1Þ defined as
1expðx=sÞ for x ¼ 0
Z 1
where s 40 and for x Z0 the support is 0 rxoN, while for x o0 Mp,q,r ¼ EðX p F q ð1FÞr Þ ¼ ½XðFÞp F q ð1FÞr dF: ð5Þ
0
the support is 0 rxr  s/x. The parameters s and x are called the
scale and shape parameters, respectively. The generalised Pareto This method was first introduced for the GPD by Hosking and
family contains the uniform distribution when x ¼  1 and the Wallis (1987), who considered the use of the PWMs M1,0,0 and
exponential distribution when x ¼0. When x o0 the distribution M1,0,1. Rasmussen (2001) discussed parameter estimators using
has a finite end point and is referred to as short-tailed. When M1,0,r and M1,q,0, where q and r are not restricted to being
x 40 the distribution is referred to as heavy-tailed. non-negative integers, known as generalised probability weighted
The mean and variance of the distribution are given by moment (GPWM) estimators. Rasmussen concludes that the use
s of M1,0,r is preferable to M1,q,0 since the estimators can be
EðXÞ ¼ for x o1 ð2Þ expressed as explicit functions of the data. The optimal values
ð1xÞ
of r used in the GPWM estimator are defined as a function of the
and non-exceedance probability, F, of interest. For the simulations
s2 described in Section 4 a value of F¼0.999 has been used.
varðXÞ ¼ for x o 1=2: ð3Þ Similar to the MOM estimator, the PWM and GPWM estima-
ð1xÞ2 ð12xÞ
tors assume a restricted parameter space, in that M1,0,r only exists
When x Z1 the mean is not defined and when x Z1/2 the if r 4 1 and x or þ1. Therefore the PWM estimator assumes that
variance is not defined. In general, for x 40 the rth central x o1 and the GPWM estimator assumes x o0.75.
moment exists only if x o1/r. MOM, PWM and GPWM estimators always exist and are easy
The N-year return is obtained as the solution of to compute. However, they can sometimes produce estimates,
PrfY 4xN g ¼ 1=Nm, where m is the number of observations per which are not consistent with observed data, in that there are
year. If we define zu ¼ PrfY 4ug, then from Eq. (1) we have samples for which x^ is negative and xðnÞ 4s ^ =x^ , violating the
8 h i definition given in Eq. (1). The probability of samples occurring,
< u þ s ðNmz Þx 1 for x a0 which result in these ‘non-feasible’ parameter estimates, has been
x u
xN ¼ ð4Þ discussed by Dupuis (1996) and Ashkar and Nwentsa Tatsambon
: u þ slogðNmz Þ for x ¼ 0
u
(2007). They show that when x ¼  0.5 about 30% of samples
1340 E.B.L. Mackay et al. / Ocean Engineering 38 (2011) 1338–1346

results in non-feasible parameter estimates, but the occurrence distribution for the penalty function. Martins and Stedinger
decreases with increasing x. (2000) considered the use of a beta distribution with a mean of
Dupuis and Tsao (1998) introduced hybrid-MOM and PWM  0.1 and upper and lower limits of 70.5. Obviously this will
estimators, which are always feasible. These hybrid estimators tend to bias the estimate of x towards a value of 0.1, giving good
take the value of the MOM or PWM estimators if the results are performance when the true value of x is  0.1, but worse
feasible and set x^ ¼ s
^ =xðnÞ otherwise. We will extend this idea to performance for other values of x. It would be interesting to
the GPWM estimate also. In the rest of the paper we will only investigate the use of other beta distributions as penalty func-
consider these hybrid estimators. tions, to find the optimal penalty to apply in different situations,
Hosking (1990) introduced another type of estimator obtained but this is beyond the scope of the current paper. In the following
by considering expectations of linear combinations of order statis- we will only consider the PML estimators proposed by Coles and
tics, known as L-moments. In hydrological studies parameter Dixon (1999).
estimates are often expressed in terms of L-moments or LH- It is worth noting here that the maximum entropy (ME)
moments (higher-order L-moments). However, the L-moments estimators proposed by Sing and Guo (1995) are equivalent to
can be expressed as linear combinations of the PWMs, so they do the ML estimators in the case that the threshold is chosen before
not lead to anything new in terms of inference and we will not the parameters are estimated. As mentioned before, we are not
consider them further. considering the case when the threshold is estimated along with
the parameters. Therefore we will not consider the ME estimators
further.
3.2. Maximum likelihood

The likelihood function of independent observations x1, y, xn 3.3. Likelihood-moment estimation


from Eq. (1) is
Recently a hybrid between likelihood and moment estimators
Y
n
has been proposed by Zhang (2007). Since this is a relatively new
Lðs, xÞ ¼ f ðxj , s, xÞ ð6Þ
j¼1 method it will be discussed in some detail here. The likelihood-
moment (LM) estimators always exist, are always feasible and are
where f¼dF/dx. The maximum likelihood (ML) estimators are the simple to compute. They are found by solving the equation
values of s and x, which maximise Eq. (6). Grimshaw (1993) and
Chaouche and Bacro (2006) have both discussed the computation 1Xn
1
0 ¼ gðbÞ ¼ ð1bxj Þp  ð8Þ
of ML estimators of GPD parameters. They show that the problem nj¼1 1þ r
can be reduced to a one-dimensional search in order to speed up
computation. However, there are some samples for which no ML for b, where
solution exists. There has been some confusion about this point in rn
p ¼ Pn ð9Þ
the literature. Chaouche and Bacro (2006) and Ashkar and j¼1 logð1bxj Þ
Nwentsa Tatsambon (2007) claim that reports of samples where
and the parameter r 4  1 is chosen before the estimation. Having
no ML solutions exist are a result of the failure of the algorithm
solved for b the GPD parameter estimates are given by
used to find them (a claim repeated by de Zea Bermudez and Kotz,
2010). However, it is relatively straightforward to generate a 1Xn

sample for which no local maximum of the likelihood function


x^ ¼ logð1bxj Þ ð10Þ
nj¼1
exists and to verify this by plotting the likelihood surface. Hosking
and Wallis (1987) showed that this occurs most frequently for
s^ ¼ x^ =b: ð11Þ
small sample sizes and negative x. For a sample size of 25 over
20% of samples with x ¼ 0.5 have no ML solution. This decreases Zhang (2007) shows that the solution to Eq. (8) is simple to
to around 2% for a sample size of 50, and no failures were found obtain since g(b) is a smooth monotone function of b with a
for sample of size 100 and larger. These results were confirmed in unique solution in (  N,1/x(n)), unless r ¼0 or x1 ¼x2 ¼ y ¼xn. He
our simulations (presented in Section 4), using the algorithm of notes that a Newton–Raphson method will usually converge
Chaouche and Bacro (2006). within 4–6 iterations to a margin of relative error less than 10  6.
Coles and Dixon (1999) have noted that the superior perfor- Zhang (2007) also shows that if r ¼ x then the asymptotic
mance of MOM and PWM estimators to the ML estimator for variance of the LM estimator is equal to that of the ML estimator.
small sample sizes is due to the assumption of a restricted The variance of s ^ and x^ in finite samples was also presented for
parameter space, corresponding to finite population moments. several fixed values r ¼  2,  0.5 and 0.25, but no results were
To incorporate similar information into a likelihood-based infer- presented for the case, where r is estimated from the data. Instead
ence they suggest the use of a likelihood function, which pena- the following question was proposed: Given an initial estimate for
lises larger estimates of x (with an infinite penalty at x ¼1), x, it can be used as the value of r in Eq. (8) to obtain a new
similar to assuming a prior distribution for x. The corresponding estimate. If this procedure is iterated and each step uses the new
penalised likelihood function is given by estimate of x as r, will the LM estimates converge to the ML
estimates when they exist?
Lpen ðs, xÞ ¼ Lðs, xÞPðxÞ ð7Þ
We have investigated the answer to this question, using the
where P(x) is the penalty function. Estimators are found as the hybrid-PWM estimate of x as the initial guess for r and iterating
values of s and x, which maximise Eq. (7). In the following we will until x^ converges to r or x^ o 1 (at which point the algorithm
refer to these estimators as the penalised maximum likelihood fails, but this happens very rarely). It was found that the iterated
(PML) estimators. LM estimates did not converge to the ML estimates and that the
Martins and Stedinger (2000) propose a similar method for performance was not significantly better than the ordinary LM
estimating the parameters of the Generalised Extreme Value estimates with the hybrid-PWM estimate of x as the initial guess
distribution, which can easily be adapted to estimate the for r. An investigation was also made of the performance of the
parameters of the GPD. In their method the likelihood function LM estimator using the hybrid-MOM estimate of x as the initial
is penalised at both high and low values of x, using a beta guess for r. The performance was very similar but with marginally
E.B.L. Mackay et al. / Ocean Engineering 38 (2011) 1338–1346 1341

larger mean square error in small samples. In the following the 4. Monte Carlo simulations
LM estimate will refer to the non-iterated version using the
hybrid-PWM estimate of x as the initial guess for r. The performance of the estimators described in Section 3 was
compared using the Monte Carlo simulation. Since the results are
invariant to the value of s, we have arbitrarily chosen s ¼1. Trials
3.4. Least-squares type methods
have been made with n¼ 25, 50, 100, 200 and 500 and x ¼  0.5,
 0.45, y, 0.45, 0.5. The GPD is a valid probability distribution for
Two types of least-squares type estimators have been pro-
any value of x; however, we have restricted our interest to the
posed for the GPD. Moharram et al. (1993) proposed estimators,
range 0.5 r x r0.5 since these are the values, which are
which are obtained by minimising the sum of squared differences
commonly observed for environmental variables. In particular
between the empirical and model quantiles. Moharram et al.
the values of x observed for significant wave height are normally
(1993) used the method to estimate the threshold as well as the
in the range  0.25r x r0.25. For each value of n and x, 10,000
scale and shape parameters and noted that their algorithm
samples were generated and the estimates of s and x were
sometimes had convergence problems. No such problems have
calculated using each method (although for n¼500 the EPM
been encountered in our simulations, where the threshold is
estimators were not computed as they are far more computa-
chosen before the parameters are estimated.
tionally intensive than the other estimators and would have taken
Using a similar idea, Luceno (2006) proposed estimators found
several days to compute).
by minimising the squared differences between empirical and
Results will be presented here in terms of the quantile
model distribution functions, given in terms of various goodness-
estimates rather than the parameter estimates, since most often
of-fit statistics. Luceno considered several goodness-of-fit statistics,
we are interested in the accuracy of the predicted extreme values
including the Cramer–von Mises statistic (CM), the Anderson–
rather than the estimated parameters of the distribution. Quantile
Darling statistic (AD) and the right-tail weighted Anderson–
estimators are obtained by rearranging Eq. (1) to obtain an
Darling statistic (ADR). In the following these estimators will be
expression for x(F) and substituting the parameter estimators in
referred to as the CM, AD and ADR estimators, respectively, and the
place of the parameters. Presenting results in terms of the
estimator proposed by Moharram et al. (1993) will be referred to as
accuracy of the quantile estimators rather than parameter esti-
the LS estimator.
mators simplifies the comparison between methods since only
one quantity is being compared rather than two (the scale and
3.5. The empirical percentile method shape parameters).
We shall consider the bias and root-mean-square error (RMSE) of
Castillo and Hadi (1997) proposed an estimator found by quantiles at a non-exceedance probability of 0.999. Under the
equating percentiles of the empirical and theoretical distribution assumption that there are approximately 10 storms per year, which
functions, known as the empirical percentile method (EPM). The exceed the threshold, this corresponds to the 100-year return value.
EPM estimators always exist and are always feasible, but they can From an extensive analysis of buoy data from the North Pacific,
be very computationally intensive for larger samples since it is Mackay et al. (2010) found that the average number of storms per
necessary to solve n  1 equations by numerical methods in order year, which exceeded the threshold value, was between 4 and 20.
to compute the estimates. The variation in this range was partly due to natural sampling
effects and partly due to the choice of threshold used for each data
set. The performance of the various estimators is not entirely
3.6. Bayesian methods
independent of quantile under consideration, but from our investi-
gations those that had the best performance predicting quantiles at
De Zea Bermudez and Kotz (2010) reviewed several Bayesian
a non-exceedance probability of 0.999 also performed well at higher
methods for estimating the parameters of the GPD. They note that
and lower non-exceedance probabilities.
although they can produce good results, they are difficult to
Because of the large number of estimators and cases consid-
implement and are often computationally intensive. Recently
ered in this study, results are only presented for a subset of cases.
Zhang and Stephens (2009) have proposed a new method, based
Tables 1–5 list the relative bias and RMSE in quantile estimates
on maximum likelihood, but which uses a data-driven prior. This
for x ¼  0.5,  0.25, 0, 0.25 and 0.5. For a subset of the estimators
prior is chosen in such a way that the estimates always exist and
results are presented graphically in Figs. 1–5 for all cases
can be expressed as explicit functions of the observations. This
considered. The results have been presented in terms of bias
enables the estimates to be computed very efficiently, since the
and RMSE rather than bias and variance since the RMSE gives a
Markov chain Monte Carlo techniques (typically used in Bayesian
better indication of absolute accuracy. However, these statistics
inference) are not required to search for solutions. Following
are related by RMSE2 ¼bias2 þvariance.
Zhang and Stephens (2009) we will simply refer to this as the
The results for x 40 can be summarised as follows:
NEW estimator.
 The MOM, GPWM and PML estimators consistently have
3.7. Robust methods amongst the lowest RMSE of all the estimators, but this is at
the expense of a strong negative bias, caused by the assump-
Several robust methods estimators have been proposed for the tion of a restricted range for the shape parameter.
parameters of the GPD, which are less sensitive to outliers (e.g.  Of the estimators with low bias for x 40, the LM estimator has
Dupuis, 1998; Peng and Welsh, 2001; Juarez and Schucany, 2004). the lowest RMSE for small sample sizes.
Davison and Smith (1990) caution against the use of robust  For large sample sizes (n4200) the NEW estimator has a low
methods in extremes, arguing that the highest observations are bias and RMSE close to that of the ML estimator, but the NEW
those which are most informative and that careful checking of the estimator has the advantage that it is very fast to compute.
data by the analyst is preferable. Moreover, Juarez and Schucany  The PWM estimator has low bias, but higher RMSE than other
(2004) have shown that these robust estimators do not perform estimators.
as well as ML estimator when there is no contamination of the  The ML, LS, CM, AD, ADR, EPM and NEW estimators all have
data. Therefore robust methods are not considered further here. strong positive bias and high RMSE for small sample sizes.
1342 E.B.L. Mackay et al. / Ocean Engineering 38 (2011) 1338–1346

Table 1
Bias and RMSE of quantile estimators at a non-exceedance probability of 0.999 for n¼ 25.

x¼ Bias [%] RMSE [%]

 0.5  0.25 0 0.25 0.5  0.5  0.25 0 0.25 0.5

MOM 4.6 1.6  6.0  24.1  45.1 25.0 31.8 44.2 56.9 74.1
PWM 10.0 9.4 9.1 6.3  4.9 36.4 47.2 67.3 101.4 117.1
GPWM  3.5  3.9  6.6  20.2  41.1 23.8 38.8 52.5 62.2 64.5
ML  5.7  7.1 4.0 53.3 147.0 41.3 39.2 123.4 863.5 1231.1
PML  7.0  9.5  14.6  26.7  46.0 18.0 30.4 47.2 59.6 63.2
LM  4.2  7.2  10.0  10.2  16.4 18.0 28.5 51.2 96.1 126.7
LS 0.8 4.1 32.5 15114.1 1029372.6 21.7 42.2 162.3 465930.7 28499135.9
CM 87.4 123.4 406.9 644.1 1216.5 1222.6 846.0 4552.2 4093.3 13621.1
AD 40.5 220.4 283.5 1310.4 2737.4 269.6 4225.3 2407.8 22091.6 55414.0
ADR 11.0 29.0 75.6 209.9 441.8 56.2 175.1 488.8 1597.9 2882.8
EPM 12.5 34.8 118.6 617.1 41810.0 34.7 85.7 308.3 3896.1 1219638.8
NEW 14.5 26.9 54.8 121.2 255.1 43.4 85.0 188.0 1053.2 4685.4

Table 2
Bias and RMSE of quantile estimators at a non-exceedance probability of 0.999 for n¼ 50.

x¼ Bias [%] RMSE [%]

 0.5  0.25 0 0.25 0.5  0.5  0.25 0 0.25 0.5

MOM 3.2 1.0  3.4  17.9  39.5 16.9 22.5 33.8 47.0 62.2
PWM 5.9 5.4 4.6 4.2  1.6 22.6 30.7 45.1 72.8 96.7
GPWM  1.6  2.9  3.7  14.1  33.3 12.6 24.0 38.0 50.4 55.3
ML  5.7  6.2  1.1 15.3 46.1 11.4 21.9 48.7 117.0 295.5
PML  5.5  6.5  7.6  14.4  32.0 11.4 21.0 36.4 51.0 53.7
LM  2.0  3.7  5.5  5.6  10.7 11.7 20.0 36.9 68.7 95.6
LS  0.6 0.5 13.2 299.7 18781.7 12.6 22.8 69.3 3885.2 531883.8
CM 15.0 32.6 87.7 146.4 212.5 56.4 173.3 788.0 1080.1 954.9
AD 10.8 22.0 57.5 99.7 138.2 39.5 114.9 325.6 503.5 543.1
ADR 5.4 10.1 22.8 59.4 106.8 26.0 46.6 86.8 233.6 418.1
EPM 5.5 14.1 52.5 244.3 1022.5 15.3 39.5 124.5 2044.9 12752.1
NEW 4.8 9.2 18.7 36.7 60.0 14.8 29.5 62.6 132.3 254.6

Table 3
Bias and RMSE of quantile estimators at a non-exceedance probability of 0.999 for n¼ 100.

x¼ Bias [%] RMSE [%]

 0.5  0.25 0 0.25 0.5  0.5  0.25 0 0.25 0.5

MOM 2.2 0.7  1.9  12.4  34.1 11.4 15.9 25.1 38.1 51.9
PWM 3.7 3.1 2.3 2.8  0.8 14.4 20.7 31.0 52.0 73.1
GPWM  0.2  1.8  1.8  10.2  27.0 7.8 14.9 27.0 39.5 47.3
ML  3.4  3.6  1.7 4.8 18.1 7.2 13.8 29.9 56.4 97.6
PML  3.4  3.8  4.0  7.1  18.9 7.2 14.0 26.6 41.2 45.7
LM  0.9  1.9  2.9  3.4  6.5 7.7 13.8 26.3 51.0 75.3
LS  0.2  0.6 6.0 89.3 1901.1 8.4 14.7 37.7 1297.4 38039.0
CM 7.6 12.0 22.5 45.3 70.7 27.7 48.4 96.0 182.6 272.6
AD 4.5 8.1 18.6 31.8 57.7 16.1 31.0 82.9 132.1 204.7
ADR 2.2 5.1 10.2 22.4 38.7 11.5 23.1 44.9 85.1 135.4
EPM 2.4 7.5 27.6 104.6 463.9 8.1 20.0 65.7 373.4 6133.1
NEW 1.9 4.1 8.7 16.8 27.2 7.7 16.2 34.1 65.2 106.2

The results for x o0 can be summarised as follows: small samples (the values of bias and RMSE shown
for n ¼25 and 50 are for those samples where the solutions
 The LM estimator consistently has amongst the lowest RMSE exist).
and a small negative bias.  The LS estimator has low bias but not the lowest RMSE.
 The MOM, PWM and GPWM estimators have low bias but  The CM, AD, ADR and EPM estimators all have strong positive
larger RMSE than other methods in small samples. bias and high RMSE for small sample sizes, decreasing at large
 The ML and PML estimators have a small negative bias sample sizes (although the EPM estimator is extremely slow to
and low variance, but sometimes solutions do not exist in compute for large sample sizes).
E.B.L. Mackay et al. / Ocean Engineering 38 (2011) 1338–1346 1343

Table 4
Bias and RMSE of quantile estimators at a non-exceedance probability of 0.999 for n¼ 200.

x¼ Bias [%] RMSE [%]

 0.5  0.25 0 0.25 0.5  0.5  0.25 0 0.25 0.5

MOM 1.3 0.4  1.0  9.5  29.7 7.6 10.9 18.6 30.2 41.0
PWM 2.8 1.6 0.9 1.3 2.2 10.0 14.0 22.2 35.8 65.2
GPWM 0.4  0.9  1.0  7.3  22.1 5.4 9.7 19.0 30.4 39.8
ML  2.1  2.0  0.9 2.6 9.3 4.6 9.3 20.3 35.9 58.5
PML  2.0  2.2  2.0  3.2  10.2 4.6 9.4 19.0 31.7 37.3
LM  0.3  1.2  1.5  1.9  1.4 5.3 9.5 19.0 35.6 65.8
LS 0.0  0.2 3.1 31.5 141.1 5.6 9.8 24.5 208.6 1056.8
CM 4.4 6.3 10.8 19.1 29.0 14.5 26.9 50.9 84.1 126.9
AD 2.3 4.2 8.7 14.9 24.9 9.2 17.9 38.3 59.0 84.3
ADR 1.6 2.7 4.6 10.9 16.4 7.2 14.0 27.5 48.5 68.1
EPM 1.4 4.9 18.3 58.3 164.4 5.0 13.2 44.4 150.5 732.6
NEW 0.8 1.6 4.5 8.6 13.9 4.5 10.1 21.3 39.7 59.6

Table 5
Bias and RMSE of quantile estimators at a non-exceedance probability of 0.999 for n¼ 500.

x¼ Bias [%] RMSE [%]

 0.5  0.25 0 0.25 0.5  0.5  0.25 0 0.25 0.5

MOM 0.9 0.2  0.3  5.3  25.0 4.7 7.1 12.0 22.6 37.2
PWM 1.4 0.9 0.5 0.6 0.3 5.8 8.9 13.6 21.9 39.7
GPWM 0.6  0.1  0.5  4.9  16.2 3.4 5.9 11.9 21.2 32.3
ML  1.0  0.9  0.4 0.7 2.3 2.6 5.7 12.3 21.0 31.6
PML  1.0  1.1  0.9  1.2  3.9 2.6 5.6 12.0 20.6 27.1
LM  0.1  0.4  0.9  0.9  1.5 3.0 5.8 11.8 22.9 40.6
LS 0.0  0.1 1.7 11.8 53.7 3.4 6.0 14.9 55.6 363.3
CM 2.2 2.6 4.1 6.9 10.6 7.7 14.6 26.6 40.2 55.4
AD 1.3 1.7 2.8 5.5 7.3 5.0 9.7 19.0 31.0 43.0
ADR 0.8 1.1 2.3 3.4 6.9 4.1 8.2 15.7 26.2 38.7
NEW 0.2 0.8 1.7 3.1 4.9 2.5 5.8 12.6 22.3 32.5

Fig. 1. Bias and RMSE of quantile estimators at a non-exceedance probability of 0.999 for n¼ 25.

 The NEW estimator has a positive bias and high RMSE for the fastest to compute and the NEW estimator has very similar
small sample sizes but has low bias and RMSE at high sample bias and RMSE, with the advantage that it is very fast to compute.
sizes and is very fast to compute. It is interesting to note that Zhang and Stephens (2009) showed
that the NEW estimator has amongst the highest efficiencies for
It is clear that there is no one estimator, which stands out as parameter estimates in small sample sizes. However, the small
being the best in all situations. For positive x the lowest RMSE is positive bias in the estimate of the shape parameter causes
achieved at expense of negative bias. This is problematic when positive bias and high RMSE in quantile estimators at low
estimating extremes as it is usually preferable to use conservative exceedance probabilities.
estimates of return values. For small sample sizes of 200 or less it
could be argued that the LM estimator has the best all round
performance, since it has amongst the lowest bias and RMSE for 5. Discussion and example
all values of x in the range considered here. The LM estimator
always exists, is always feasible and is reasonably fast to The results presented in the previous section have compared
compute. For large samples, although the ML is optimal it is not the accuracy of estimators under the assumption that threshold
1344 E.B.L. Mackay et al. / Ocean Engineering 38 (2011) 1338–1346

Fig. 2. Bias and RMSE of quantile estimators at a non-exceedance probability of 0.999 for n ¼50.

Fig. 3. Bias and RMSE of quantile estimators at a non-exceedance probability of 0.999 for n¼ 100.

Fig. 4. Bias and RMSE of quantile estimators at a non-exceedance probability of 0.999 for n¼ 200.

Fig. 5. Bias and RMSE of quantile estimators at a non-exceedance probability of 0.999 for n¼ 500.

exceedances follow a generalised Pareto distribution. In practical importance if it is to be used to extrapolate outside the range of
situations a threshold must be selected for which the GPD is a observations. As mentioned earlier on, the method preferred by
valid approximating distribution for the data. Selecting a thresh- most practitioners is to calculate parameter estimates for a range
old and demonstrating that the GPD is a valid model is of critical of thresholds and examine the variation of certain statistics with
E.B.L. Mackay et al. / Ocean Engineering 38 (2011) 1338–1346 1345

Fig. 6. Variation of parameter and return value estimates with threshold.

the threshold. If the GPD is a valid model for a threshold u0 then accompanying text). Obviously, using an estimator with a low
for u 4u0 the estimates of the shape parameter, normalised scale sensitivity to threshold choice and low RMSE will aid the
parameter (sn ¼ s  xu) and return values should tend to constant identification of a valid threshold.
values (Coles, 2001). Confidence bounds estimated using a bootstrap technique give
The differences between several estimators are illustrated with an estimate of the variance of the return values and do not take
an example using buoy measurements of significant wave height. into account the bias, since this is not known in real situations.
The data comes from NDBC buoy 46002, located in the North East The RMSE calculated in the previous section gives a better
Pacific, 500 km off the coast of Oregon. The data covers the period indication of the absolute accuracy of the return values from
1978–2008 and consists of 20 min samples at hourly intervals. various methods, but it does not account for uncertainty due to
Following the recommendation of Forristall et al. (1996), a 3 h violation of the model assumptions, such as that of stationary
moving average filter has been applied to the data to reduce the distribution.
effect of sampling variability. Storm peaks are identified as local
maxima with a minimum separation of 36 hours.
Fig. 6 shows the variation of the estimates of the shape 6. Conclusions
parameter, normalised scale parameter and 100-year return values
using the MOM, PWM and LM estimators with threshold value, as The most commonly used estimators for the GPD in oceano-
well as the number of exceedances at each threshold value. It is graphic studies are the MOM, PWM and ML estimators, all of
clear that the PWM estimator is very sensitive to the threshold which have disadvantages. The ML estimator is asymptotically
value, making it difficult to identify a clear region of stability in the efficient, but for small samples the ML estimator does not always
parameters. The MOM estimator is slightly more sensitive than the exist and when it does exist it has a larger bias and variance than
LM estimator, displaying a slightly higher variation with threshold other estimators. The MOM and PWM estimators always exist and
value. These three estimators have been chosen to illustrate the are simple to compute, but are sensitive to the threshold value.
differences in sensitivity to threshold and it should be noted that, The LM estimator introduced by Zhang (2007) always exists, is
although not shown, the GPWM and PML estimators gave very simple to compute and has a low sensitivity to threshold choice.
similar results to the LM estimator. In this study we have extended the results of Zhang (2007) to
The data set used in this example is quite long and often return show that the LM estimator using the hybrid-PWM estimate as
values must be estimated using shorter data sets. Despite the use the first guess performs well for all sample sizes and values of
of a large data set there is considerable variation in these shape parameter. The bias and RMSE of the LM estimator are
estimates with the threshold and there is no clear choice of consistently the lowest or close to the lowest of all the estimators
threshold above which the estimates tend towards constant tested. For large sample sizes the NEW estimator proposed by
values. The addition of confidence bounds can aid threshold Zhang and Stephens (2009) has a bias and variance very close to
selection. Mackay et al. (2010) used a modified bootstrap techni- those of the ML estimator, but the NEW estimator has the
que to estimate confidence bounds for the parameters, which also advantage that it is very fast to compute. It is recommended that
smoothes the variation with threshold (see their Figs. 2 and 3 and for small sample sizes the LM estimator, using the hybrid-PWM
1346 E.B.L. Mackay et al. / Ocean Engineering 38 (2011) 1338–1346

estimator as a first guess, should be used in preference to the Ferreira, J.A., Guedes Soares, C., 1998. An application of the peaks over threshold
other estimators considered in this study. In the fortunate situa- method to predict extremes of significant wave height. J. Offshore Mech. Arctic
Eng. 120, 165–176.
tion that a large sample of size 500 or upwards is available, the Forristall, G.Z., Heideman, J.C., Leggett, I.M., Roskam, B., Vanderschuren, L., 1996.
NEW estimator will give accurate results for very reasonable Effect of sampling variability on hindcast and measured wave heights. J.
computational cost. Waterway, Port, Coastal Ocean Eng. 122 (5), 216–225.
Grimshaw, S.D., 1993. Computing maximum likelihood estimates for the general-
ized Pareto distribution. Technometrics 35, 185–191.
References Hosking, J.R.M., 1990. L-moments: analysis and estimation of distributions using
linear combinations of order statistics. J. R. Stat. Soc. B 52 (1), 105–124.
Hosking, J.R.M., Wallis, J.R., 1987. Parameter and quantile estimation for the
Ashkar, F., Nwentsa Tatsambon, C., 2007. Revisiting some estimation methods for generalized Pareto distribution. Technometrics 29, 339–349.
the generalized Pareto distribution. J. Hydrol. 346, 136–143. Juarez, S.F., Schucany, W.R., 2004. Robust and efficient estimation for the general-
Beirlant, J., Goegebeur, Y., Segers, J., Teugels, J., 2004. Statistics of Extremes. John ized Pareto distribution. Extremes 7 (3), 237–251.
Wiley and Sons 490 pp. Luceno, A., 2006. Fitting the generalized Pareto distribution to data using
Caires, S., Sterl, A., 2005. 100-Year return value estimates for ocean wind speed maximum goodness-of-fit estimators. Comput. Stat. Data Anal. 51, 904–917.
and significant wave height from the ERA-40 data. J. Clim. 18 (7), 1032–1048. Mackay, E.B.L., Challenor, P.G., Bahaj, A., 2010. On the use of discrete seasonal and
Castillo, E., Hadi, A.S., 1997. Fitting the generalized Pareto distribution to data. J. directional models for the estimation of extreme wave conditions. Ocean Eng.
Am. Stat. Assoc. 92 (440), 1609–1620. 37, 425–442.
Chaouche, A., Bacro, J.N., 2006. Statistical inference for the generalized Pareto Martins, E.S., Stedinger, J.R., 2000. Generalized maximum-likelihood generalized
distribution: maximum likelihood revisited. Commun. Stat.: Theory Methods extreme-value quantile estimators for hydrological data. Water Resour. Res.
35 (5), 785–802. 36 (3), 737–744.
Coles, S., 2001. An Introduction to the Statistical Modelling of Extreme Values. Mathiesen, M., Goda, Y., Hawkes, P., Mansard, E., Martin, M.J., Peltier, E.,
Springer-Verlag, London. Thompson, E., van Vledder, G., 1994. Recommended practice for extreme
Coles, S., Dixon, M.J., 1999. Likelihood-based inference for extreme value models. wave analysis. J. Hydraul. Res. 32 (6), 803–814.
Extremes 2 (1), 5–23. Moharram, S.H., Gosain, A.K., Kapoor, P.N., 1993. A comparative study for the
Davison, A.C., Smith, R.L., 1990. Models for exceedances over high thresholds. J. R. estimators of the generalized Pareto distribution. J. Hydrol. 150, 169–185.
Stat. Soc. B 52 (3), 393–442. Naess, A., 1998. Statistical extrapolation of extreme value data based on the Peaks
de Zea Bermudez, P., Kotz, S., 2010. Parameter estimation of the generalized Pareto Over Threshold method. J. Offshore Mech. Arctic Eng. 120, 91–96.
distribution—Parts I & II. J. Stat. Plann. Inference 140, 1353–1388. Peng, L., Welsh, A.H., 2001. Robust estimation of the generalized Pareto distribu-
Deidda, R., Puliga, M., 2009. Performances of some parameter estimators of the tion. Extremes 4 (1), 53–65.
generalized Pareto distribution over rounded-off samples. Phys. Chem. Earth Rasmussen, P.F., 2001. Generalized probability weighted moments: application to
34, 626–634. the generalized Pareto distribution. Water Resour. Res. 37 (6), 1745–1751.
Dupuis, D.J., 1996. Estimating the probability of obtaining nonfeasible parameter Singh, V.P., Guo, H., 1995. Parameter estimation for 3-parameter generalized
estimates of the generalized Pareto distribution. J. Stat. Comput. Simulation Pareto distribution by the principle of maximum entropy (POME). Hydrol. Sci.
54, 197–209. J. 40, 165–181.
Dupuis, D.J., 1998. Exceedances over high thresholds: a guide to threshold Zhang, J., 2007. Likelihood moment estimation for the generalised Pareto distribu-
selection. Extremes 1 (3), 251–261. tion. Aust. N. Z. J. Stat. 49 (1), 69–77.
Dupuis, D.J., Tsao, M., 1998. A hybrid estimator for generalized Pareto and Zhang, J., Stephens, M.A., 2009. A new and efficient estimation method for the
extreme-value distributions. Commun. Stat.: Theory Methods 27, 925–941. generalized Pareto distribution. Technometrics 51 (3), 316–325.

Você também pode gostar