Package fExtremes

February 19, 2015

Version 3010.81
Revision 5532
Date 2013-12-17
Title Rmetrics - Extreme Financial Market Data
Author Diethelm Wuertz and many others, see the SOURCE file
Depends R (>= 2.4.0), timeDate, timeSeries, fBasics, fGarch, fTrading
Imports methods
Suggests RUnit, tcltk
Maintainer Yohan Chalabi <yohan.chalabi@rmetrics.org>
Description Environment for teaching ``Financial Engineering and
Computational Finance''
Note Several parts are still preliminary and may be changed in the
future. this typically includes function and argument names, as
well as defaults for arguments and return values.
LazyData yes
License GPL (>= 2)

URL http://www.rmetrics.org
NeedsCompilation no
Repository CRAN
Date/Publication 2013-12-17 23:16:44

R topics documented:
DataPreprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
ExtremeIndex . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
ExtremesData . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
GevDistribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
GevMdaEstimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
GevModelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
GevRisk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

1
2 DataPreprocessing

GpdDistribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
GpdModelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
gpdRisk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
TimeSeriesData . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
ValueAtRisk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

Index 36

DataPreprocessing Extremes Data Preprocessing

Description
A collection and description of functions for data preprocessing of extreme values. This includes
tools to separate data beyond a threshold value, to compute blockwise data like block maxima, and
to decluster point process data.

The functions are:

blockMaxima Block Maxima from a vector or a time series,

findThreshold Upper threshold for a given number of extremes,
pointProcess Peaks over Threshold from a vector or a time series,
deCluster Declusters clustered point process data.

Usage
blockMaxima(x, block = c("monthly", "quarterly"), doplot = FALSE)
findThreshold(x, n = floor(0.05*length(as.vector(x))), doplot = FALSE)
pointProcess(x, u = quantile(x, 0.95), doplot = FALSE)
deCluster(x, run = 20, doplot = TRUE)

Arguments
block the block size. A numeric value is interpreted as the number of data values in
each successive block. All the data is used, so the last block may not contain
block observations. If the data has a times attribute containing (in an ob-
ject of class "POSIXct", or an object that can be converted to that class, see
as.POSIXct) the times/dates of each observation, then block may instead take
the character values "month", "quarter", "semester" or "year". By default
monthly blocks from daily data are assumed.
doplot a logical value. Should the results be plotted? By default TRUE.
n a numeric value or vector giving number of extremes above the threshold. By
default, n is set to an integer representing 5% of the data from the whole data set
x.
run parameter to be used in the runs method; any two consecutive threshold ex-
ceedances separated by more than this number of observations/days are consid-
ered to belong to different clusters.
DataPreprocessing 3

u a numeric value at which level the data are to be truncated. By default the
threshold value which belongs to the 95% quantile, u=quantile(x,0.95).
x a numeric data vector from which findThreshold and blockMaxima determine
the threshold values and block maxima values. For the function deCluster
the argument x represents a numeric vector of threshold exceedances with a
times attribute which should be a numeric vector containing either the indices
or the times/dates of each exceedance (if times/dates, the attribute should be an
object of class "POSIXct" or an object that can be converted to that class; see
as.POSIXct).

Details

Computing Block Maxima:

The function blockMaxima calculates block maxima from a vector or a time series, whereas the
function blocks is more general and allows for the calculation of an arbitrary function FUN on
blocks.

Finding Thresholds:

The function findThreshold finds a threshold so that a given number of extremes lie above. When
the data are tied a threshold is found so that at least the specified number of extremes lie above.

De-Clustering Point Processes:

The function deCluster declusters clustered point process data so that Poisson assumption is more
tenable over a high threshold.

Value

blockMaxima
returns a timeSeries object or a numeric vector of block maxima data.
findThreshold
returns a numeric value or vector of suitable thresholds.
pointProcess
returns a timeSeries object or a numeric vector of peaks over a threshold.
deCluster
returns a timeSeries object or a numeric vector for the declustered point process.

Author(s)

Some of the functions were implemented from Alec Stephensons R-package evir ported from
Alexander McNeils S library EVIS, Extreme Values in S, some from Alec Stephensons R-package
ismev based on Stuart Coles code from his book, Introduction to Statistical Modeling of Extreme
Values and some were written by Diethelm Wuertz.
4 ExtremeIndex

References
Coles S. (2001); Introduction to Statistical Modelling of Extreme Values, Springer.
Embrechts, P., Klueppelberg, C., Mikosch, T. (1997); Modelling Extremal Events, Springer.

Examples

## findThreshold -
# Threshold giving (at least) fifty exceedances for Danish data:
x = as.timeSeries(data(danishClaims))
findThreshold(x, n = c(10, 50, 100))

## blockMaxima -
# Block Maxima (Minima) for left tail of BMW log returns:
BMW = as.timeSeries(data(bmwRet))
colnames(BMW) = "BMW.RET"
head(BMW)
x = blockMaxima( BMW, block = 65)
head(x)
y = blockMaxima(-BMW, block = 65)
head(y)
y = blockMaxima(-BMW, block = "monthly")
head(y)

## pointProcess -
# Return Values above threshold in negative BMW log-return data:
PP = pointProcess(x = -BMW, u = quantile(as.vector(x), 0.75))
PP
nrow(PP)

## deCluster -
# Decluster the 200 exceedances of a particular
DC = deCluster(x = PP, run = 15, doplot = TRUE)
DC
nrow(DC)

ExtremeIndex Extremal Index Estimation

Description
A collection and description of functions to simulate time series with a known extremal index, and
to estimate the extremal index by four different kind of methods, the blocks method, the reciprocal
mean cluster size method, the runs method, and the method of Ferro and Segers.

The functiona are:

thetaSim Simulates a time Series with known theta,

ExtremeIndex 5

blockTheta Computes theta from Block Method,

clusterTheta Computes theta from Reciprocal Cluster Method,
runTheta Computes theta from Run Method,
ferrosegersTheta Computes Theta according to Ferro and Seegers,
exindexPlot Calculate and Plot Theta(1,2,3),
exindexesPlot Calculate Theta(1,2) and Plot Theta(1).

Usage
## S4 method for signature 'fTHETA'
show(object)

thetaSim(model = c("max", "pair"), n = 1000, theta = 0.5)

blockTheta(x, block = 22, quantiles = seq(0.950, 0.995, length = 10),

title = NULL, description = NULL)
clusterTheta(x, block = 22, quantiles = seq(0.950, 0.995, length = 10),
title = NULL, description = NULL)
runTheta(x, block = 22, quantiles = seq(0.950, 0.995, length = 10),
title = NULL, description = NULL)
ferrosegersTheta(x, quantiles = seq(0.950, 0.995, length = 10),
title = NULL, description = NULL)

exindexPlot(x, block = c("monthly", "quarterly"), start = 5, end = NA,

doplot = TRUE, plottype = c("thresh", "K"), labels = TRUE, ...)

exindexesPlot(x, block = 22, quantiles = seq(0.950, 0.995, length = 10),

doplot = TRUE, labels = TRUE, ...)

Arguments
block [*Theta] -
an integer value, the block size. Currently only integer specified block sizes are
supported.
[exindex*Plot] -
the block size. Either "monthly", "quarterly" or an integer value. An integer
value is interpreted as the number of data values in each successive block. The
default value is "monthly" which correpsond for daily data to an approximately
22-day periods.
description a character string which allows for a brief description.
doplot a logical, should the results be plotted?
labels whether or not axes should be labelled. If set to FALSE then user specified lables
can be passed through the "..." argument.
model [thetaSim] -
a character string denoting the name of the model. Either "max" or "pair",
the first representing the maximimum Frechet series, and the second the paired
exponential series.
6 ExtremeIndex

n [thetaSim] -
an integer value, the length of the time series to be generated.
object an object of class "fTHETA" as returned by the functions *Theta.
plottype [exindexPlot] -
whether plot is to be by increasing threshold (thresh) or increasing K value (K).
quantiles [exindexesPlot] -
a numeric vector of quantile values.
start, end [exindexPlot] -
start is the lowest value of K at which to plot a point, and end the highest value;
K is the number of blocks in which a specified threshold is exceeded.
theta [thetaSim] -
a numeric value between 0 and 1 setting the value of the extremal index for the
maximum Frechet time series. (Not used in the case of the paired exponential
series.)
title a character string which allows for a project title.
x a timeSeries object or any other object which can be transformed by the func-
tion as.vector into a numeric vector. "monthly" and "quarterly" blocks
require x to be an object of class "timeSeries".
... additional arguments passed to the plot function.

Value

exindexPlot
returns a data frame of results with the following columns: N, K, un, theta2, and theta. A plot
with K on the lower x-axis and threshold Values on the upper x-axis versus the extremal index is
displayed.
exindexesPlot
returns a data.frame with four columns: thresholds, theta1, theta2, and theta3. A plot with
quantiles on the x-axis and versus the extremal indexes is displayed.

Author(s)

Alexander McNeil, for parts of the exindexPlot function, and

Diethelm Wuertz for the exindexesPlot function.

References

Embrechts, P., Klueppelberg, C., Mikosch, T. (1997); Modelling Extremal Events, Springer. Chap-
ter 8, 413429.

See Also

hillPlot, gevFit.
ExtremesData 7

Examples
## Extremal Index for the right and left tails
## of the BMW log returns:
data(bmwRet)
par(mfrow = c(2, 2), cex = 0.7)
exindexPlot( as.timeSeries(bmwRet), block = "quarterly")
exindexPlot(-as.timeSeries(bmwRet), block = "quarterly")

## Extremal Index for the right and left tails

## of the BMW log returns:
exindexesPlot( as.timeSeries(bmwRet), block = 65)
exindexesPlot(-as.timeSeries(bmwRet), block = 65)

ExtremesData Explorative Data Analysis

Description
A collection and description of functions for explorative data analysis. The tools include plot func-
tions for emprical distributions, quantile plots, graphs exploring the properties of exceedences over
a threshold, plots for mean/sum ratio and for the development of records.

The functions are:

emdPlot Plot of empirical distribution function,

qqparetoPlot Exponential/Pareto quantile plot,
mePlot Plot of mean excesses over a threshold,
mrlPlot another variant, mean residual life plot,
mxfPlot another variant, with confidence intervals,
msratioPlot Plot of the ratio of maximum and sum,
recordsPlot Record development compared with iid data,
ssrecordsPlot another variant, investigates subsamples,
sllnPlot verifies Kolmogorovs strong law of large numbers,
lilPlot verifies Hartman-Wintners law of the iterated logarithm,
xacfPlot ACF of exceedences over a threshold,
normMeanExcessFit fits mean excesses with a normal density,
ghMeanExcessFit fits mean excesses with a GH density,
hypMeanExcessFit fits mean excesses with a HYP density,
nigMeanExcessFit fits mean excesses with a NIG density,
ghtMeanExcessFit fits mean excesses with a GHT density.

Usage
emdPlot(x, doplot = TRUE, plottype = c("xy", "x", "y", " "),
labels = TRUE, ...)

qqparetoPlot(x, xi = 0, trim = NULL, threshold = NULL, doplot = TRUE,

8 ExtremesData

labels = TRUE, ...)

mePlot(x, doplot = TRUE, labels = TRUE, ...)

mrlPlot(x, ci = 0.95, umin = mean(x), umax = max(x), nint = 100, doplot = TRUE,
plottype = c("autoscale", ""), labels = TRUE, ...)
mxfPlot(x, u = quantile(x, 0.05), doplot = TRUE, labels = TRUE, ...)

msratioPlot(x, p = 1:4, doplot = TRUE, labels = TRUE, ...)

recordsPlot(x, ci = 0.95, doplot = TRUE, labels = TRUE, ...)

ssrecordsPlot(x, subsamples = 10, doplot = TRUE, plottype = c("lin", "log"),
labels = TRUE, ...)

sllnPlot(x, doplot = TRUE, labels = TRUE, ...)

lilPlot(x, doplot = TRUE, labels = TRUE, ...)

xacfPlot(x, u = quantile(x, 0.95), lag.max = 15, doplot = TRUE,

which = c("all", 1, 2, 3, 4), labels = TRUE, ...)

normMeanExcessFit(x, doplot = TRUE, trace = TRUE, ...)

ghMeanExcessFit(x, doplot = TRUE, trace = TRUE, ...)
hypMeanExcessFit(x, doplot = TRUE, trace = TRUE, ...)
nigMeanExcessFit(x, doplot = TRUE, trace = TRUE, ...)
ghtMeanExcessFit(x, doplot = TRUE, trace = TRUE, ...)

Arguments
ci [recordsPlot] -
a confidence level. By default 0.95, i.e. 95%.
doplot a logical value. Should the results be plotted? By default TRUE.
labels a logical value. Whether or not x- and y-axes should be automatically labelled
and a default main title should be added to the plot. By default TRUE.
lag.max [xacfPlot] -
maximum number of lags at which to calculate the autocorrelation functions.
The default value is 15.
nint [mrlPlot] -
the number of intervals, see umin and umax. The default value is 100.
p [msratioPlot] -
the power exponents, a numeric vector. By default a sequence from 1 to 4 in
unit integer steps.
plottype [emdPlot] -
which axes should be on a log scale: "x" x-axis only; "y" y-axis only; "xy"
both axes; "" neither axis.
[msratioPlot] -
a logical, if set to "autoscale", then the scale of the plots are automatically
determined, any other string allows user specified scale information through the
... argument.
ExtremesData 9

[ssrecordsPlot] -
one from two options can be select either "lin" or "log". The default creates a
linear plot.
subsamples [ssrecordsPlot] -
the number of subsamples, by default 10, an integer value.
threshold, trim
[qPlot][xacfPlot] -
a numeric value at which data are to be left-truncated, value at which data are to
be right-truncated or the thresold value, by default 95%.
trace a logical flag, by default TRUE. Should the calculations be traced?
u a numeric value at which level the data are to be truncated. By default the
threshold value which belongs to the 95% quantile, u=quantile(x,0.95).
umin, umax [mrlPlot] -
range of threshold values. If umin and/or umax are not available, then by default
they are set to the following values: umin=mean(x) and umax=max(x).
which [xacfPlot] -
a numeric or character value, if which="all" then all four plots are displayed,
if which is an integer between one and four, then the first, second, third or fourth
plot will be displayed.
x, y numeric data vectors or in the case of x an object to be plotted.
xi the shape parameter of the generalized Pareto distribution.
... additional arguments passed to the FUN or plot function.

Details
Empirical Distribution Function:

The function emdPlot is a simple explanatory function. A straight line on the double log scale
indicates Pareto tail behaviour.

QuantileQuantile Pareto Plot:

qqparetoPlot creates a quantile-quantile plot for threshold data. If xi is zero the reference dis-
tribution is the exponential; if xi is non-zero the reference distribution is the generalized Pareto
with that parameter value expressed by xi. In the case of the exponential, the plot is interpreted
as follows: Concave departures from a straight line are a sign of heavy-tailed behaviour, convex
departures show thin-tailed behaviour.

Mean Excess Function Plot:

Three variants to plot the mean excess function are available: A sample mean excess plot over
increasing thresholds, and two mean excess function plots with confidence intervals for discrimina-
tion in the tails of a distribution. In general, an upward trend in a mean excess function plot shows
heavy-tailed behaviour. In particular, a straight line with positive gradient above some threshold is a
sign of Pareto behaviour in tail. A downward trend shows thin-tailed behaviour whereas a line with
10 ExtremesData

zero gradient shows an exponential tail. Here are some hints: Because upper plotting points are the
average of a handful of extreme excesses, these may be omitted for a prettier plot. For mrlPlot and
mxfPlot the upper tail is investigated; for the lower tail reverse the sign of the data vector.

Plot of the Maximum/Sum Ratio:

The ratio of maximum and sum is a simple tool for detecting heavy tails of a distribution and
for giving a rough estimate of the order of its finite moments. Sharp increases in the curves of a
msratioPlot are a sign for heavy tail behaviour.

Plot of the Development of Records:

These are functions that investigate the development of records in a dataset and calculate the ex-
pected behaviour for iid data. recordsPlot counts records and reports the observations at which
they occur. In addition subsamples can be investigated with the help of the function ssrecordsPlot.

Plot of Kolmogorovs and Hartman-Winterns Laws:

The function sllnPlot verifies Kolmogorovs strong law of large numbers, and the function lilPlot
verifies Hartman-Wintners law of the iterated logarithm.

ACF Plot of Exceedences over a Thresold:

This function plots the autocorrelation functions of heights and distances of exceedences over a
threshold.

Value
The functions return a plot.

Note
The plots are labeled by default with a x-label, a y-label and a main title. If the argument labels
is set to FALSE neither a x-label, a y-label nor a main title will be added to the graph. To add user
defined label strings just use the function title(xlab="...", ylab="...", main="...").

Author(s)
Some of the functions were implemented from Alec Stephensons R-package evir ported from
Alexander McNeils S library EVIS, Extreme Values in S, some from Alec Stephensons R-package
ismev based on Stuart Coles code from his book, Introduction to Statistical Modeling of Extreme
Values and some were written by Diethelm Wuertz.

References
Coles S. (2001); Introduction to Statistical Modelling of Extreme Values, Springer.
Embrechts, P., Klueppelberg, C., Mikosch, T. (1997); Modelling Extremal Events, Springer.
GevDistribution 11

Examples

## Danish fire insurance data:

data(danishClaims)
danishClaims = as.timeSeries(danishClaims)

## emdPlot -
# Show Pareto tail behaviour:
par(mfrow = c(2, 2), cex = 0.7)
emdPlot(danishClaims)

## qqparetoPlot -
# QQ-Plot of heavy-tailed Danish fire insurance data:
qqparetoPlot(danishClaims, xi = 0.7)

## mePlot -
# Sample mean excess plot of heavy-tailed Danish fire:
mePlot(danishClaims)

## ssrecordsPlot -
# Record fire insurance losses in Denmark:
ssrecordsPlot(danishClaims, subsamples = 10)

GevDistribution Generalized Extreme Value Distribution

Description
Density, distribution function, quantile function, random number generation, and true moments for
the GEV including the Frechet, Gumbel, and Weibull distributions.

The GEV distribution functions are:

dgev density of the GEV distribution,

pgev probability function of the GEV distribution,
qgev quantile function of the GEV distribution,
rgev random variates from the GEV distribution,
gevMoments computes true mean and variance,
gevSlider displays density or rvs from a GEV.

Usage
dgev(x, xi = 1, mu = 0, beta = 1, log = FALSE)
pgev(q, xi = 1, mu = 0, beta = 1, lower.tail = TRUE)
qgev(p, xi = 1, mu = 0, beta = 1, lower.tail = TRUE)
rgev(n, xi = 1, mu = 0, beta = 1)

gevMoments(xi = 0, mu = 0, beta = 1)
12 GevDistribution

gevSlider(method = c("dist", "rvs"))

Arguments
log a logical, if TRUE, the log density is returned.
lower.tail a logical, if TRUE, the default, then probabilities are P[X <= x], otherwise,
P[X > x].
method a character sgtring denoting what should be displayed. Either the density and
"dist" or random variates "rvs".
n the number of observations.
p a numeric vector of probabilities. [hillPlot] -
probability required when option quantile is chosen.
q a numeric vector of quantiles.
x a numeric vector of quantiles.
xi, mu, beta xi is the shape parameter, mu the location parameter, and beta is the scale pa-
rameter. The default values are xi=1, mu=0, and beta=1. Note, if xi=0 the
distribution is of type Gumbel.

Value
d* returns the density,
p* returns the probability,
q* returns the quantiles, and
r* generates random variates.

All values are numeric vectors.

Author(s)
Alec Stephenson for Rs evd and evir package, and
Diethelm Wuertz for this R-port.

References
Coles S. (2001); Introduction to Statistical Modelling of Extreme Values, Springer.
Embrechts, P., Klueppelberg, C., Mikosch, T. (1997); Modelling Extremal Events, Springer.

Examples
## rgev -
# Create and plot 1000 Weibull distributed rdv:
r = rgev(n = 1000, xi = -1)
plot(r, type = "l", col = "steelblue", main = "Weibull Series")
grid()

## dgev -
GevMdaEstimation 13

# Plot empirical density and compare with true density:

hist(r[abs(r)<10], nclass = 25, freq = FALSE, xlab = "r",
xlim = c(-5,5), ylim = c(0,1.1), main = "Density")
box()
x = seq(-5, 5, by = 0.01)
lines(x, dgev(x, xi = -1), col = "steelblue")

## pgev -
# Plot df and compare with true df:
plot(sort(r), (1:length(r)/length(r)),
xlim = c(-3, 6), ylim = c(0, 1.1),
cex = 0.5, ylab = "p", xlab = "q", main = "Probability")
grid()
q = seq(-5, 5, by = 0.1)
lines(q, pgev(q, xi = -1), col = "steelblue")

## qgev -
# Compute quantiles, a test:
qgev(pgev(seq(-5, 5, 0.25), xi = -1), xi = -1)

## gevMoments:
# Returns true mean and variance:
gevMoments(xi = 0, mu = 0, beta = 1)

## Slider:
# gevSlider(method = "dist")
# gevSlider(method = "rvs")

GevMdaEstimation Generalized Extreme Value Modelling

Description
A collection and description functions to estimate the parameters of the GEV distribution. To model
the GEV three types of approaches for parameter estimation are provided: Maximum likelihood
estimation, probability weighted moment method, and estimation by the MDA approach. MDA in-
cludes functions for the Pickands, Einmal-Decker-deHaan, and Hill estimators together with several
plot variants.

Maximum Domain of Attraction estimators:

hillPlot shape parameter and Hill estimate of the tail index,

shaparmPlot variation of shape parameter with tail depth.

Usage
hillPlot(x, start = 15, ci = 0.95,
doplot = TRUE, plottype = c("alpha", "xi"), labels = TRUE, ...)
shaparmPlot(x, p = 0.01*(1:10), xiRange = NULL, alphaRange = NULL,
14 GevMdaEstimation

doplot = TRUE, plottype = c("both", "upper"))

shaparmPickands(x, p = 0.05, xiRange = NULL,

doplot = TRUE, plottype = c("both", "upper"), labels = TRUE, ...)
shaparmHill(x, p = 0.05, xiRange = NULL,
doplot = TRUE, plottype = c("both", "upper"), labels = TRUE, ...)
shaparmDEHaan(x, p = 0.05, xiRange = NULL,
doplot = TRUE, plottype = c("both", "upper"), labels = TRUE, ...)

Arguments
alphaRange, xiRange
[saparmPlot] -
plotting ranges for alpha and xi. By default the values are automatically se-
lected.
ci [hillPlot] -
probability for asymptotic confidence band; for no confidence band set ci to
zero.
doplot a logical. Should the results be plotted?
[shaparmPlot] -
a vector of logicals of the same lengths as tails defining for wich tail depths plots
should be created, by default plots will be generated for a tail depth of 5 percent.
By default c(FALSE, FALSE, FALSE, FALSE, TRUE, FALSE, FALSE, FALSE, FALSE, FALSE).
labels [hillPlot] -
whether or not axes should be labelled.
plottype [hillPlot] -
whether alpha, xi (1/alpha) or quantile (a quantile estimate) should be plot-
ted.
p [qgev] -
a numeric vector of probabilities. [hillPlot] -
probability required when option quantile is chosen.
start [hillPlot] -
lowest number of order statistics at which to plot a point.
x [dgev][devd] -
a numeric vector of quantiles.
[gevFit] -
data vector. In the case of method="mle" the interpretation depends on the value
of block: if no block size is specified then data are interpreted as block maxima;
if block size is set, then data are interpreted as raw data and block maxima are
calculated.
[hillPlot][shaparmPlot] -
the data from which to calculate the shape parameter, a numeric vector.
[print][plot] -
a fitted object of class "gevFit".
... [gevFit] -
control parameters optionally passed to the optimization function. Parameters
GevMdaEstimation 15

for the optimization function are passed to components of the control argument
of optim.
[hillPlot] -
other graphics parameters.
[plot][summary] -
arguments passed to the plot function.

Details
Parameter Estimation:

gevFit and gumbelFit estimate the parameters either by the probability weighted moment method,
method="pwm" or by maximum log likelihood estimation method="mle". The summary method
produces diagnostic plots for fitted GEV or Gumbel models.

Methods:

print.gev, plot.gev and summary.gev are print, plot, and summary methods for a fitted object of
class gev. Concerning the summary method, the data are converted to unit exponentially distributed
residuals under null hypothesis that GEV fits. Two diagnostics for iid exponential data are offered.
The plot method provides two different residual plots for assessing the fitted GEV model. Two
diagnostics for iid exponential data are offered.

Return Level Plot:

gevrlevelPlot calculates and plots the k-block return level and 95% confidence interval based
on a GEV model for block maxima, where k is specified by the user. The k-block return level is that
level exceeded once every k blocks, on average. The GEV likelihood is reparameterized in terms
of the unknown return level and profile likelihood arguments are used to construct a confidence
interval.

Hill Plot:

The function hillPlot investigates the shape parameter and plots the Hill estimate of the tail index
of heavy-tailed data, or of an associated quantile estimate. This plot is usually calculated from the
alpha perspective. For a generalized Pareto analysis of heavy-tailed data using the gpdFit function,
it helps to plot the Hill estimates for xi.

Shape Parameter Plot:

The function shaparmPlot investigates the shape parameter and plots for the upper and lower tails
the shape parameter as a function of the taildepth. Three approaches are considered, the Pickands
estimator, the Hill estimator, and the Decker-Einmal-deHaan estimator.

Value
gevSim
returns a vector of data points from the simulated series.
16 GevMdaEstimation

gevFit
returns an object of class gev describing the fit.

print.summary
prints a report of the parameter fit.

summary
performs diagnostic analysis. The method provides two different residual plots for assessing the
fitted GEV model.

gevrlevelPlot
returns a vector containing the lower 95% bound of the confidence interval, the estimated return
level and the upper 95% bound.

hillPlot
displays a plot.

shaparmPlot
returns a list with one or two entries, depending on the selection of the input variable both.tails.
The two entries upper and lower determine the position of the tail. Each of the two variables is
again a list with entries pickands, hill, and dehaan. If one of the three methods will be discarded
the printout will display zeroes.

Note
GEV Parameter Estimation:

If method "mle" is selected the parameter fitting in gevFit is passed to the internal function
gev.mle or gumbel.mle depending on the value of gumbel, FALSE or TRUE. On the other hand, if
method "pwm" is selected the parameter fitting in gevFit is passed to the internal function gev.pwm
or gumbel.pwm again depending on the value of gumbel, FALSE or TRUE.

Author(s)
Alec Stephenson for Rs evd and evir package, and
Diethelm Wuertz for this R-port.

References
Coles S. (2001); Introduction to Statistical Modelling of Extreme Values, Springer.
Embrechts, P., Klueppelberg, C., Mikosch, T. (1997); Modelling Extremal Events, Springer.

Examples

## Load Data:
GevModelling 17

x = as.timeSeries(data(danishClaims))
colnames(x) <- "Danish"
head(x)

## hillPlot -
# Hill plot of heavy-tailed Danish fire insurance data
par(mfrow = c(1, 1))
hillPlot(x, plottype = "xi")
grid()

GevModelling Generalized Extreme Value Modelling

Description
A collection and description functions to estimate the parameters of the GEV distribution. To model
the GEV three types of approaches for parameter estimation are provided: Maximum likelihood
estimation, probability weighted moment method, and estimation by the MDA approach. MDA in-
cludes functions for the Pickands, Einmal-Decker-deHaan, and Hill estimators together with several
plot variants.

The GEV modelling functions are:

gevSim generates data from the GEV distribution,

gumbelSim generates data from the Gumbel distribution,
gevFit fits data to the GEV distribution,
gumbelFit fits data to the Gumbel distribution,
print print method for a fitted GEV object,
plot plot method for a fitted GEV object,
summary summary method for a fitted GEV object,
gevrlevelPlot k-block return level with confidence intervals.

Usage

gevSim(model = list(xi = -0.25, mu = 0, beta = 1), n = 1000, seed = NULL)

gumbelSim(model = list(mu = 0, beta = 1), n = 1000, seed = NULL)

gevFit(x, block = 1, type = c("mle", "pwm"), title = NULL, description = NULL, ...)
gumbelFit(x, block = 1, type = c("mle", "pwm"), title = NULL, description = NULL, ...)

## S4 method for signature 'fGEVFIT'

show(object)
## S3 method for class 'fGEVFIT'
plot(x, which = "ask", ...)
## S3 method for class 'fGEVFIT'
summary(object, doplot = TRUE, which = "all", ...)
18 GevModelling

Arguments
block block size.
description a character string which allows for a brief description.
doplot a logical. Should the results be plotted?
[shaparmPlot] -
a vector of logicals of the same lengths as tails defining for wich tail depths plots
should be created, by default plots will be generated for a tail depth of 5 percent.
By default c(FALSE, FALSE, FALSE, FALSE, TRUE, FALSE, FALSE, FALSE, FALSE, FALSE).
model [gevSim][gumbelSim] -
a list with components shape, location and scale giving the parameters of
the GEV distribution. By default the shape parameter has the value -0.25, the
location is zero and the scale is one. To fit random deviates from a Gumbel
distribution set shape=0.
n [gevSim][gumbelSim] -
number of generated data points, an integer value.
[rgev] -
the number of observations.
object [summary][grlevelPlot] -
a fitted object of class "gevFit".
seed [gevSim] -
an integer value to set the seed for the random number generator.
title [gevFit] -
a character string which allows for a project title.
type a character string denoting the type of parameter estimation, either by maximum
likelihood estimation "mle", the default value, or by the probability weighted
moment menthod "pwm".
which [plot][summary] -
a vector of logicals, one for each plot, denoting which plot should be displayed.
Alkternatively if which="ask" the user will be interactively asked which of the
plots should be desplayed. By default which="all".
x [dgev][devd] -
a numeric vector of quantiles.
[gevFit] -
data vector. In the case of method="mle" the interpretation depends on the value
of block: if no block size is specified then data are interpreted as block maxima;
if block size is set, then data are interpreted as raw data and block maxima are
calculated.
[hillPlot][shaparmPlot] -
the data from which to calculate the shape parameter, a numeric vector.
[print][plot] -
a fitted object of class "gevFit".
xi, mu, beta [*gev] -
xi is the shape parameter, mu the location parameter, and sigma is the scale
parameter. The default values are xi=1, mu=0, and beta=1. Note, if xi=0 the
distribution is of type Gumbel.
GevModelling 19

... [gevFit] -
control parameters optionally passed to the optimization function. Parameters
for the optimization function are passed to components of the control argument
of optim.
[hillPlot] -
other graphics parameters.
[plot][summary] -
arguments passed to the plot function.

Details

Parameter Estimation:

gevFit and gumbelFit estimate the parameters either by the probability weighted moment method,
method="pwm" or by maximum log likelihood estimation method="mle". The summary method
produces diagnostic plots for fitted GEV or Gumbel models.

Methods:

print.gev, plot.gev and summary.gev are print, plot, and summary methods for a fitted object of
class gev. Concerning the summary method, the data are converted to unit exponentially distributed
residuals under null hypothesis that GEV fits. Two diagnostics for iid exponential data are offered.
The plot method provides two different residual plots for assessing the fitted GEV model. Two
diagnostics for iid exponential data are offered.

Return Level Plot:

gevrlevelPlot calculates and plots the k-block return level and 95% confidence interval based
on a GEV model for block maxima, where k is specified by the user. The k-block return level is that
level exceeded once every k blocks, on average. The GEV likelihood is reparameterized in terms
of the unknown return level and profile likelihood arguments are used to construct a confidence
interval.

Hill Plot:

The function hillPlot investigates the shape parameter and plots the Hill estimate of the tail index
of heavy-tailed data, or of an associated quantile estimate. This plot is usually calculated from the
alpha perspective. For a generalized Pareto analysis of heavy-tailed data using the gpdFit function,
it helps to plot the Hill estimates for xi.

Shape Parameter Plot:

The function shaparmPlot investigates the shape parameter and plots for the upper and lower tails
the shape parameter as a function of the taildepth. Three approaches are considered, the Pickands
estimator, the Hill estimator, and the Decker-Einmal-deHaan estimator.
20 GevModelling

Value
gevSim
returns a vector of data points from the simulated series.

gevFit
returns an object of class gev describing the fit.

print.summary
prints a report of the parameter fit.

summary
performs diagnostic analysis. The method provides two different residual plots for assessing the
fitted GEV model.

gevrlevelPlot
returns a vector containing the lower 95% bound of the confidence interval, the estimated return
level and the upper 95% bound.

hillPlot
displays a plot.

shaparmPlot
returns a list with one or two entries, depending on the selection of the input variable both.tails.
The two entries upper and lower determine the position of the tail. Each of the two variables is
again a list with entries pickands, hill, and dehaan. If one of the three methods will be discarded
the printout will display zeroes.

Note
GEV Parameter Estimation:

If method "mle" is selected the parameter fitting in gevFit is passed to the internal function
gev.mle or gumbel.mle depending on the value of gumbel, FALSE or TRUE. On the other hand, if
method "pwm" is selected the parameter fitting in gevFit is passed to the internal function gev.pwm
or gumbel.pwm again depending on the value of gumbel, FALSE or TRUE.

Author(s)
Alec Stephenson for Rs evd and evir package, and
Diethelm Wuertz for this R-port.

References
Coles S. (2001); Introduction to Statistical Modelling of Extreme Values, Springer.
Embrechts, P., Klueppelberg, C., Mikosch, T. (1997); Modelling Extremal Events, Springer.
GevRisk 21

Examples
## gevSim -
# Simulate GEV Data, use default length n=1000
x = gevSim(model = list(xi = 0.25, mu = 0 , beta = 1), n = 1000)
head(x)

## gumbelSim -
# Simulate GEV Data, use default length n=1000
x = gumbelSim(model = list(xi = 0.25, mu = 0 , beta = 1))

## gevFit -
# Fit GEV Data by Probability Weighted Moments:
fit = gevFit(x, type = "pwm")
print(fit)

## summary -
# Summarize Results:
par(mfcol = c(2, 2))
summary(fit)

GevRisk Generalized Extreme Value Modelling

Description
A collection and description functions to estimate the parameters of the GEV distribution. To model
the GEV three types of approaches for parameter estimation are provided: Maximum likelihood
estimation, probability weighted moment method, and estimation by the MDA approach. MDA in-
cludes functions for the Pickands, Einmal-Decker-deHaan, and Hill estimators together with several
plot variants.

The GEV modelling functions are:

gevrlevelPlot k-block return level with confidence intervals.

Usage
gevrlevelPlot(object, kBlocks = 20, ci = c(0.90, 0.95, 0.99),
plottype = c("plot", "add"), labels = TRUE,...)

Arguments
add [gevrlevelPlot] -
whether the return level should be added graphically to a time series plot; if
FALSE a graph of the profile likelihood curve showing the return level and its
confidence interval is produced.
ci [hillPlot] -
22 GevRisk

probability for asymptotic confidence band; for no confidence band set ci to

zero.
kBlocks [gevrlevelPlot] -
specifies the particular return level to be estimated; default set arbitrarily to 20.
labels [hillPlot] -
whether or not axes should be labelled.
object [summary][grlevelPlot] -
a fitted object of class "gevFit".
plottype [hillPlot] -
whether alpha, xi (1/alpha) or quantile (a quantile estimate) should be plot-
ted.
... arguments passed to the plot function.

Details
Parameter Estimation:

gevFit and gumbelFit estimate the parameters either by the probability weighted moment method,
method="pwm" or by maximum log likelihood estimation method="mle". The summary method
produces diagnostic plots for fitted GEV or Gumbel models.

Methods:

print.gev, plot.gev and summary.gev are print, plot, and summary methods for a fitted object of
class gev. Concerning the summary method, the data are converted to unit exponentially distributed
residuals under null hypothesis that GEV fits. Two diagnostics for iid exponential data are offered.
The plot method provides two different residual plots for assessing the fitted GEV model. Two
diagnostics for iid exponential data are offered.

Return Level Plot:

gevrlevelPlot calculates and plots the k-block return level and 95% confidence interval based
on a GEV model for block maxima, where k is specified by the user. The k-block return level is that
level exceeded once every k blocks, on average. The GEV likelihood is reparameterized in terms
of the unknown return level and profile likelihood arguments are used to construct a confidence
interval.

Hill Plot:

The function hillPlot investigates the shape parameter and plots the Hill estimate of the tail index
of heavy-tailed data, or of an associated quantile estimate. This plot is usually calculated from the
alpha perspective. For a generalized Pareto analysis of heavy-tailed data using the gpdFit function,
it helps to plot the Hill estimates for xi.

Shape Parameter Plot:

GevRisk 23

The function shaparmPlot investigates the shape parameter and plots for the upper and lower tails
the shape parameter as a function of the taildepth. Three approaches are considered, the Pickands
estimator, the Hill estimator, and the Decker-Einmal-deHaan estimator.

Value

gevSim
returns a vector of data points from the simulated series.

gevFit
returns an object of class gev describing the fit.

print.summary
prints a report of the parameter fit.

summary
performs diagnostic analysis. The method provides two different residual plots for assessing the
fitted GEV model.

gevrlevelPlot
returns a vector containing the lower 95% bound of the confidence interval, the estimated return
level and the upper 95% bound.

hillPlot
displays a plot.

shaparmPlot
returns a list with one or two entries, depending on the selection of the input variable both.tails.
The two entries upper and lower determine the position of the tail. Each of the two variables is
again a list with entries pickands, hill, and dehaan. If one of the three methods will be discarded
the printout will display zeroes.

Note

GEV Parameter Estimation:

If method "mle" is selected the parameter fitting in gevFit is passed to the internal function
gev.mle or gumbel.mle depending on the value of gumbel, FALSE or TRUE. On the other hand, if
method "pwm" is selected the parameter fitting in gevFit is passed to the internal function gev.pwm
or gumbel.pwm again depending on the value of gumbel, FALSE or TRUE.

Author(s)

Alec Stephenson for Rs evd and evir package, and

Diethelm Wuertz for this R-port.
24 GpdDistribution

References
Coles S. (2001); Introduction to Statistical Modelling of Extreme Values, Springer.
Embrechts, P., Klueppelberg, C., Mikosch, T. (1997); Modelling Extremal Events, Springer.

Examples

## Load Data:
# BMW Stock Data - negative returns
x = -as.timeSeries(data(bmwRet))
colnames(x)<-"BMW"
head(x)

## gevFit -
# Fit GEV to monthly Block Maxima:
fit = gevFit(x, block = "month")
print(fit)

## gevrlevelPlot -
# Return Level Plot:
gevrlevelPlot(fit)

GpdDistribution Generalized Pareto Distribution

Description
A collection and description of functions to compute the generalized Pareto distribution. The func-
tions compute density, distribution function, quantile function and generate random deviates for the
GPD. In addition functions to compute the true moments and to display the distribution and random
variates changing parameters interactively are available.

The GPD distribution functions are:

dgpd Density of the GPD Distribution,

pgpd Probability function of the GPD Distribution,
qgpd Quantile function of the GPD Distribution,
rgpd random variates from the GEV distribution,
gpdMoments computes true mean and variance,
gpdSlider displays density or rvs from a GPD.

Usage
dgpd(x, xi = 1, mu = 0, beta = 1, log = FALSE)
pgpd(q, xi = 1, mu = 0, beta = 1, lower.tail = TRUE)
qgpd(p, xi = 1, mu = 0, beta = 1, lower.tail = TRUE)
rgpd(n, xi = 1, mu = 0, beta = 1)
GpdDistribution 25

gpdMoments(xi = 1, mu = 0, beta = 1)
gpdSlider(method = c("dist", "rvs"))

Arguments
log a logical, if TRUE, the log density is returned.
lower.tail a logical, if TRUE, the default, then probabilities are P[X <= x], otherwise,
P[X > x].
method [gpdSlider] -
a character string denoting what should be displayed. Either the density and
"dist" or random variates "rvs".
n [rgpd][gpdSim\ -
the number of observations to be generated.
p a vector of probability levels, the desired probability for the quantile estimate
(e.g. 0.99 for the 99th percentile).
q [pgpd] -
a numeric vector of quantiles.
x [dgpd] -
a numeric vector of quantiles.
xi, mu, beta xi is the shape parameter, mu the location parameter, and beta is the scale pa-
rameter.

Value
All values are numeric vectors:
d* returns the density,
p* returns the probability,
q* returns the quantiles, and
r* generates random deviates.

Author(s)
Alec Stephenson for the functions from Rs evd package,
Alec Stephenson for the functions from Rs evir package,
Alexander McNeil for the EVIS functions underlying the evir package,
Diethelm Wuertz for this R-port.

References
Embrechts, P., Klueppelberg, C., Mikosch, T. (1997); Modelling Extremal Events, Springer.

Examples
## rgpd -
par(mfrow = c(2, 2), cex = 0.7)
r = rgpd(n = 1000, xi = 1/4)
plot(r, type = "l", col = "steelblue", main = "GPD Series")
26 GpdModelling

grid()

## dgpd -
# Plot empirical density and compare with true density:
# Omit values greater than 500 from plot
hist(r, n = 50, probability = TRUE, xlab = "r",
col = "steelblue", border = "white",
xlim = c(-1, 5), ylim = c(0, 1.1), main = "Density")
box()
x = seq(-5, 5, by = 0.01)
lines(x, dgpd(x, xi = 1/4), col = "orange")

## pgpd -
# Plot df and compare with true df:
plot(sort(r), (1:length(r)/length(r)),
xlim = c(-3, 6), ylim = c(0, 1.1), pch = 19,
cex = 0.5, ylab = "p", xlab = "q", main = "Probability")
grid()
q = seq(-5, 5, by = 0.1)
lines(q, pgpd(q, xi = 1/4), col = "steelblue")

## qgpd -
# Compute quantiles, a test:
qgpd(pgpd(seq(-1, 5, 0.25), xi = 1/4 ), xi = 1/4)

GpdModelling GPD Distributions for Extreme Value Theory

Description
A collection and description to functions to compute the generalized Pareto distribution and to es-
timate its parameters. The functions compute density, distribution function, quantile function and
generate random deviates for the GPD. Two approaches for parameter estimation are provided:
Maximum likelihood estimation and the probability weighted moment method.

The GPD modelling functions are:

gpdSim generates data from the GPD,

gpdFit fits empirical or simulated data to the distribution,
print print method for a fitted GPD object of class ...,
plot plot method for a fitted GPD object,
summary summary method for a fitted GPD object.

Usage

gpdSim(model = list(xi = 0.25, mu = 0, beta = 1), n = 1000,

seed = NULL)
gpdFit(x, u = quantile(x, 0.95), type = c("mle", "pwm"), information =
GpdModelling 27

c("observed", "expected"), title = NULL, description = NULL, ...)

## S4 method for signature 'fGPDFIT'

show(object)
## S3 method for class 'fGPDFIT'
plot(x, which = "ask", ...)
## S3 method for class 'fGPDFIT'
summary(object, doplot = TRUE, which = "all", ...)

Arguments
description a character string which allows for a brief description.
doplot a logical. Should the results be plotted?
information whether standard errors should be calculated with "observed" or "expected"
information. This only applies to the maximum likelihood method; for the
probability-weighted moments method "expected" information is used if pos-
sible.
model [gpdSim] -
a list with components shape, location and scale giving the parameters of
the GPD distribution. By default the shape parameter has the value 0.25, the
location is zero and the scale is one.
n [rgpd][gpdSim\ -
the number of observations to be generated.
object [summary] -
a fitted object of class "gpdFit".
seed [gpdSim] -
an integer value to set the seed for the random number generator.
title a character string which allows for a project title.
type a character string selecting the desired estimation mehtod, either "mle" for the
maximum likelihood mehtod or "pwm" for the probability weighted moment
method. By default, the first will be selected. Note, the function gpd uses "ml".
u the threshold value.
which if which is set to "ask" the function will interactively ask which plot should
be displayed. By default this value is set to FALSE and then those plots will be
displayed for which the elements in the logical vector which ar set to TRUE; by
default all four elements are set to "all".
x [dgpd] -
a numeric vector of quantiles.
[gpdFit] -
the data vector. Note, there are two different names for the first argument x
and data depending which function name is used, either gpdFit or the EVIS
synonyme gpd.
[print][plot] -
a fitted object of class "gpdFit".
xi, mu, beta xi is the shape parameter, mu the location parameter, and beta is the scale pa-
rameter.
28 GpdModelling

... control parameters and plot parameters optionally passed to the optimization
and/or plot function. Parameters for the optimization function are passed to
components of the control argument of optim.

Details
Generalized Pareto Distribution:

Compute density, distribution function, quantile function and generates random variates for the
Generalized Pareto Distribution.
Simulation:

gpdSim simulates data from a Generalized Pareto distribution.

Parameter Estimation:

gpdFit fits the model parameters either by the probability weighted moment method or the maxim
log likelihood method. The function returns an object of class "gpd" representing the fit of a gen-
eralized Pareto model to excesses over a high threshold. The fitting functions use the probability
weighted moment method, if method method="pwm" was selected, and the the general purpose opti-
mization function optim when the maximum likelihood estimation, method="mle" or method="ml"
is chosen.

Methods:

print.gpd, plot.gpd and summary.gpd are print, plot, and summary methods for a fitted ob-
ject of class gpdFit. The plot method provides four different plots for assessing fitted GPD model.

gpd* Functions:

gpdqPlot calculates quantile estimates and confidence intervals for high quantiles above the thresh-
old in a GPD analysis, and adds a graphical representation to an existing plot. The GPD approxima-
tion in the tail is used to estimate quantile. The "wald" method uses the observed Fisher information
matrix to calculate confidence interval. The "likelihood" method reparametrizes the likelihood
in terms of the unknown quantile and uses profile likelihood arguments to construct a confidence
interval.

gpdquantPlot creates a plot showing how the estimate of a high quantile in the tail of a dataset
based on the GPD approximation varies with threshold or number of extremes. For every model
gpdFit is called. Evaluation may be slow. Confidence intervals by the Wald method may be fastest.

gpdriskmeasures makes a rapid calculation of point estimates of prescribed quantiles and expected
shortfalls using the output of the function gpdFit. This function simply calculates point estimates
and (at present) makes no attempt to calculate confidence intervals for the risk measures. If confi-
dence levels are required use gpdqPlot and gpdsfallPlot which interact with graphs of the tail of
a loss distribution and are much slower.
GpdModelling 29

gpdsfallPlot calculates expected shortfall estimates, in other words tail conditional expectation
and confidence intervals for high quantiles above the threshold in a GPD analysis. A graphical rep-
resentation to an existing plot is added. Expected shortfall is the expected size of the loss, given that
a particular quantile of the loss distribution is exceeded. The GPD approximation in the tail is used
to estimate expected shortfall. The likelihood is reparametrised in terms of the unknown expected
shortfall and profile likelihood arguments are used to construct a confidence interval.

gpdshapePlot creates a plot showing how the estimate of shape varies with threshold or number of
extremes. For every model gpdFit is called. Evaluation may be slow.

gpdtailPlot produces a plot of the tail of the underlying distribution of the data.

Value
gpdSim
returns a vector of datapoints from the simulated series.
gpdFit
returns an object of class "gpd" describing the fit including parameter estimates and standard errors.
gpdQuantPlot
returns invisible a table of results.
gpdShapePlot
returns invisible a table of results.
gpdTailPlot
returns invisible a list object containing details of the plot is returned invisibly. This object should
be used as the first argument of gpdqPlot or gpdsfallPlot to add quantile estimates or expected
shortfall estimates to the plot.

Author(s)
Alec Stephenson for the functions from Rs evd package,
Alec Stephenson for the functions from Rs evir package,
Alexander McNeil for the EVIS functions underlying the evir package,
Diethelm Wuertz for this R-port.

References
Embrechts, P., Klueppelberg, C., Mikosch, T. (1997); Modelling Extremal Events, Springer.
Hosking J.R.M., Wallis J.R., (1987); Parameter and quantile estimation for the generalized Pareto
distribution, Technometrics 29, 339349.

Examples
## gpdSim -
x = gpdSim(model = list(xi = 0.25, mu = 0, beta = 1), n = 1000)
## gpdFit -
par(mfrow = c(2, 2), cex = 0.7)
fit = gpdFit(x, u = min(x), type = "pwm")
print(fit)
30 gpdRisk

summary(fit)

gpdRisk GPD Distributions for Extreme Value Theory

Description
A collection and description to functions to compute tail risk under the GPD approach.

The GPD modelling functions are:

gpdQPlot estimation of high quantiles,

gpdQuantPlot variation of high quantiles with threshold,
gpdRiskMeasures prescribed quantiles and expected shortfalls,
gpdSfallPlot expected shortfall with confidence intervals,
gpdShapePlot variation of shape with threshold,
gpdTailPlot plot of the tail,
tailPlot ,
tailSlider ,
tailRisk .

Usage

gpdQPlot(x, p = 0.99, ci = 0.95, type = c("likelihood", "wald"),

like.num = 50)
gpdQuantPlot(x, p = 0.99, ci = 0.95, models = 30, start = 15, end = 500,
doplot = TRUE, plottype = c("normal", "reverse"), labels = TRUE,
...)
gpdSfallPlot(x, p = 0.99, ci = 0.95, like.num = 50)
gpdShapePlot(x, ci = 0.95, models = 30, start = 15, end = 500,
doplot = TRUE, plottype = c("normal", "reverse"), labels = TRUE,
...)
gpdTailPlot(object, plottype = c("xy", "x", "y", ""), doplot = TRUE,
extend = 1.5, labels = TRUE, ...)

gpdRiskMeasures(object, prob = c(0.99, 0.995, 0.999, 0.9995, 0.9999))

tailPlot(object, p = 0.99, ci = 0.95, nLLH = 25, extend = 1.5, grid =

TRUE, labels = TRUE, ...)
tailSlider(x)
tailRisk(object, prob = c(0.99, 0.995, 0.999, 0.9995, 0.9999), ...)

Arguments
ci the probability for asymptotic confidence band; for no confidence band set to
zero.
gpdRisk 31

doplot a logical. Should the results be plotted?

extend optional argument for plots 1 and 2 expressing how far x-axis should extend as
a multiple of the largest data value. This argument must take values greater than
1 and is useful for showing estimated quantiles beyond data.
grid ...
labels optional argument for plots 1 and 2 specifying whether or not axes should be
labelled.
like.num the number of times to evaluate profile likelihood.
models the number of consecutive gpd models to be fitted.
nLLH ...
object [summary] -
a fitted object of class "gpdFit".
p a vector of probability levels, the desired probability for the quantile estimate
(e.g. 0.99 for the 99th percentile).
reverse should plot be by increasing threshold (TRUE) or number of extremes (FALSE).
prob a numeric value.
plottype a character string.
start, end the lowest and maximum number of exceedances to be considered.
type a character string selecting the desired estimation mehtod, either "mle" for the
maximum likelihood mehtod or "pwm" for the probability weighted moment
method. By default, the first will be selected. Note, the function gpd uses "ml".
x [dgpd] -
a numeric vector of quantiles.
[gpdFit] -
the data vector. Note, there are two different names for the first argument x
and data depending which function name is used, either gpdFit or the EVIS
synonyme gpd.
[print][plot] -
a fitted object of class "gpdFit".
... control parameters and plot parameters optionally passed to the optimization
and/or plot function. Parameters for the optimization function are passed to
components of the control argument of optim.

Details
Generalized Pareto Distribution:

Compute density, distribution function, quantile function and generates random variates for the
Generalized Pareto Distribution.
Simulation:

gpdSim simulates data from a Generalized Pareto distribution.

32 gpdRisk

Parameter Estimation:

gpdFit fits the model parameters either by the probability weighted moment method or the maxim
log likelihood method. The function returns an object of class "gpd" representing the fit of a gen-
eralized Pareto model to excesses over a high threshold. The fitting functions use the probability
weighted moment method, if method method="pwm" was selected, and the the general purpose opti-
mization function optim when the maximum likelihood estimation, method="mle" or method="ml"
is chosen.

Methods:

print.gpd, plot.gpd and summary.gpd are print, plot, and summary methods for a fitted ob-
ject of class gpdFit. The plot method provides four different plots for assessing fitted GPD model.

gpd* Functions:

gpdqPlot calculates quantile estimates and confidence intervals for high quantiles above the thresh-
old in a GPD analysis, and adds a graphical representation to an existing plot. The GPD approxima-
tion in the tail is used to estimate quantile. The "wald" method uses the observed Fisher information
matrix to calculate confidence interval. The "likelihood" method reparametrizes the likelihood
in terms of the unknown quantile and uses profile likelihood arguments to construct a confidence
interval.

gpdquantPlot creates a plot showing how the estimate of a high quantile in the tail of a dataset
based on the GPD approximation varies with threshold or number of extremes. For every model
gpdFit is called. Evaluation may be slow. Confidence intervals by the Wald method may be fastest.

gpdriskmeasures makes a rapid calculation of point estimates of prescribed quantiles and expected
shortfalls using the output of the function gpdFit. This function simply calculates point estimates
and (at present) makes no attempt to calculate confidence intervals for the risk measures. If confi-
dence levels are required use gpdqPlot and gpdsfallPlot which interact with graphs of the tail of
a loss distribution and are much slower.

gpdsfallPlot calculates expected shortfall estimates, in other words tail conditional expectation
and confidence intervals for high quantiles above the threshold in a GPD analysis. A graphicalx
representation to an existing plot is added. Expected shortfall is the expected size of the loss, given
that a particular quantile of the loss distribution is exceeded. The GPD approximation in the tail
is used to estimate expected shortfall. The likelihood is reparametrised in terms of the unknown
expected shortfall and profile likelihood arguments are used to construct a confidence interval.

gpdshapePlot creates a plot showing how the estimate of shape varies with threshold or number of
extremes. For every model gpdFit is called. Evaluation may be slow.

gpdtailPlot produces a plot of the tail of the underlying distribution of the data.
gpdRisk 33

Value

gpdSim
returns a vector of datapoints from the simulated series.
gpdFit
returns an object of class "gpd" describing the fit including parameter estimates and standard errors.
gpdQuantPlot
returns invisible a table of results.
gpdShapePlot
returns invisible a table of results.
gpdTailPlot
returns invisible a list object containing details of the plot is returned invisibly. This object should
be used as the first argument of gpdqPlot or gpdsfallPlot to add quantile estimates or expected
shortfall estimates to the plot.

Author(s)

Alec Stephenson for the functions from Rs evd package,

Alec Stephenson for the functions from Rs evir package,
Alexander McNeil for the EVIS functions underlying the evir package,
Diethelm Wuertz for this R-port.

References

Embrechts, P., Klueppelberg, C., Mikosch, T. (1997); Modelling Extremal Events, Springer.
Hosking J.R.M., Wallis J.R., (1987); Parameter and quantile estimation for the generalized Pareto
distribution, Technometrics 29, 339349.

Examples

## Load Data:
danish = as.timeSeries(data(danishClaims))

## Tail Plot:
x = as.timeSeries(data(danishClaims))
fit = gpdFit(x, u = 10)
tailPlot(fit)

## Try Tail Slider:

# tailSlider(x)

## Tail Risk:
tailRisk(fit)
34 ValueAtRisk

TimeSeriesData Time Series Data Sets

Description
Data sets used in the examples of the timeSeries packages.

ValueAtRisk Value-at-Risk

Description
A collection and description of functions to compute Value-at-Risk and conditional Value-at-Risk

The functiona are:

VaR Computes Value-at-Risk,

CVaR Computes conditional Value-at-Risk.

Usage

VaR(x, alpha = 0.05, type = "sample", tail = c("lower", "upper"))

CVaR(x, alpha = 0.05, type = "sample", tail = c("lower", "upper"))

Arguments

x an uni- or multivariate timeSeries object

alpha a numeric value, the confidence interval.
type a character string, the type to calculate the value-at-risk.
tail a character string denoting which tail will be considered, either "lower" or
"upper". If tail="lower", then alpha will be converted to alpha=1-alpha.

Value

VaR
CVaR

returns a numeric vector or value with the (conditional) value-at-risk for each time series column.

Author(s)

Diethelm Wuertz for this R-port.

ValueAtRisk 35

See Also
hillPlot, gevFit.
Index

Topic distribution fGPDFIT-class (GpdModelling), 26

GpdDistribution, 24 findThreshold (DataPreprocessing), 2
GpdModelling, 26 fTHETA (ExtremeIndex), 4
gpdRisk, 30 fTHETA-class (ExtremeIndex), 4
Topic hplot
ExtremeIndex, 4 GevDistribution, 11
ExtremesData, 7 gevFit (GevModelling), 17
Topic models GevMdaEstimation, 13
GevDistribution, 11 GevModelling, 17
GevMdaEstimation, 13 gevMoments (GevDistribution), 11
GevModelling, 17 GevRisk, 21
GevRisk, 21 gevrlevelPlot (GevRisk), 21
ValueAtRisk, 34 gevSim (GevModelling), 17
Topic programming gevSlider (GevDistribution), 11
DataPreprocessing, 2 ghMeanExcessFit (ExtremesData), 7
ghtMeanExcessFit (ExtremesData), 7
as.POSIXct, 2, 3 GpdDistribution, 24
gpdFit (GpdModelling), 26
blockMaxima (DataPreprocessing), 2 GpdModelling, 26
blockTheta (ExtremeIndex), 4 gpdMoments (GpdDistribution), 24
bmwRet (TimeSeriesData), 34 gpdQPlot (gpdRisk), 30
clusterTheta (ExtremeIndex), 4 gpdQuantPlot (gpdRisk), 30
CVaR (ValueAtRisk), 34 gpdRisk, 30
gpdRiskMeasures (gpdRisk), 30
danishClaims (TimeSeriesData), 34 gpdSfallPlot (gpdRisk), 30
DataPreprocessing, 2 gpdShapePlot (gpdRisk), 30
deCluster (DataPreprocessing), 2 gpdSim (GpdModelling), 26
dgev (GevDistribution), 11 gpdSlider (GpdDistribution), 24
dgpd (GpdDistribution), 24 gpdTailPlot (gpdRisk), 30
gumbelFit (GevModelling), 17
emdPlot (ExtremesData), 7 gumbelSim (GevModelling), 17
exindexesPlot (ExtremeIndex), 4
exindexPlot (ExtremeIndex), 4 hillPlot (GevMdaEstimation), 13
ExtremeIndex, 4 hypMeanExcessFit (ExtremesData), 7
ExtremesData, 7
lilPlot (ExtremesData), 7
ferrosegersTheta (ExtremeIndex), 4
fGEVFIT (GevModelling), 17 mePlot (ExtremesData), 7
fGEVFIT-class (GevModelling), 17 mrlPlot (ExtremesData), 7
fGPDFIT (GpdModelling), 26 msratioPlot (ExtremesData), 7

36
INDEX 37

mxfPlot (ExtremesData), 7

nigMeanExcessFit (ExtremesData), 7
normMeanExcessFit (ExtremesData), 7

pgev (GevDistribution), 11
pgpd (GpdDistribution), 24
plot.fGEVFIT (GevModelling), 17
plot.fGPDFIT (GpdModelling), 26
pointProcess (DataPreprocessing), 2

qgev (GevDistribution), 11
qgpd (GpdDistribution), 24
qqparetoPlot (ExtremesData), 7

recordsPlot (ExtremesData), 7
rgev (GevDistribution), 11
rgpd (GpdDistribution), 24
runTheta (ExtremeIndex), 4

shaparmDEHaan (GevMdaEstimation), 13
shaparmHill (GevMdaEstimation), 13
shaparmPickands (GevMdaEstimation), 13
shaparmPlot (GevMdaEstimation), 13
show,fGEVFIT-method (GevModelling), 17
show,fGPDFIT-method (GpdModelling), 26
show,fTHETA-method (ExtremeIndex), 4
sllnPlot (ExtremesData), 7
ssrecordsPlot (ExtremesData), 7
summary.fGEVFIT (GevModelling), 17
summary.fGPDFIT (GpdModelling), 26

tailPlot (gpdRisk), 30
tailRisk (gpdRisk), 30
tailSlider (gpdRisk), 30
thetaSim (ExtremeIndex), 4
TimeSeriesData, 34

ValueAtRisk, 34
VaR (ValueAtRisk), 34

xacfPlot (ExtremesData), 7