Changes in version 2025-11-10 New: - Option tails for thselect.vmetrics - Function thselect.egp for threshold selection with tests similar to Northop and Coleman routine. - Function thselect.cbm for threshold selection with semiparametric bootstrap mean square error estimation of shape parameter. Changes: - thselect.vmetric argument ci has been renamed to level (backward compatible). - Arguments method for tstab.egp and tstab.gp now has "lrt" instead of "profile" (choice backward compatible). - thselect.sdinfo gains an argument k to select the number of gaps, and returns a threshold thresh0 and k as part of the object output. The print method has been modified to reflect this. Fixes: - Function pegp did not work, and did not correctly handle lower.tail argument. - Functions degp and regp correctly handle the case shape=0 for models pt-beta and pt-gamma. - thselect.mdps now return shape parameter (rather than reciprocal shape), consistent with other semiparametric methods - thselect.pickands now correctly handles user-provided threshold vectors. - Warnings are not produced when fit.egp fails to obtain standard errors. - The print method for thselect.egp is more explicit about the lack of threshold returned if the hypothesis is rejected at the highest threshold considered (which returns NA). - thselect.ncpgp now passes ordered thresholds to inner function to avoid warning messages. - Function shape.hill now provides more informative error messages when passing vectors of invalid number of exceedances k. Changes in version 2025-10-22 Major release with breaking changes (mostly function names and arguments) for more coherence. Many functions now flagged as deprecated, but changes to the syntax are often backward compatible. New: - All functions producing plot now export objects with associated S3 routines. - Stein's weighted generalized Pareto estimator with fit.wgpd and suggested weights Stein_weights. - Multiple shape estimators (Hill, Pickands, moment estimator, generalized quantile) via fit.shape and specific routines shape.hill, shape.erm, shape.genjack, shape.lthill, shape.trunhill, shape.rbm, shape.moment, shape.pickands, shape.osz, shape.genquant and shape.vries. - Second order regular variation parameter estimation via fit.rho (four estimators available, via rho.dk, rho.fagh, rho.gbw and rho.ghp). - L-moment estimator for generalized Pareto distribution via gpd.lmom. - Estimator of tail dependence of Krupskii and Joe (kjtail) and through xdep.eta. - Weissman quantile estimator (qweissman) - Function dgeoaniso for geometric anisotropy, using the parametrization of Rai and Brown (2025) - Function maxstable for changes of parameters for GEV. - Distributions for extended generalized Pareto (d/p/q/r egp), with new models and profile likelihood (egp.pll). This also affects the plotting routines. The nomenclature has been modified for argument 'model' in egp.fit, egp.ll and egp.retlev with backward compatibility. - New methods for threshold selection (16 in total) thselect.alrs, thselect.bab, thselect.cv, theselect.expgqt, thselect.gbw, thselect.mdps, thselect.mrl (which supersedes automrl), thselect.pec, thselect.pickands, thselect.rbm, thselect.samsee, - Threshold stability plots for different estimators tstab.hill, tstab.lthill (with different diagnostics), tstab.mrl Changes: - Function PickandsXU is deprecated and replaced with shape.osz - Function fit.gpd with method obre now returns ordered exceedances and weights. - spunif now correctly handles missing values. - d/p/q/r for gev and gp now fully vectorized with respect to arguments, including shape parameter - confint.eprof has an argument for boundary non-regular null distribution (for the most common case). - Breaking changes to fit.egp, egp.retlev, tstab.egp. The former now allows fixed parameters (for profiling). - All functions for threshold selection are now named using the prefix thselect. Their arguments are somewhat standardized, and they each have distinct plot and print methods with automatic selection. - All threshold stability plots now have the same syntax, with prefix tstab. - Measures of extremal dependence now via functions prefixed with xdep., including xdep.chi, xdep.eta, xdep.chibar and xdep.asym. The first three call taildep, the last one supersedes xasym. - Function lambdadep deprecated, replaced with adf to better comply with current nomenclature. - Function automrl deprecated in favour of thselect.mrl. - Function chibar removed in favour of xdep.chibar. - Arguments of extcoef for marginal transformation margtrans have been modified (backward compatible changes). - Function fit.egp now includes alternative arguments to select the optimization routine - Internal function (exported, but hidden) jac has been replaced by jac_gpd_pareto to avoid ambiguity - Tests for max stability (maxstabtest) and independence via score (scoreindep) now renamed test.maxstab and test.scoreindep - Change to name infomat.test now thselect.sdinfo, vmetric now thselect.vmetric, W.diag now thselect.wseq, NC.diag now thselect.ncpgp - Arguments of xasym modified to make syntax more uniform. - Function mvrnorm renamed to rmvnorm. Fixes: - Function adf now allow for values above 1 for bivariate angular dependence function. - Function .infomat needs not have data argument. - Fix generalized Pareto fit with fixed parameters Changes in version 2024-07-09 Fixes: - gpd.ll and gpd.nll correctly return a non-NA value when xi=-1 (#18, reported by Paddy O'Toole) - tstab.gpd now correctly works when input leads to boundary values, and profile likelihood intervals are now correctly truncated in the range of admissible values - fit.gpd now relies on gpd.ll to return the negative log likelihood value nllh - Threshold stability plots now compute limits of plot safely, handling missing values in parameter estimates. Intervals are truncated to admissible values - For multivariate censored likelihood, marginal thresholds equal to the functional threshold (default option) doesn't return an error. Changes in version 2023-11-30 New: - Pickand's U-statistic estimator of the shape - New datasets: Newlyn wave height, pandemics, nutrients Fixes: - gp.tstab optimization bounds for profile confidence interval fixed. - rparp failed when a single draw was accepted (due to matrix being cast to vectors) - fit.extgp now handles model 0 (generalized Pareto), fixing #17 Changes in version 2023-04-23 New: - gev and gp distribution functions (d/p/q/r) introduced to remove dependency on evd package. - Ramos and Ledford score test of independence Fixes: - taildep function now correctly check choice of estimators against list of methods - Fix argument size in likmgp and clikmgp for the logistic model - Correctly return a covariance matrix with NAs when the matrix is singular or xi=-1 Changes: - taildep arguments leads to faster calculations with betacop option. - Replace nloptr dependency by Rsolnp Changes in version 2022-04-25 New: - bivariate coefficient of extremal asymmetry - four new families of max-stable models (pairwise beta, pairwise exponential, weighted Dirichlet and weighted exponential) for rmev, following Ballani and Schlather (2011) - maximum likelihood estimation routines (fit.gpd, fit.gev, etc.) now accept fixed parameters - mean residual life plots with weighted least square fit - coefficient of variation tests for threshold selection - Varty et al. metric based threshold selection diagnostic - anova method for mev_gpd and mev_gev objects - rmev and rmevspec now accept a distance matrix in place of coordinates for spatial models. - new datasets - website with vignettes Changes - Functions W.diag and NC.diag now have S3 plot and print methods - Changes to arguments (backward compatible) to xdat throughout - Many dependencies used by single functions are now listed in Suggest. Fixes: - ext.coef correctly handles arrays with missing values (reported by M. Jousset) - Optimization method in fit.gev now uses the PWM solution of Hosking (1985) as starting value - gev.pll now returns confidence intervals for param = "quant" (reported by D. Dupuis) - Fixes to NHPP order statistics density (returns -Inf outside of domain, also correctly evaluate for boundary case when xi=-1) - Optimization routines fit.pp, fit.gev and fit.rlarg now return correct MLE when solution lies on boundary (xi=-1) and are more robust to failure (gradients for nlminb return large values rather than NA which caused the algorithm to stop). - Grimshaw's algorithm sometimes returned incorrect value because of too low tolerance for eta near zero. Set back to default settings. - fit.gpd(..., method = "obre") now returns additional failure messages if the algorithm drifts towards infeasible values. - rparp now correctly handles xi=0 - Extended GP model now has 'step' for discretization, and a valid distribution function that returns real arguments whenever the input is finite (#9) - W.diag and egp.fitrange include arguments for changing 'par' (#10) - smith.penult now computes reciprocal hazard and it's derivative on the log scale (when possible) to avoid numerical overflow. Changes in version 2019-12-17 Fixes: - rlarg.infomat incorrect sign for expected information for d=1 - rmev function did not work properly for alog and aneglog (reported by Michael Lalancette) - Remove class()!= "matrix due to changes in R 4.0.0 - egp.retlev now returns invisible object (new ordering and format). Changes: - New S3 methods for objects returned by "fit" routines, for use in lax package Changes in version 2019-06-24 New: - Function taildep is multivariate equivalent to evd::chiplot - Function rparpcs for simulating from elliptical Pareto processes associated with max - rgparp for simulation of generalized R Pareto processes - Exponent measure for Brown-Resnick and extremal Student model - spunif for semi-parametric transform to uniform (tail modelled using GP) - rparp now has attributes to give acceptance rate of accept/reject procedure. - Exponent measure estimators extcoef - New method for robust OBRE estimates of GP parameters - Functions fit.gev, fit.gpd, fit.rlarg and fit.pp for maximum likelihood estimation - Default printing and plot optims (p-p and q-q plots) for mev_ objects - Changes to optimization routine for pp in W.diag function, use of expected information matrix, change to default tuning parameter - New datasets: eskrain, maiquetia, nidd, venice, w1500m Fixes: - Scaling of Brown-Resnick is now consistent with literature (half of semivariogram) - Degrees of freedom argument in C++ code for rmev with extremal Student family was incorrect - rparp now has a hard bound to ensure the vector of simulated vectors fits in memory. - Change to Grimshaw routine to ensure shape not less than -1, constrained optimization method - Information matrix, score for GEV now interpolated in a neighborhood of zero to preserve continuity and avoid numerical overflow. - Information matrix of GEV: scale factor off. All scores and information matrix have been checked and limits as xi -> 0 implemented. - ext.index: number of exceedances off by one, causing Inf in weight vector for lm (thanks to @MCristinaMiranda). Warnings now silenced by default. Changes: - coordinates for rmev, rparp, etc. now use coord instead of loc to avoid confusion with location parameter in rgparp - Removed dependency to ismev, rootsolve (replaced with nleqslv routine) and numDeriv. - Change to plot for profile log-likelihood methods - smith.penult function has new arguments (backward compatible). The quantiles 'u' are now returned for Smith penultimate approximations with method = "pot" Changes in version 2018-02-23 New: - Function rparp for simulation from R-Pareto Processes via rejection sampling - Function gev.pll and gpd.pll for penalized profile likelihood and tangent exponential model approximations - New functions chibar, angextrapo and lambdadep for bivariate and multivariate model estimation, based on work of Tawn et al. - Dirichlet mixture smoothing for empirical angular distribution of de Carvalho et al. (2013) - Functions gev.mle and gpd.mle for maximum likelihood estimates of transformed parameters - Functions gev.abias and gpd.abias for asymptotic bias of block maxima for fixed sample sizes or fixed thresholds Changes: - Functions rmev, rmevspec, etc. now only accept variogram functions vario that have distance as argument - Simulation from rmev and rmevspec now faster due to refactoring of code - Function smith.penult now has a 'family' as argument for specifying distributions via a string - Function gev.tem and gpd.tem are now wrappers for gev.pll and gpd.pll, respectively. Routine should be more robust - TEM corrections now handle more options - Clarifications in the vignette about the asymmetric negative logistic model (thanks to A. Stephenson) Fixes: - Fixed incorrect scaling in infomat.test (thanks to P. Northrop) - Model "br" now simulates from stationary version only if argument 'sigma' is provided, and otherwise samples intrinsically Gaussian processes - Display of p-value matrix for infomat.test Changes in version 2017-02-01 New: - Added 'negdir' model to rmev families - Changes to angmeas to include different weighting if the region of interest is 'max' or 'min' Fixes: - Fixed bug affecting angmeas in the bivariate case that would cause the method to crash Changes: - Fixed argument matching in function fit.egp with model egp2 Changes in version 2016-08-24 New: - Bias-correction, TEM added - Penultimate approximations (Papastathopoulos & Tawn, 2013), Naveau et al. (2014) and Smith (1987) Changes: - rmev, etc.: model "br" is now distinct from "hr" Fixes: - fixed invalid random number generation from logistic model for near-independence cases Changes in version 2016-06-07 New: - ext.index Extremal index estimates based on interexceedance and gap times - infomat.test Information matrix test of Suveges and Davison (2010) Fixes: - fixed an error in the acceptance rate for the gp.fit MCMC Changes in version 2016-03-15 Fixes: - fixed an error in the normal sampler (affecting version 1.5 and 1.6). All simulations of Brown-Resnick or extremal-Student were affected by the mistake Changes in version 2016-03-08 New: - Empirical and Euclidean likelihood estimation of spectral measure Changes: - gp.fit ample changes to the function, in particular a fix for the printing method, handling of errors and inclusion of the Zhang (2010) method and MCMC algorithm for the latter. This function is still preliminary and may updated in the nearby future to include further possibilities. Changes in version 2016-02-16 New: - Wadsworth (2015) Technometrics's proposal for threshold selection based on NHPP superposition - Northrop & Coleman diagnostic (2014) Extremes for shape equality and p-value path Fixes: - fixed error for simulation on grids Changes: - Check for marginal mean constraint for the Dirichlet mixture now has tolerance Changes in version 2015-10-05 Changes: - Extremal Dirichlet model now implemented with "ef" - Added Smith and asymmetric (negative) logistic model (differs from bivariate setting for aneglog, given that the later is not a valid DF according to Stephenson). - rmev can now return arrays for random fields on regular grids ("hr","exstud" and "smith" models). Changes in version 2015-08-23 Changes: - Added the negative bilogistic and the scaled Dirichlet models. - Extremal Dirichlet model implemented with "sm" only. Changes in version 2015-08-19 Changes: - Implementation of sampler from spectral distribution, moving rdirspec and rbilogspec to internal functions along with other functions. - Fixed a typo in rPextstud setting arguments of newly created arma vector to zero Changes in version 2014-08-16 Initial submission Changes in version 2.2 (2026-04-15) New: - New datasets, thames and cheeseborowind - Function rmar1 for first-order maximum autoregressive processes - Function gev.boot for Gaussian approximation to GEV, to be used in parametric bootstrap schemes - Functions test.blocksize, build.blocks, fit.gevblock, qqplot.blocksize for tests of max-stability based on comparison of block maxima, from Belzile and Davison (2026). - Tests of Durbin for uniformity (test.unif) and quantile-quantile plots for uniforms with binomial confidence intervals (qqplot.unif). Changes: - Parameter neval for thselect.vmetric replaced by pp to let users directly select probability points, as suggested in Collings et al. (2025) - Range for shape plots with semiparametric methods has been reduced - Function fit.gpd now allows for returnsamp parameter to disable return of vector of exceedances. - New method "pwm" implemented for fit.gpd. Bug fixes: - Package recompiled with Rcpp > 1.1.1 to avoid Rf_error - xdep.xindex is now more robust to boundary cases when theta=1, and returns upper confidence intervals even when the standard error cannot be obtained. The profile likelihood method also works for these cases.