svsample_roll: Rolling Estimation of Stochastic Volatility Models

Description

svsample_roll performs rolling window estimation based on svsample. A convenience function for backtesting purposes.

Usage

svsample_roll(
  y,
  designmatrix = NA,
  n_ahead = 1,
  forecast_length = 500,
  n_start = NULL,
  refit_every = 1,
  refit_window = c("moving", "expanding"),
  calculate_quantile = c(0.01),
  calculate_predictive_likelihood = TRUE,
  keep_draws = FALSE,
  parallel = c("no", "multicore", "snow"),
  n_cpus = 1L,
  cl = NULL,
  ...
)
svtsample_roll(
  y,
  designmatrix = NA,
  n_ahead = 1,
  forecast_length = 500,
  n_start = NULL,
  refit_every = 1,
  refit_window = c("moving", "expanding"),
  calculate_quantile = c(0.01),
  calculate_predictive_likelihood = TRUE,
  keep_draws = FALSE,
  parallel = c("no", "multicore", "snow"),
  n_cpus = 1L,
  cl = NULL,
  ...
)
svlsample_roll(
  y,
  designmatrix = NA,
  n_ahead = 1,
  forecast_length = 500,
  n_start = NULL,
  refit_every = 1,
  refit_window = c("moving", "expanding"),
  calculate_quantile = c(0.01),
  calculate_predictive_likelihood = TRUE,
  keep_draws = FALSE,
  parallel = c("no", "multicore", "snow"),
  n_cpus = 1L,
  cl = NULL,
  ...
)
svtlsample_roll(
  y,
  designmatrix = NA,
  n_ahead = 1,
  forecast_length = 500,
  n_start = NULL,
  refit_every = 1,
  refit_window = c("moving", "expanding"),
  calculate_quantile = c(0.01),
  calculate_predictive_likelihood = TRUE,
  keep_draws = FALSE,
  parallel = c("no", "multicore", "snow"),
  n_cpus = 1L,
  cl = NULL,
  ...
)

Value

The value returned is a list object of class svdraws_roll

holding a list item for every time window. The elements of these list items are

indices: a list object containing two elements: train is the vector of indices used for fitting the model, and test is the vector of indices used for prediction. The latter is mainly useful if a designmatrix is provided.
quantiles: the input parameter calculate_quantiles.
refit_every: the input parameter refit_every.
predictive_likelihood: present only if calculate_predictive_likelihood is TRUE. Then it is a number, the expected predictive density of the observation. The expecation is taken over the joint n_ahead predictive distribution of all model parameters.
predictive_quantile: present only if calculate_quantile is a non-empty vector. Then it is a vector of quantiles from the n_ahead predictive distribution of y. It is based on MCMC simulation by using predict.
fit: present only if keep_draws is TRUE. Then it is an svdraws object as returned by svsample.
prediction: present only if keep_draws is TRUE. Then it is an svpredict object as returned by predict.svdraws.

To display the output, use print and summary. The print method simply prints a short summary of the setup; the summary method displays the summary statistics of the backtesting.

Arguments

y: numeric vector containing the data (usually log-returns), which must not contain zeros. Alternatively, y can be an svsim object. In this case, the returns will be extracted and a message is signalled.
designmatrix: regression design matrix for modeling the mean. Must have length(y) rows. Alternatively, designmatrix may be a string of the form "arX", where X is a nonnegative integer. To fit a constant mean model, use designmatrix = "ar0" (which is equivalent to designmatrix = matrix(1, nrow = length(y))). To fit an AR(1) model, use designmatrix = "ar1", and so on. If some elements of designmatrix are NA, the mean is fixed to zero (pre-1.2.0 behavior of stochvol).
n_ahead: number of time steps to predict from each time window.
forecast_length: the time horizon at the end of the data set that is used for backtesting.
n_start: optional the starting time point for backtesting. Computed from forecast_length if omitted.
refit_every: the SV model is refit every refit_every time steps. Only the value 1 is allowed.
refit_window: one of "moving" or "expanding". If "expanding", then the start of the time window stays at the beginning of the data set. If "moving", then the length of the time window is constant throughout backtesting.
calculate_quantile: vector of numbers between 0 and 1. These quantiles are predicted using predict.svdraws for each time window.
calculate_predictive_likelihood: boolean. If TRUE, the n_ahead predictive density is evaluated at the n_ahead time observation after each time window.
keep_draws: boolean. If TRUE, the svdraws and the svpredict objects are kept from each time window.
parallel: one of "no" (default), "multicore", or "snow", indicating what type of parallellism is to be applied. Option "multicore" is not available on Windows.
n_cpus: optional positive integer, the number of CPUs to be used in case of parallel computations. Defaults to 1L. Ignored if parameter cl is supplied and parallel != "snow".
cl: optional so-called SNOW cluster object as implemented in package parallel. Ignored unless parallel == "snow".
...: Any extra arguments will be forwarded to svsample, controlling the prior setup, the starting values for the MCMC chains, the number of independent MCMC chains, thinning and other expert settings.

Details

Functions svtsample_roll, svlsample_roll, and svtlsample_roll are wrappers around svsample_roll with convenient default values for the SV model with t-errors, leverage, and both t-errors and leverage, respectively.

Examples

Run this code

# \donttest{
# Simulate from the true model
sim <- svsim(200)

# Perform rolling estimation using the vanilla SV
# model and default priors
roll <- svsample_roll(sim, draws = 5000, burnin = 2000,
                      keep_draws = TRUE,
                      forecast_length = 10,
                      n_ahead = 1, refit_every = 1,
                      refit_window = "moving",
                      calculate_predictive_likelihood = TRUE,
                      calculate_quantile = c(0.01, 0.05))

# Perform rolling estimation by making use
# of two CPU cores, advanced priors, and multiple
# chains with pre-set initial values. Let us combine
# that with an AR(2) specification
prior_beta <- sv_multinormal(c(1,0,-1), rbind(c(1, 0, 0.1),
                                              c(0, 0.3, -0.04),
                                              c(0.1, -0.04, 0.1)))
priorspec <- specify_priors(rho = sv_beta(4, 4),
                            latent0_variance = sv_constant(1),
                            beta = prior_beta,
                            nu = sv_exponential(0.05))
startpara <- list(list(mu = -9, phi = 0.3),
                  list(mu = -11, sigma = 0.1, phi = 0.95),
                  list(phi = 0.99))
roll <- svsample_roll(sim, draws = 5000, burnin = 2000,
                      designmatrix = "ar2",
                      priorspec = priorspec,
                      startpara = startpara,
                      parallel = "snow", n_cpus = 2,
                      n_chains = 3,
                      keep_draws = TRUE,
                      forecast_length = 10,
                      n_ahead = 1, refit_every = 1,
                      refit_window = "expanding",
                      calculate_predictive_likelihood = TRUE,
                      calculate_quantile = c(0.01, 0.05))
# }