CHANGES FROM VERSION 2.2-3 ARE AS FOLLOWS Removed defunct url from description file. Internal coding change to fpot.norm in order to avoid note produced under R CMD check. CHANGES FROM VERSION 2.2-2 ARE AS FOLLOWS Added a Suggests field to the description file in order to avoid warnings under future versions of R CMD check. CHANGES FROM VERSION 2.2-1 ARE AS FOLLOWS Fitted objects now contain the variance covariance matrix, which can be accessed by the function vcov. New function confint (methods confint.profile.evd and confint.evd) for calculating Wald and profile confidence intervals. The internal function pcint has now been deleted, and plot.profile.evd no longer returns profile confidence intervals invisibly. The function amvnonpar now implements non-parametric estimators for dependence functions of extreme value distributions of any dimension. The function amvevd now implements parametric evaluation of dependence functions of logistic and asymmetric logistic models of any dimension. I have removed the bootstrap confidence intervals option from chiplot. This had problems with zero empirical cdf estimates for some replications. A solution would be to use smoothed estimates, but I decided against implementing this. Argument trunc added to chiplot to control whether estimates are truncated at theoretical upper and lower bounds. The old interfaces to abvnonpar and amvnonpar have been removed. I have included a CITATION file in the /inst directory so that typing citation("evd") returns an appropriate citation. The internal functions ccop and ccop.case have been converted into a new documented function ccbvevd. The recommended R package "stats" is now a required package for evd. The files INDEX and data/00Index are now unnecessary (the information is now automatically generated from documentation files) and have been removed. The defunct function abvpar has been removed. CHANGES FROM VERSION 2.2-0 ARE AS FOLLOWS New function hbvevd for plotting the spectral density of bivariate extreme value models. The spectral density is also now plotted from plot.bvevd. Plotting diagnostics are now implemented for bivariate threshold methods using the method function plot.bvpot. The fbvpot function is now consistent with fbvevd; the dsm argument has been removed, and the asymmetric mixed model has been added. Also, the Husler-Reiss model has been added for (undocumented) Poisson process likelihood fitting. The new argument boot in the function chiplot allows bootstrap confidence intervals. The new argument spcase plots lines corresponding to special cases for comparison. The functions mrlplot and tcplot have a new argument pscale allowing the x-axis to be the threshold exceedance probability rather than the threshold. The default plotting charater in bivariate density plots is now a circle, to be consistent with quantile curves plots. Also, all plotting functions now produce all available plots by default. The function tcplot has a new argument vci allowing control over the plotting style of the confidence intervals. Other arguments allow more control over labels and limits. The clusters function with plot = TRUE would produce an error if no data points were above the threshold. This has been fixed. The demo soe9 has been expanded to include all of Chapter Nine of the text Statistics of Extremes. CHANGES FROM VERSION 2.1-7 ARE AS FOLLOWS Added the asymmetric mixed bivariate model to the function fbvevd. A demo soe9 has been included, giving examples from Chapter Nine of the book statistics of extremes. An associated file demo.txt has been added to the /inst directory. The atvnonpar and atvpar functions have been renamed to amvnonpar and amvevd. The abvpar function has been renamed to abvevd, but the former still runs, with a warning. An argument rev has been added to abvpar and abvnonpar, allowing the evaluation of A(1-t) rather than A(t). It can also be passed from plot.bvevd. The argument epmar has been added to abvnonpar and atvnonpar. It allows empirical estimation of the margins. The formerly internal function mtransform is now documented. It transforms to and from exponential distributions under the gev model. The profile plots now plot profile log-likelihoods rather than profile deviances. The plotting functions plot.uvevd and plot.bvevd have an additional argument cilwd to control the line width of confidence intervals. The functions abvpar and abvnonpar now have default axes labels and a border line width argument. The function abvnonpar is simpler; the "tdo" method is now undocumented, old methods "deheuvels" and "halltajvidi" can now be implemented using the argument madj. For back compatability, the old interface can still be used. The weight function option for method "cfg" has been removed, and the "cfg" definition now has a simpler equivalent representation. Similar changes have been made to atvnonpar (now renamed to amvnonpar). In fbvevd the dependence summaries, and hence the argument dsm, have been removed, though the former Dependence One value is now automatically given in the output. The argument warn.inf is now undocumented. The marginal option "exponential" to the function evmc has been changed to "rweibull" (negative exponential) because the former swaps the bivariate tails around. The function anova now has a logical argument half to deal with non-regular cases where the asymptotic distribution of the deviance difference is known to be one half of a chi-squared. Also, it no longer errors when testing bilog vs log or negbilog vs bilog. Added the dataset lossalae on general liability claims. CHANGES FROM VERSION 2.1-6 ARE AS FOLLOWS The exi function has been simplified; it now returns only a single estimate, but a new estimation method based on inter-exceedance times has been added. A new function exiplot plots estimates of the extremal index. The rlow argument (lower clustering interval) is now hidden in the documentation for fpot, clusters, exi and exiplot. Due to numerical problems, the function rmvevd could return Inf when the dependence parameters were small. This has been fixed. The maintainer thanks Mohammed Mehdi Gholam Rezaee for reporting this. The anova function errored when used inside another function. This is now corrected. The maintainer thanks William Valdar for reporting this. In function abvnonpar the method "hall" is now "halltajvidi", and an additional comment has been added to the help file, following a request from Nader Tajvidi. Typographic sign error in User's Guide corrected. The maintainer thanks C L Wong for reporting this. Spelling errors in uccle help file corrected. The maintainer thanks Tobias Verbeke for reporting this. Apologies to residents of Belgium! CHANGES FROM VERSION 2.1-5 ARE AS FOLLOWS The data passed to the bivariate fitting function fbvevd can now be a data frame with a third column of mode logical. See the corresponding documentation for more details. A new dataset called sealevel2 has been introduced. CHANGES FROM VERSION 2.1-4 ARE AS FOLLOWS The lower.tail argument for pbvevd and pmvevd did not properly produce survivor functions when set to FALSE. This has been fixed. CHANGES FROM VERSION 2.1-3 ARE AS FOLLOWS A file was accidently included in Version 2.1-3 which meant that the package would not easily install under C compilers other than gcc. This has been fixed. CHANGES FROM VERSION 2.1-2 ARE AS FOLLOWS The function dextreme errored when the argument largest was FALSE. This has been fixed. The maintainer thanks Brian Tolley for reporting this. CHANGES FROM VERSION 2.1-1 ARE AS FOLLOWS The function fbvevd now has arguments cshape, cscale and cloc for fitting common marginal parameters. The function fbvpot now has arguments cshape and cscale for fitting common marginal parameters. The Husler-Reiss model is now implemented in fbvpot. New function chiplot for plotting estimates of the dependence measures chi and chi-bar for bivariate data. The latex file for The Users' Guide is now included in the doc directory along with the pdf file. The default x-label "theshold" in mrlplot has been corrected to "threshold". CHANGES FROM VERSION 2.1-0 ARE AS FOLLOWS New function fbvpot for fitting bivariate threshold models. A plot method for these models will be implemented at a later date. New argument sym in function fbvevd, to allow the fitting of dependence structures under a symmetry constraint. The argument also exists in the new function fbvpot. CHANGES FROM VERSION 2.0-1 ARE AS FOLLOWS Univariate threshold models are now implemented. The main new function is fpot, which calculates maximum likelihood estimates under the generalized Pareto and point process representations. Density, distribution, quantile and simulation functions for the generalized Pareto distribution have been added. The new function clusters identifies extreme clusters. A related function exi calculates estimates of a quantity known as the extremal index. The new plotting functions mrlplot and tcplot aid threshold selection. The class structure has changed slightly. Models fitted using fextreme and forder still have class c("extreme", "evd"). Models fitted using fbvevd still have class c("bvevd", "evd"). Models fitted using the new function fpot have class c("pot", "uvevd", "evd"). Models fitted using fgev now have class c("gev", "uvevd", "evd"). The method function plot.gev is now plot.uvevd. This operates on both the gev and pot classes, due to the new class structure. The defunct functions fgumbel, frweibull, ffrechet (each defunct since version 1.2-0) and fbvall (defunct since version 2.0-0) have been removed. CHANGES FROM VERSION 2.0-0 ARE AS FOLLOWS Internal change to avoid warnings under R version 1.7.0 when calling the multivariate asymmetric logistic distribution, density, quantile and generation functions. The tests directory has been removed. The function evmc unintentionally reversed asymmetric dependence structures. This has been fixed. The maintainer thanks Chris Ferro for reporting this. CHANGES FROM VERSION 1.2-3 ARE AS FOLLOWS The function fgev.quantile is now internal; the functionality has been absorbed into fgev. It is now possible to parameterize gev model fits using the endpoint of the distribution by setting prob to zero or one in fgev. New functions for generating stochastic processes associated with extreme value theory. marma, mar and mma generate max autoregressive moving average processes. evmc generates first order Markov chains with bivariate extreme value dependence structures. The functions fbvlog, fbvalog, etc are now internal. The new function fbvevd should be used for the fitting of all bivariate models. The rbvevd, dbvevd and pbvevd functions replace individual functions for bivariate models. Similarly, rmvevd and dmvevd functions replace individual functions for multivariate models. The function abvpar replaces individual functions for plotting and calculating the dependence functions of parametric bivariate models. New functions atvnonpar and atvpar, which calculate and plot dependence functions of trivariate extreme value distributions, using non-parametric estimates and parametric models (at given parameter values). Fitted bivariate models (i.e. objects of class bvevd) now include Akaike's Information Criterion and, optionally, various dependence structure summaries. The argument dsm controls this option. New lower.tail argument in bivariate and multivariate distribution functions. The function pcint is now internal. Profile confidence intervals are now invisibly returned from plot.profile.evd. Also, the argument ci of plot.profile.evd can now be a vector. Functions called from plot.evd (dens, pp, qq, rl) and plot.bvevd (bvdens, bvdp, bvcpp) are now internal. The default labelling of the x-axis for the return level plot has been changed from the technically correct "-1/log(1-1/Return Period)" to the more widely used "Return Period". There now exists a density function for the multivariate asymmetric logistic model. The density function can be calculated by calling dmvevd with model = "alog". The argument mar of rmvevd, pmvevd and dmvevd can now be a list with d elements, where d is the dimension of the distribution. An extraction function logLik.evd has been added so that the function AIC.default in R base can be used. The function profile2d.evd has a new method argument (to be consistent with profile.evd) and new arguments xaxs and yaxs (to override the default behaviour of the function image). The function plot.bvevd has new arguments blty (border line type) and grid. In the function abvnonpar the logical argument convex replaces the numeric argument modify. The functions fext, rext, etc are now called fextreme, rextreme, etc. The objects formery of class "evd" are now of class c("gev","evd") or c("extreme","evd"). The plot.evd function is now plot.gev. The row names of all data.frame datasets are now the years of observation. Furthermore, the period of observation for the venice data has been corrected in the help file. The arguments mesh and conf in profile.evd now work as documented. The fbvall function is defunct. New datasets: failure, fox, lisbon, ocmulgee, oldage, sask and uccle. The CHANGES file has been moved to the top-level directory. CHANGES FROM VERSION 1.2-2 ARE AS FOLLOWS Two more datasets - venice and portpirie, intended for use in the evdbayes package. Fixed minor problem with extraction functions for bvall objects when only one model is fitted. Contours for bivariate density plots are now chosen by the contour function. Error messages in internal functions are more informative. CHANGES FROM VERSION 1.2-1 ARE AS FOLLOWS New methods hall and tdo for calculating non-parametric dependence function estimates. BUG fix: the non-parameteric dependence function estimator of Caperaa et al (the default) was plotting/calculating A(1-x) rather than A(x). The recommended citation for the package is now the article included in R-News Volume 2/2. Extra graphical arguments have been included in the abv[...] functions. CHANGES FROM VERSION 1.2-0 ARE AS FOLLOWS The profile.evd function now has a method argument to specify the optimization method. The default method is now BFGS. A BUG existed in profile.evd; a fatal error would result if the mesh argument did not have a names attribute. This has been fixed. Internal: General code cleaning/optimizing. New internal functions have been created. CHANGES FROM VERSION 1.1-0 ARE AS FOLLOWS Class orientated objects have been introduced, including print methods and extraction functions. Diagnostics plots can be implemented using plot, or the lower level functions dens, rl, pp, qq (univariate) and bvdens, bvcpp, bvdp (bivariate). Parameters can be profiled using the functions profile and profile2d. Profile deviance surfaces can be plotted using plot. Profile confidence intervals can be calculated using pcint. Fitted models can be compared using the function anova. New function fgev.quantile to fit the GEV distribution, re-parameterizing using a quantile. This allows profile deviances of extreme quantiles to be plotted. Extra argument corr for fitting functions. If corr is TRUE the correlation matrix is calculated. By default corr is FALSE. Extra argument warn.inf for fitting functions. When warn.inf is TRUE (the default) a warning is given if the negative log-likelihood is infinite at the starting values. The ffrechet, fgumbel and frweibull functions are defunct. Internal: Function ccop for calculating conditional copulas. CHANGES FROM VERSION 1.0-0 ARE AS FOLLOWS Automated starting values for fitting functions. Compiled code is now used within the bivariate fitting routines and all bivariate simulation functions. They are consequently much faster than those in Version 1.0-0. Simulation, distribution, density and fitting functions for the bivariate bilogistic, bivariate negative bilogistic and bivariate Coles-Tawn models. A new function fbvall which fits all bivariate models simultaneously. For every model it returns maximum likelihood estimates, standard errors, criteria for model comparisons based on the deviance (e.g. AIC), and summaries of the dependence structure (see the help file for details). A new function abvnonpar which calculates or plots non-parametric estimates of the dependence function. All functions explicitly allow for missing values. This includes bivariate fitting functions, where missing values can occur on either or both margins within any observation. Extra argument nsloc for univariate fitting functions, and arguments nsloc1 and nsloc2 for bivariate fitting functions. These allow non-stationary fitting using linear models for the location parameters. Extra argument std.err for fitting functions. If std.err is FALSE the standard errors are not calculated. By default std.err is TRUE. Extra argument method for fitting functions. The method argument explicitly specifies the optimization method to be passed to the function optim. The default method is now BFGS for all fitting functions except fext and forder, where the default method is still Nelder-Mead. The sealevel data frame has been expanded to include observations from 1912 to 1992. There are 39 missing values. Explicit error handling in fitting functions when the observed information matrix is singular. Artificial constraints are now placed on dependence parameters within bivariate fitting functions to prevent numerical problems. The asy argument for the bivariate asymmetric logistic and bivariate asymmetric negative logistic models now defaults to c(1,1). The default values of the marginal parameters for the bivariate and multivariate functions in Version 1.0-0 are different. This is counter intuitive, so the default value for each margin within the multivariate functions is now c(0,1,0) (Gumbel), which is consistent with the functions for bivariate models. The bivariate density functions contained a fairly minor BUG; if one of the marginal parameter arguments were passed a three column matrix and some of the rows of the matrix produced zero density, a fatal error would result. This has now been fixed. Internal: The .C interface is called with PACKAGE = "evd". Internal: R_alloc replaces malloc in C routines.