| Title: | Threshold Estimation Approaches |
|---|---|
| Description: | Different approaches for selecting the threshold in generalized Pareto distributions. Most of them are based on minimizing the AMSE-criterion or at least by reducing the bias of the assumed GPD-model. Others are heuristically motivated by searching for stable sample paths, i.e. a nearly constant region of the tail index estimator with respect to k, which is the number of data in the tail. The third class is motivated by graphical inspection. In addition, a sequential testing procedure for GPD-GoF-tests is also implemented here. |
| Authors: | Johannes Ossberger |
| Maintainer: | Johannes Ossberger <[email protected]> |
| License: | GPL-3 |
| Version: | 1.1 |
| Built: | 2024-09-17 05:59:03 UTC |
| Source: | CRAN |
This package contains implementations of many of the threshold estimation approaches proposed in the literature. The estimation of the threshold is of great interest in statistics of extremes. Estimating the threshold is equivalent to choose the optimal sample fraction in tail index estimation. The sample fraction is given by k/n with n the sample size and k the number of extremes in the data or, if you wish, the exceedances over a high unknown threshold u.
| Package: | tea |
| Type: | Package |
| Version: | 1.1 |
| Date: | 2020-04-17 |
| License: | GPL-3 |
Johannes Ossberger
Maintainer: Johannes Ossberger <[email protected]>
Caeiro and Gomes (2016) <doi:10.1201/b19721-5>
Cebrian et al. (2003) <doi:10.1080/10920277.2003.10596098>
Danielsson et al. (2001) <doi:10.1006/jmva.2000.1903>
Danielsson et al. (2016) <doi:10.2139/ssrn.2717478>
De Sousa and Michailidis (2004) <doi:10.1198/106186004X12335>
Drees and Kaufmann (1998) <doi:10.1016/S0304-4149(98)00017-9>
Hall (1990) <doi:10.1016/0047-259X(90)90080-2>
Hall and Welsh (1985) <doi:10.1214/aos/1176346596>
Kratz and Resnick (1996) <doi:10.1080/15326349608807407>
Gomes et al. (2011) <doi:10.1080/03610918.2010.543297>
Gomes et al. (2012) <doi:10.1007/s10687-011-0146-6>
Gomes et al. (2013) <doi:10.1080/00949655.2011.652113>
G'Sell et al. (2016) <doi:10.1111/rssb.12122>
Guillou and Hall <doi:10.1111/1467-9868.00286>
Reiss and Thomas (2007) <doi:10.1007/978-3-0348-6336-0>
Resnick and Starica (1997) <doi:10.1017/S0001867800027889>
Thompson et al. (2009) <doi:10.1016/j.coastaleng.2009.06.003>
Plots the Alternative Hill Plot and an averaged version of it against the upper order statistics.
althill(data, u = 2, kmin = 5, conf.int = FALSE)althill(data, u = 2, kmin = 5, conf.int = FALSE)
data |
vector of sample data |
u |
gives the amount of which the Hill estimator is averaged. Default ist set to |
kmin |
gives the minimal |
conf.int |
|
The Alternative Hill Plot is just a normal Hill Plot scaled to the [0,1] interval which can make interpretation much easier. See references for more information.
The normal black line gives a simple Hill Plot scaled to [0,1]. The red dotted line is an averaged version that smoothes the Hill Plot by taking the mean of k(u-1) subsequent Hill estimations with respect to k. See references for more information.
Resnick, S. and Starica, C. (1997). Smoothing the Hill estimator. Advances in Applied Probability, 271–293.
data=rexp(500) althill(data)data=rexp(500) althill(data)
Plots an averaged version of the classical Hill Plot
avhill(data, u = 2, kmin = 5, conf.int = FALSE)avhill(data, u = 2, kmin = 5, conf.int = FALSE)
data |
vector of sample data |
u |
gives the amount of which the Hill estimator is averaged. Default ist set to |
kmin |
gives the minimal |
conf.int |
|
The Averaged Hill Plot is a smoothed version of the classical Hill Plot by taking the mean of values of the Hill estimator for subsequent k, i.e. upper order statistics. For more information see references.
The normal black line gives the classical Hill Plot. The red dotted line is an averaged version that smoothes the Hill Plot by taking the mean of k(u-1) subsequent Hill estimations with respect to k. See references for more information.
Resnick, S. and Starica, C. (1997). Smoothing the Hill estimator. Advances in Applied Probability, 271–293.
data(danish) avhill(danish)data(danish) avhill(danish)
Gives the optimal number of upper order statistics k for the Hill estimator by minimizing the AMSE-criterion.
dAMSE(data)dAMSE(data)
data |
vector of sample data |
The optimal number of upper order statistics is equivalent to the number of extreme values or, if you wish, the number of exceedances in the context of a POT-model like the generalized Pareto distribution. This number is identified by minimizing the AMSE criterion with respect to k. The optimal number, denoted k0 here, can then be associated with the unknown threshold u of the GPD by choosing u as the n-k0th upper order statistic. For more information see references.
second.order.par |
gives an estimation of the second order parameter |
k0 |
optimal number of upper order statistics, i.e. number of exceedances or data in the tail |
threshold |
the corresponding threshold |
tail.index |
the corresponding tail index |
Caeiro, J. and Gomes, M.I. (2016). Threshold selection in extreme value analysis. Extreme Value Modeling and Risk Analysis:Methids and Applications, 69–86.
data(danish) dAMSE(danish)data(danish) dAMSE(danish)
An Implementation of the procedure proposed in Danielsson et al. (2001) for selecting the optimal sample fraction in tail index estimation.
danielsson(data, B = 500, epsilon = 0.9)danielsson(data, B = 500, epsilon = 0.9)
data |
vector of sample data |
B |
number of Bootstrap replications |
epsilon |
gives the amount of the first resampling size |
The Double Bootstrap procedure simulates the AMSE criterion of the Hill estimator using an auxiliary statistic. Minimizing this statistic gives a consistent estimator of the sample fraction k/n with k the optimal number of upper order statistics. This number, denoted k0 here, is equivalent to the number of extreme values or, if you wish, the number of exceedances in the context of a POT-model like the generalized Pareto distribution. k0 can then be associated with the unknown threshold u of the GPD by choosing u as the n-k0th upper order statistic. For more information see references.
second.order.par |
gives an estimation of the second order parameter |
k0 |
optimal number of upper order statistics, i.e. number of exceedances or data in the tail |
threshold |
the corresponding threshold |
tail.index |
the corresponding tail index |
Danielsson, J. and Haan, L. and Peng, L. and Vries, C.G. (2001). Using a bootstrap method to choose the sample fraction in tail index estimation. Journal of Multivariate analysis, 2, 226-248.
data=rexp(100) danielsson(data, B=200)data=rexp(100) danielsson(data, B=200)
These data describe large fire insurance claims in Denmark from Thursday 3rd January 1980 until Monday 31st December 1990. The data are contained in a numeric vector. They were supplied by Mette Rytgaard of Copenhagen Re
data("danish")data("danish")
The format is: atomic [1:2167] 1.68 2.09 1.73 1.78 4.61 ... - attr(*, "times")= POSIXt[1:2167], format: "1980-01-03 01:00:00" "1980-01-04 01:00:00" ...
The data is taken from package evir.
data(danish)data(danish)
An Implementation of the procedure proposed in Drees & Kaufmann (1998) for selecting the optimal sample fraction in tail index estimation.
DK(data, r = 1)DK(data, r = 1)
data |
vector of sample data |
r |
tuning parameter for the stopping criterion. |
The procedure proposed in Drees & Kaufmann (1998) is based on bias reduction. A stopping criterion with respect to k is implemented to find the optimal tail fraction, i.e. k/n with k the optimal number of upper order statistics. This number, denoted k0 here, is equivalent to the number of extreme values or, if you wish, the number of exceedances in the context of a POT-model like the generalized Pareto distribution. k0 can then be associated with the unknown threshold u of the GPD by choosing u as the n-k0th upper order statistic. If the above mentioned stopping criterion exceedes a certain value r, the bias of the assumed extreme model has become prominent and therefore k should not be chosen higher. For more information see references.
second.order.par |
gives an estimation of the second order parameter |
k0 |
optimal number of upper order statistics, i.e. number of exceedances or data in the tail |
threshold |
the corresponding threshold |
tail.index |
the corresponding tail |
Drees, H. and Kaufmann, E. (1998). Selecting the optimal sample fraction in univariate extreme value estimation. Stochastic Processes and their Applications, 75(2), 149–172.
data(danish) DK(danish)data(danish) DK(danish)
An Implementation of the so called Eye-balling Technique proposed in Danielsson et al. (2016)
eye(data, ws = 0.01, epsilon = 0.3, h = 0.9)eye(data, ws = 0.01, epsilon = 0.3, h = 0.9)
data |
vector of sample data |
ws |
size of the moving window. |
epsilon |
size of the range in which the estimates can vary |
h |
percentage of data inside the moving window that should lie in the tolerable range |
The procedure searches for a stable region in the Hill-Plot by defining a moving window. Inside this window the estimates of the Hill estimator with respect to k have to be in a pre-defined range around the first estimate within this window. It is sufficient to claim that only h percent of the estimates within this window lie in this range. The smallest k that accomplishes this is then the optimal number of upper order statistics, i.e. data in the tail.
k0 |
optimal number of upper order statistics, i.e. number of exceedances or data in the tail |
threshold |
the corresponding threshold |
tail.index |
the corresponding tail index by plugging in |
Danielsson, J. and Ergun, L.M. and de Haan, L. and de Vries, C.G. (2016). Tail Index Estimation: Quantile Driven Threshold Selection.
data(danish) eye(danish)data(danish) eye(danish)
Performs a sequential Mann-Kendall Plot also known as Gerstengarbe Plot.
ggplot(data, nexceed = min(data) - 1)ggplot(data, nexceed = min(data) - 1)
data |
vector of sample data |
nexceed |
number of exceedances. Default is the minimum of the data to make sure the whole dataset is considered. |
The Gerstengarbe Plot, referring to Gerstengarbe and Werner (1989), is a sequential version of the Mann-Kendall-Test. This test searches for change points within a time series. This method is adopted for finding a threshold in a POT-model. The basic idea is that the differences of order statistics of a given dataset behave different between the body and the tail of a heavy-tailed distribution. So there should be a change point if the POT-model holds. To identify this change point the sequential test is done twice, for the differences from start to the end of the dataset and vice versa. The intersection point of these two series can then be associated with the change point of the sample data. For more informations see references.
k0 |
optimal number of upper order statistics, i.e. the change point of the dataset |
threshold |
the corresponding threshold |
tail.index |
the corresponding tail index |
Ana Cebrian Johannes Ossberger
Great thanks to A. Cebrian for providing a basic version of this code.
Gerstengarbe, F.W. and Werner, P.C. (1989). A method for statistical definition of extreme-value regions and their application to meteorological time series. Zeitschrift fuer Meteorologie, 39(4), 224–226.
Cebrian, A., and Denuit, M. and Lambert, P. (2003). Generalized pareto fit to the society of actuaries large claims database. North American Actuarial Journal, 7(3), 18–36.
data(danish) ggplot(danish)data(danish) ggplot(danish)
An Implementation of the procedure proposed in Guillou & Hall(2001) for selecting the optimal threshold in extreme value analysis.
GH(data)GH(data)
data |
vector of sample data |
The procedure proposed in Guillou & Hall (2001) is based on bias reduction. Due to the fact that the log-spacings of the order statistics are approximately exponentially distributed if the tail of the underlying distribution follows a Pareto distribution, an auxilliary statistic with respect to k is implemented with the same properties. The method then behaves like an asymptotic test for mean 0. If some critical value crit is exceeded the hypothesis of zero mean is rejected. Thus the bias has become too large and the assumed exponentiality and therefore the assumed Pareto tail can not be hold. From this an optimal number of k can be found such that the critical value is not exceeded. This optimal number, denoted k0 here, is equivalent to the number of extreme values or, if you wish, the number of exceedances in the context of a POT-model like the generalized Pareto distribution. k0 can then be associated with the unknown threshold u of the GPD by
coosing u as the n-k0th upper order statistic. For more information see references.
k0 |
optimal number of upper order statistics, i.e. number of exceedances or data in the tail |
threshold |
the corresponding threshold |
tail.index |
the corresponding tail index |
Guillou, A. and Hall, P. (2001). A Diagnostic for Selecting the Threshold in Extreme Value Analysis. Journal of the Royal Statistical Society, 63(2), 293–305.
data(danish) GH(danish)data(danish) GH(danish)
An Implementation of the procedure proposed in Gomes et al. (2012) and Caeiro et al. (2016) for selecting the optimal sample fraction in tail index estimation.
gomes(data, B = 1000, epsilon = 0.995)gomes(data, B = 1000, epsilon = 0.995)
data |
vector of sample data |
B |
number of Bootstrap replications |
epsilon |
gives the amount of the first resampling size |
The Double Bootstrap procedure simulates the AMSE criterion of the Hill estimator using an auxiliary statistic. Minimizing this statistic gives a consistent estimator of the sample fraction k/n with k the optimal number of upper order statistics. This number, denoted k0 here, is equivalent to the number of extreme values or, if you wish, the number of exceedances in the context of a POT-model like the generalized Pareto distribution. k0 can then be associated with the unknown threshold u of the GPD by choosing u as the n-k0th upper order statistic. For more information see references.
second.order.par |
gives an estimation of the second order parameter |
k0 |
optimal number of upper order statistics, i.e. number of exceedances or data in the tail |
threshold |
the corresponding threshold |
tail.index |
the corresponding tail |
Gomes, M.I. and Figueiredo, F. and Neves, M.M. (2012). Adaptive estimation of heavy right tails: resampling-based methods in action. Extremes, 15, 463–489.
Caeiro, F. and Gomes, I. (2016). Threshold selection in extreme value analysis. Extreme Value Modeling and Risk Analysis: Methods and Applications, 69–86.
data(danish) gomes(danish)data(danish) gomes(danish)
Density, distribution function, quantile function and random number generation for the Generalized Pareto distribution with location, scale, and shape parameters.
dgpd(x, loc = 0, scale = 1, shape = 0, log.d = FALSE) rgpd(n, loc = 0, scale = 1, shape = 0) qgpd(p, loc = 0, scale = 1, shape = 0, lower.tail = TRUE, log.p = FALSE) pgpd(q, loc = 0, scale = 1, shape = 0, lower.tail = TRUE, log.p = FALSE)dgpd(x, loc = 0, scale = 1, shape = 0, log.d = FALSE) rgpd(n, loc = 0, scale = 1, shape = 0) qgpd(p, loc = 0, scale = 1, shape = 0, lower.tail = TRUE, log.p = FALSE) pgpd(q, loc = 0, scale = 1, shape = 0, lower.tail = TRUE, log.p = FALSE)
x |
Vector of observations. |
loc, scale, shape
|
Location, scale, and shape parameters. Can be vectors, but the lengths must be appropriate. |
log.d |
Logical; if TRUE, the log density is returned. |
n |
Number of observations. |
p |
Vector of probabilities. |
lower.tail |
Logical; if TRUE (default), probabilities are P[X <= x], otherwise, P[X > x]. |
log.p |
Logical; if TRUE, probabilities p are given as log(p). |
q |
Vector of quantiles. |
The Generalized Pareto distribution function is given (Pickands, 1975) by
defined
on , with location ,
scale , and shape parameter .
Brian Bader, Jun Yan. "eva: Extreme Value Analysis with Goodness-of-Fit Testing." R package version (2016)
Pickands III, J. (1975). Statistical inference using extreme order statistics. Annals of Statistics, 119-131.
dgpd(2:4, 1, 0.5, 0.01) dgpd(2, -2:1, 0.5, 0.01) pgpd(2:4, 1, 0.5, 0.01) qgpd(seq(0.9, 0.6, -0.1), 2, 0.5, 0.01) rgpd(6, 1, 0.5, 0.01) ## Generate sample with linear trend in location parameter rgpd(6, 1:6, 0.5, 0.01) ## Generate sample with linear trend in location and scale parameter rgpd(6, 1:6, seq(0.5, 3, 0.5), 0.01) p <- (1:9)/10 pgpd(qgpd(p, 1, 2, 0.8), 1, 2, 0.8) ## [1] 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 ## Incorrect syntax (parameter vectors are of different lengths other than 1) # rgpd(1, 1:8, 1:5, 0) ## Also incorrect syntax # rgpd(10, 1:8, 1, 0.01)dgpd(2:4, 1, 0.5, 0.01) dgpd(2, -2:1, 0.5, 0.01) pgpd(2:4, 1, 0.5, 0.01) qgpd(seq(0.9, 0.6, -0.1), 2, 0.5, 0.01) rgpd(6, 1, 0.5, 0.01) ## Generate sample with linear trend in location parameter rgpd(6, 1:6, 0.5, 0.01) ## Generate sample with linear trend in location and scale parameter rgpd(6, 1:6, seq(0.5, 3, 0.5), 0.01) p <- (1:9)/10 pgpd(qgpd(p, 1, 2, 0.8), 1, 2, 0.8) ## [1] 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 ## Incorrect syntax (parameter vectors are of different lengths other than 1) # rgpd(1, 1:8, 1:5, 0) ## Also incorrect syntax # rgpd(10, 1:8, 1, 0.01)
Fits exceedances above a chosen threshold to the Generalized Pareto model. Various estimation procedures can be used, including maximum likelihood, probability weighted moments, and maximum product spacing. It also allows generalized linear modeling of the parameters.
gpdFit(data, threshold = NA, nextremes = NA, npp = 365, method = c("mle", "mps", "pwm"), information = c("expected", "observed"), scalevars = NULL, shapevars = NULL, scaleform = ~1, shapeform = ~1, scalelink = identity, shapelink = identity, start = NULL, opt = "Nelder-Mead", maxit = 10000, ...)gpdFit(data, threshold = NA, nextremes = NA, npp = 365, method = c("mle", "mps", "pwm"), information = c("expected", "observed"), scalevars = NULL, shapevars = NULL, scaleform = ~1, shapeform = ~1, scalelink = identity, shapelink = identity, start = NULL, opt = "Nelder-Mead", maxit = 10000, ...)
data |
Data should be a numeric vector from the GPD. |
threshold |
A threshold value or vector of the same length as the data. |
nextremes |
Number of upper extremes to be used (either this or the threshold must be given, but not both). |
npp |
Length of each period (typically year). Is used in return level estimation. Defaults to 365. |
method |
Method of estimation - maximum likelihood (mle), maximum product spacing (mps), and probability weighted moments (pwm). Uses mle by default. For pwm, only the stationary model can be fit. |
information |
Whether standard errors should be calculated via observed or expected (default) information. For probability weighted moments, only expected information will be used if possible. For non-stationary models, only observed information is used. |
scalevars, shapevars
|
A dataframe of covariates to use for modeling of the each parameter. Parameter intercepts are automatically handled by the function. Defaults to NULL for the stationary model. |
scaleform, shapeform
|
An object of class ‘formula’ (or one that can be coerced into that class), specifying the model of each parameter. By default, assumes stationary (intercept only) model. See details. |
scalelink, shapelink
|
A link function specifying the relationship between the covariates and each parameter. Defaults to the identity function. For the stationary model, only the identity link should be used. |
start |
Option to provide a set of starting parameters to optim; a vector of scale and shape, in that order. Otherwise, the routine attempts to find good starting parameters. See details. |
opt |
Optimization method to use with optim. |
maxit |
Number of iterations to use in optimization, passed to optim. Defaults to 10,000. |
... |
Additional arguments to pass to optim. |
The base code for finding probability weighted moments is taken from the R package evir. See citation.
In the stationary case (no covariates), starting parameters for mle and mps estimation are the probability weighted moment estimates.
In the case where covariates are used, the starting intercept parameters are the probability weighted moment estimates from the
stationary case and the parameters based on covariates are initially set to zero. For non-stationary parameters, the
first reported estimate refers to the intercept term. Covariates are centered and scaled automatically to speed up optimization,
and then transformed back to original scale.
Formulas for generalized linear modeling of the parameters should be given in the form '~ var1 + var2 + '. Essentially,
specification here is the same as would be if using function ‘lm’ for only the right hand side of the equation. Interactions,
polynomials, etc. can be handled as in the ‘formula’ class.
Intercept terms are automatically handled by the function. By default, the link functions are the identity function and
the covariate dependent scale parameter estimates are forced to be positive. For some link function and for
example, scale parameter , the link is written as .
Maximum likelihood estimation and maximum product spacing estimation can be used in all cases. Probability weighted moments
can only be used for stationary models.
A class object ‘gpdFit’ describing the fit, including parameter estimates and standard errors.
Brian Bader, Jun Yan. "eva: Extreme Value Analysis with Goodness-of-Fit Testing." R package version (2016)
## Fit data using the three different estimation procedures set.seed(7) x <- rgpd(2000, loc = 0, scale = 2, shape = 0.2) ## Set threshold at 4 mle_fit <- gpdFit(x, threshold = 4, method = "mle") pwm_fit <- gpdFit(x, threshold = 4, method = "pwm") mps_fit <- gpdFit(x, threshold = 4, method = "mps") ## Look at the difference in parameter estimates and errors mle_fit$par.ests pwm_fit$par.ests mps_fit$par.ests mle_fit$par.ses pwm_fit$par.ses mps_fit$par.ses ## A linear trend in the scale parameter set.seed(7) n <- 300 x2 <- rgpd(n, loc = 0, scale = 1 + 1:n / 200, shape = 0) covs <- as.data.frame(seq(1, n, 1)) names(covs) <- c("Trend1") result1 <- gpdFit(x2, threshold = 0, scalevars = covs, scaleform = ~ Trend1) ## Show summary of estimates result1## Fit data using the three different estimation procedures set.seed(7) x <- rgpd(2000, loc = 0, scale = 2, shape = 0.2) ## Set threshold at 4 mle_fit <- gpdFit(x, threshold = 4, method = "mle") pwm_fit <- gpdFit(x, threshold = 4, method = "pwm") mps_fit <- gpdFit(x, threshold = 4, method = "mps") ## Look at the difference in parameter estimates and errors mle_fit$par.ests pwm_fit$par.ests mps_fit$par.ests mle_fit$par.ses pwm_fit$par.ses mps_fit$par.ses ## A linear trend in the scale parameter set.seed(7) n <- 300 x2 <- rgpd(n, loc = 0, scale = 1 + 1:n / 200, shape = 0) covs <- as.data.frame(seq(1, n, 1)) names(covs) <- c("Trend1") result1 <- gpdFit(x2, threshold = 0, scalevars = covs, scaleform = ~ Trend1) ## Show summary of estimates result1
An Implementation of the procedure proposed in Hall (1990) for selecting the optimal sample fraction in tail index estimation
hall(data, B = 1000, epsilon = 0.955, kaux = 2 * sqrt(length(data)))hall(data, B = 1000, epsilon = 0.955, kaux = 2 * sqrt(length(data)))
data |
vector of sample data |
B |
number of Bootstrap replications |
epsilon |
gives the amount of the first resampling size |
kaux |
tuning parameter for the hill estimator |
The Bootstrap procedure simulates the AMSE criterion of the Hill estimator. The unknown theoretical parameter of the inverse tail index gamma is replaced by a consistent estimation using a tuning parameter kaux for the Hill estimator. Minimizing this statistic gives a consistent estimator of the sample fraction k/n with k the optimal number of upper order statistics. This number, denoted k0 here, is equivalent to the number of extreme values or, if you wish, the number of exceedances in the context of a POT-model like the generalized Pareto distribution. k0 can then be associated with the unknown threshold u of the GPD by choosing u as the n-k0th upper order statistic. For more information see references.
k0 |
optimal number of upper order statistics, i.e. number of exceedances or data in the tail |
threshold |
the corresponding threshold |
tail.index |
the corresponding tail index |
Hall, P. (1990). Using the Bootstrap to Estimate Mean Squared Error and Select Smoothing Parameter in Nonparametric Problems. Journal of Multivariate Analysis, 32, 177–203.
data(danish) hall(danish)data(danish) hall(danish)
An Implementation of the procedure proposed in Caeiro & Gomes (2012) for selecting the optimal sample fraction in tail index estimation
Himp(data, B = 1000, epsilon = 0.955)Himp(data, B = 1000, epsilon = 0.955)
data |
vector of sample data |
B |
number of Bootstrap replications |
epsilon |
gives the amount of the first resampling size |
This procedure is an improvement of the one introduced in Hall (1990) by overcoming the restrictive assumptions through estimation of the necessary parameters. The Bootstrap procedure simulates the AMSE criterion of the Hill estimator using an auxiliary statistic. Minimizing this statistic gives a consistent estimator of the sample fraction k/n with k the optimal number of upper order statistics. This number, denoted k0 here, is equivalent to the number of extreme values or, if you wish, the number of exceedances in the context of a POT-model like the generalized Pareto distribution. k0 can then be associated with the unknown threshold u of the GPD by choosing u as the n-k0th upper order statistic. For more information see references.
second.order.par |
gives an estimation of the second order parameter |
k0 |
optimal number of upper order statistics, i.e. number of exceedances or data in the tail |
threshold |
the corresponding threshold |
tail.index |
the corresponding tail index |
Hall, P. (1990). Using the Bootstrap to Estimate Mean Squared Error and Select Smoothing Parameter in Nonparametric Problems. Journal of Multivariate Analysis, 32, 177–203.
Caeiro, F. and Gomes, M.I. (2014). On the bootstrap methodology for the estimation of the tail sample fraction. Proceedings of COMPSTAT, 545–552.
data(danish) Himp(danish)data(danish) Himp(danish)
An Implementation of the procedure proposed in Hall & Welsh (1985) for obtaining the optimal number of upper order statistics k for the Hill estimator by minimizing the AMSE-criterion.
HW(data)HW(data)
data |
vector of sample data |
The optimal number of upper order statistics is equivalent to the number of extreme values or, if you wish, the number of exceedances in the context of a POT-model like the generalized Pareto distribution. This number is identified by minimizing the AMSE criterion with respect to k. The optimal number, denoted k0 here, can then be associated with the unknown threshold u of the GPD by choosing u as the n-k0th upper order statistic. For more information see references.
second.order.par |
gives an estimation of the second order parameter |
k0 |
optimal number of upper order statistics, i.e. number of exceedances or data in the tail |
threshold |
the corresponding threshold |
tail.index |
the corresponding tail index |
Hall, P. and Welsh, A.H. (1985). Adaptive estimates of parameters of regular variation. The Annals of Statistics, 13(1), 331–341.
data(danish) HW(danish)data(danish) HW(danish)
An Implementation of the procedure proposed in Danielsson et al. (2016) for selecting the optimal threshold in extreme value analysis.
mindist(data, ts = 0.15, method = "mad")mindist(data, ts = 0.15, method = "mad")
data |
vector of sample data |
ts |
size of the upper tail the procedure is applied to. Default is 15 percent of the data |
method |
should be one of |
The procedure proposed in Danielsson et al. (2016) minimizes the distance between the largest upper order statistics of the dataset, i.e. the empirical tail, and the theoretical tail of a Pareto distribution. The parameter of this distribution are estimated using Hill's estimator. Therefor one needs the optimal number of upper order statistics k. The distance is then minimized with respect to this k. The optimal number, denoted k0 here, is equivalent to the number of extreme values or, if you wish, the number of exceedances in the context of a POT-model like the generalized Pareto distribution. k0 can then be associated with the unknown threshold u of the GPD by saying u is the n-k0th upper order statistic. For the distance metric in use one could choose the mean absolute deviation called mad here, or the maximum absolute deviation, also known as the "Kolmogorov-Smirnov" distance metric (ks). For more information see references.
k0 |
optimal number of upper order statistics, i.e. number of exceedances or data in the tail |
threshold |
the corresponding threshold |
tail.index |
the corresponding tail index by plugging in |
Danielsson, J. and Ergun, L.M. and de Haan, L. and de Vries, C.G. (2016). Tail Index Estimation: Quantile Driven Threshold Selection.
data(danish) mindist(danish,method="mad")data(danish) mindist(danish,method="mad")
An Implementation of the heuristic algorithm for choosing the optimal sample fraction proposed in Caeiro & Gomes (2016), among others.
PS(data, j = 1)PS(data, j = 1)
data |
vector of sample data |
j |
digits to round to. Should be |
The algorithm searches for a stable region of the sample path, i.e. the plot of a tail index estimator with respect to k. This is done in two steps. First the estimation of the tail index for every k is rounded to j digits and the longest set of equal consecutive values is chosen. For this set the estimates are rounded to j+2 digits and the mode of this subset is determined. The corresponding biggest k-value, denoted k0 here, is the optimal number of data in the tail.
k0 |
optimal number of upper order statistics, i.e. number of exceedances or data in the tail |
threshold |
the corresponding threshold |
tail.index |
the corresponding tail index |
Caeiro, J. and Gomes, M.I. (2016). Threshold selection in extreme value analysis. Extreme Value Modeling and Risk Analysis:Methids and Applications, 69–86.
Gomes, M.I. and Henriques-Rodrigues, L. and Fraga Alves, M.I. and Manjunath, B. (2013). Adaptive PORT-MVRB estimation: an empirical comparison of two heuristic algorithms. Journal of Statistical Computation and Simulation, 83, 1129–1144.
Gomes, M.I. and Henriques-Rodrigues, L. and Miranda, M.C. (2011). Reduced-bias location-invariant extreme value index estimation: a simulation study. Communications in Statistic-Simulation and Computation, 40, 424–447.
data(danish) PS(danish)data(danish) PS(danish)
Plots the QQ-Estimator against the upper order statistics
qqestplot(data, kmin = 5, conf.int = FALSE)qqestplot(data, kmin = 5, conf.int = FALSE)
data |
vector of sample data |
kmin |
gives the minimal |
conf.int |
|
The QQ-Estimator is a Tail Index Estimator based on regression diagnostics. Assuming a Pareto tail behaviour of the data at hand a QQ-Plot of the theoretical quantiles of an exponential distribution against the empirical quantiles of the log-data should lead to a straight line above some unknown upper order statistic k. The slope of this line is an estimator for the tail index. Computing this estimator via linear regression for every k the plot should stabilize for the correct number of upper order statistics, denoted k0 here.
The plot shows the values of the QQ-Estimator with respect to k. See references for more information.
Kratz, M. and Resnick, S.I. (1996). The QQ-estimator and heavy tails. Stochastic Models, 12(4), 699–724.
data(danish) qqestplot(danish)data(danish) qqestplot(danish)
Plots the empirical observations above a given threshold against the theoretical quantiles of a generalized Pareto distribution.
qqgpd(data, nextremes, scale, shape)qqgpd(data, nextremes, scale, shape)
data |
vector of sample data |
nextremes |
number of exceedances |
scale |
scale parameter of GPD |
shape |
shape parameter of GPD |
If the fitted GPD model provides a reasonable approximation of the underlying sample data the empirical and theoretical quantiles should coincide. So plotting them against each other should result in a straight line. Deviations from that line speak for a bad model fit and against a GPD assumption.
The straight red line gives the line of agreement. The dashed lines are simulated 95 percent confidence intervals. Therefor the fitted GPD model is simulated 1000 times using Monte Carlo. The sample size of each simulation equals the number of exceedances.
data=rexp(1000) #GPD with scale=1, shape=0 qqgpd(data,1000,1,0)data=rexp(1000) #GPD with scale=1, shape=0 qqgpd(data,1000,1,0)
An implementation of the minimization criterion proposed in Reiss & Thomas (2007).
RT(data, beta = 0, kmin = 2)RT(data, beta = 0, kmin = 2)
data |
vector of sample data |
beta |
a factor for weighting the expression below. Default is set to |
kmin |
gives a minimum value for |
The procedure proposed in Reiss & Thomas (2007) chooses the lowest upper order statistic k to minimize the expression
1/k sum_i=1^k i^beta |gamma_i-median(gamma_1,...,gamma_k)|
or an alternative of that by replacing the absolute deviation with a squared deviation and the median just with gamma_k, where gamma denotes the Hill estimator
k0 |
optimal number of upper order statistics, i.e. number of exceedances or data in the tail for both metrics, i.e. the absolute and squared deviation. |
threshold |
the corresponding thresholds. |
tail.index |
the corresponding tail indices |
Reiss, R.-D. and Thomas, M. (2007). Statistical Analysis of Extreme Values: With Applications to Insurance, Finance, Hydrology and Other Fields. Birkhauser, Boston.
data(danish) RT(danish)data(danish) RT(danish)
An implementation of the so called sum plot proposed in de Sousa & Michailidis (2004)
sumplot(data, kmin = 5)sumplot(data, kmin = 5)
data |
vector of sample data |
kmin |
gives the minimal |
The sum plot is based on the plot (k,S_k) with S_k:=k*gamma_k where gamma_k denotes the Hill estimator. So the sum plot and the Hill plot are statistically equivalent. The sum plot should be approximately linear for the k-values where gamma_k=gamma. So the linear part of the graph can be used as an estimator of the (inverse) tail index. The sum plot leads to the estimation of the slope while the classical Hill plot leads to estimation of the intercept. The optimal number of order statistics, also known as the threshold, can then be derived as the value k where the plot differs from a straight line with slope gamma. See references for more information.
The plot shows the values of S_k=k*gamma_k for different k. See references for more information.
De Sousa, Bruno and Michailidis, George (2004). A diagnostic plot for estimating the tail index of a distribution. Journal of Computational and Graphical Statistics 13(4), 1–22.
data(danish) sumplot(danish)data(danish) sumplot(danish)
An implementation of the sequential testing procedure proposed in Thompson et al. (2009) for automated threshold selection
TH(data, thresholds)TH(data, thresholds)
data |
vector of sample data |
thresholds |
a sequence of pre-defined thresholds to check for GPD assumption |
The procedure proposed in Thompson et al. (2009) is based on sequential goodness of fit testing. First, one has to choose a equally spaced grid of posssible thresholds. The authors recommend 100 thresholds between the 50 percent and 98 percent quantile of the data, provided there are enough observations left (about 100 observations above the last pre-defined threshold). Then the parameters of a GPD for each threshold are estimated. One can show that the differences of subsequent scale parameters are approximately normal distributed. So a Pearson chi-squared test for normality is applied to all the differences, striking the smallest thresholds out until the test is not rejected anymore.
threshold |
the threshold used for the test |
num.above |
the number of observations above the given threshold |
p.values |
raw p-values for the thresholds tested |
ForwardStop |
transformed p-values according to the ForwardStop criterion. See G'Sell et al (2016) for more information |
StrongStop |
transformed p-values according to the StrongStop criterion. See G'Sell et al (2016) for more information |
est.scale |
estimated scale parameter for the given threshold |
est.shape |
estimated shape parameter for the given threshold |
Thompson, P. and Cai, Y. and Reeve, D. (2009). Automated threshold selection methods for extreme wave analysis. Coastal Engineering, 56(10), 1013–1021.
G'Sell, M.G. and Wager, S. and Chouldechova, A. and Tibshirani, R. (2016). Sequential selection procedures and false discovery rate control. Journal of the Royal Statistical Society: Series B (Statistical Methodology) 78(2), 423–444.
data=rexp(1000) u=seq(quantile(data,.1),quantile(data,.9),,100) A=TH(data,u);Adata=rexp(1000) u=seq(quantile(data,.1),quantile(data,.9),,100) A=TH(data,u);A