bayesparams: Parameters for the semi-parametric approach

In tsxtreme: Bayesian Modelling of Extremal Dependence in Time Series

Description

Create, test or show objects of class "bayesparams".

Usage

bayesparams(prop.a = 0.02, prop.b = 0.02,
  prior.mu = c(0, 10), prior.nu = c(2, 1/2), prior.eta = c(2, 2),
  trunc = 100, comp.saved = 15, maxit = 30000,
  burn = 5000, thin = 1,
  adapt = 5000, batch.size = 125,
  mode = 1)

is.bayesparams(x)

Arguments

prop.a, prop.b	standard deviation for the Gaussian proposal of the Heffernan–Tawn parameters.
prior.mu	mean and standard deviation of the Gaussian prior for the components' means.
prior.nu	shape and rate of the inverse gamma prior for the components' variances.
prior.eta	shape and scale of the gamma prior for the precision parameter of the Dirichlet process.
trunc	integer; value of the truncation for the approximation of the infinite sum in the stick-breaking representation.
comp.saved	number of first components to be saved and returned.
maxit	maximum number of iterations.
burn	number of first iterations to discard.
thin	positive integer; spacing between iterations to be saved. Default is 1, i.e., all iterations are saved.
adapt	integer; number of iterations during which an adaption algorithm is applied to the proposal variances of α and β.
batch.size	size of batches used in the adaption algorithm. It has no effect if adapt==0.
mode	verbosity; 0 for debug mode, 1 (default) for standard output, and 2 for silent.
x	an arbitrary R object.

Details

prop.a is a vector of length 5 with the standard deviations for each region of the RAMA for the (Gaussian) proposal for α. If a scalar is given, 5 identical values are assumed.

prop.b is a vector of length 3 with the standard deviations for each region of the RAMA for the (Gaussian) proposal for β. If a scalar is provided, 3 identical values are assumed.

comp.saved has no impact on the calculations: its only purpose is to prevent from storing huge amounts of empty components.

The regional adaption scheme targets a 0.44 acceptance probability. It splits [-1;1] in 5 regions for α and [0;1] in 3 regions for β. The decision to increase/decrease the proposal standard deviation is based on the past batch.size MCMC iterations, so too low values yield inefficient adaption, while too large values yield slow adaption.

Default values for the hyperparameters are chosen to get reasonably uninformative priors.

See Also

bayesfit, depmeasure

Examples

is.bayesparams(bayesparams()) # TRUE
## use defaults, change max number of iteration of MCMC
par <- bayesparams(maxit=1e5)

tsxtreme documentation built on April 24, 2021, 1:07 a.m.