utilities_svdraws.R

#  #####################################################################################
#  R package stochvol by
#     Gregor Kastner Copyright (C) 2013-2018
#     Gregor Kastner and Darjus Hosszejni Copyright (C) 2019-
#
#  This file is part of the R package stochvol: Efficient Bayesian
#  Inference for Stochastic Volatility Models.
#
#  The R package stochvol is free software: you can redistribute it
#  and/or modify it under the terms of the GNU General Public License
#  as published by the Free Software Foundation, either version 2 or
#  any later version of the License.
#
#  The R package stochvol is distributed in the hope that it will be
#  useful, but WITHOUT ANY WARRANTY; without even the implied warranty
#  of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
#  General Public License for more details.
#
#  You should have received a copy of the GNU General Public License
#  along with the R package stochvol. If that is not the case, please
#  refer to <http://www.gnu.org/licenses/>.
#  #####################################################################################

read_argument_chain <- function (chain) {
  if (!(is.character(chain) && length(chain) == 1L) &&
      !(is.numeric(chain) && as.integer(chain) > 0 && length(chain) == 1L)) {
    stop("Parameter 'chain' for the extractor functions has to be either a chain index (single positive integer) or one of 'concatenated' (default) or 'all'.")
  }
  if (is.character(chain)) {
    match.arg(chain, c("concatenated", "all"))
  } else {
    as.integer(chain)
  }
}

general_extractor <- function (x, chain, name) {
  chain <- read_argument_chain(chain)
  dat <- x[[name]]
  if (is.integer(chain)) {
    dat[[chain]]
  } else if (chain == "all") {
    dat
  } else {
    pars <- coda::mcpar(dat[[1]])
    coda::mcmc(join_mcmclist(dat), pars[1], pars[2], pars[3])
  }
}

#' Common Extractors for 'svdraws' and 'svpredict' Objects
#'
#' Some simple extractors returning the corresponding element of an
#' \code{svdraws} and \code{svpredict} object.
#'
#' @name extractors
#' @aliases para latent latent0 priors thinning runtime svbeta observations vola
#' @param x \code{svdraws} object.
#' @param y \code{svpredict} object.
#' @param chain \emph{optional} either a positive integer or the string
#' \code{"concatenated"} (default) or the string \code{"all"}.
#' @return The return value depends on the actual funtion.
#' \item{para(x, chain = "concatenated")}{extracts the parameter draws.}
#' \item{latent(x, chain = "concatenated")}{extracts the latent contemporaneous
#' log-volatility draws.}
#' \item{latent0(x, chain = "concatenated")}{extracts the latent initial log-volatility draws.}
#' \item{svbeta(x, chain = "concatenated")}{extracts the linear regression coefficient draws.}
#' \item{svtau(x, chain = "concatenated")}{extracts the tau draws.}
#' \item{vola(x, chain = "concatenated")}{extracts standard deviation draws.}
#' \item{priors(x)}{extracts the prior
#' parameters used and returns them in a \code{prior_spec} object as generated by
#' \link{specify_priors}.}
#' \item{thinning(x)}{extracts the thinning parameters used and returns them in
#' a \code{list}.}
#' \item{runtime(x)}{extracts the runtime and returns it as a
#' \code{proc_time} object.}
#' \item{sampled_parameters(x)}{returns the names of time independent model
#' parameters that were actually sampled by \code{svsample}.}
#' \item{predlatent(y, chain = "concatenated")}{extracts the predicted latent contemporaneous
#' log-volatility draws.}
#' \item{predvola(y, chain = "concatenated")}{extracts predicted standard deviation draws.}
#' \item{predy(y, chain = "concatenated")}{extracts the predicted observation draws.}
#' Functions that have input parameter \code{chain} return
#' an \code{mcmc.list} object if \code{chain=="all"} and
#' return an \code{mcmc} object otherwise. If \code{chain} is
#' an integer, then the specified chain is selected from
#' all chains. If \code{chain} is \code{"concatenated"},
#' then all chains are merged into one \code{mcmc} object.
#' @keywords utilities
#' @seealso \link{specify_priors}, \link{svsample}, \link{predict.svdraws}
#' @example inst/examples/extractors.R
#' @export
para <- function (x, chain = "concatenated") {
  general_extractor(x, chain, "para")
}

#' @rdname extractors
#' @export
latent0 <- function (x, chain = "concatenated") {
  general_extractor(x, chain, "latent0")
}

#' @rdname extractors
#' @export
latent <- function (x, chain = "concatenated") {
  general_extractor(x, chain, "latent")
}

#' @rdname extractors
#' @export
vola <- function (x, chain = "concatenated") {
  if (identical(chain, "a") || identical(chain, "al") || identical(chain, "all")) {  # very elegant match.arg
    vol <- list()
    for (chain in seq_len(coda::nchain(x$latent))) {
      tau_chain <- if (!is.null(x$tau)) x$tau[[chain]] else 1
      pars <- mcpar(x$latent[[chain]])
      vol[[chain]] <- coda::mcmc(exp(x$latent[[chain]]/2) * sqrt(tau_chain), pars[1], pars[2], pars[3])
    }
    coda::mcmc.list(vol)
  } else {
    lat <- latent(x, chain)
    tau <- svtau(x, chain)
    pars <- mcpar(x$latent[[1]])
    coda::mcmc(exp(lat/2) * sqrt(tau), pars[1], pars[2], pars[3])
  }
}

#' @rdname extractors
#' @export
svbeta <- function (x, chain = "concatenated") {
  chain <- read_argument_chain(chain)
  if (is.null(x$beta)) {
    if (chain == "all") {
      NULL
    } else {
      0
    }
  } else {
    general_extractor(x, chain, "beta")
  }
}

#' @rdname extractors
#' @export
svtau <- function (x, chain = "concatenated") {
  chain <- read_argument_chain(chain)
  if (is.null(x$tau)) {
    if (chain == "all") {
      NULL
    } else {
      1
    }
  } else {
    general_extractor(x, chain, "tau")
  }
}

#' @rdname extractors
#' @export
priors <- function (x) {
  x$priors
}

#' @rdname extractors
#' @export
thinning <- function (x) {
  x$thinning
}

#' @rdname extractors
#' @export
runtime <- function (x) {
 x$runtime
}

#' @rdname extractors
#' @export
sampled_parameters <- function (x) {
  pr <- if (inherits(x, "svdraws")) {
    priors(x)
  } else if (inherits(x, "sv_priorspec")) {
    x
  } else {
    stop("Argument 'x' must be either an svdraws object or an sv_priorspec object.")
  }
  res <- c()
  if (!inherits(pr$mu, "sv_constant")) {
    res <- c(res, "mu")
  }
  if (!inherits(pr$phi, "sv_constant")) {
    res <- c(res, "phi")
  }
  if (!inherits(pr$sigma2, "sv_constant")) {
    res <- c(res, "sigma")
  }
  if (!inherits(pr$nu, "sv_constant") && !inherits(pr$nu, "sv_infinity")) {
    res <- c(res, "nu")
  }
  if (!inherits(pr$rho, "sv_constant")) {
    res <- c(res, "rho")
  }
  res
}

#' @rdname extractors
#' @export
predy <- function (y, chain = "concatenated") {
  general_extractor(y, chain, "y")
}

#' @rdname extractors
#' @export
predlatent <- function (y, chain = "concatenated") {
  general_extractor(y, chain, "h")
}

#' @rdname extractors
#' @export
predvola <- function (y, chain = "concatenated") {
  general_extractor(y, chain, "vol")
}

#' @export
"[.svdraws" <- function (x, n, drop=FALSE) {
  if (drop) {
    stop("Parameter 'drop' is only allowed to be FALSE")
  }
  res <- x
  mcmcnames <- c("para", "latent", "latent0", "beta", "tau")
  for (mcmcname in mcmcnames) {
    if (!is.null(res[[mcmcname]])) {
      if (length(n) > 1) {
        res[[mcmcname]] <- mcmc.list(res[[mcmcname]][n])
      } else {
        res[[mcmcname]] <- mcmc.list(list(res[[mcmcname]][[n]]))
      }
    }
  }
  res
}

#' @export
"[.svpredict" <- function (x, n, drop=FALSE) {
  if (drop) {
    stop("Parameter 'drop' is only allowed to be FALSE")
  }
  res <- x
  mcmcnames <- c("y", "h")
  for (mcmcname in mcmcnames) {
    if (!is.null(res[[mcmcname]])) {
      if (length(n) > 1) {
        res[[mcmcname]] <- mcmc.list(res[[mcmcname]][n])
      } else {
        res[[mcmcname]] <- mcmc.list(list(res[[mcmcname]][[n]]))
      }
    }
  }
  res
}

#' @method as.array svdraws
#' @export
as.array.svdraws <- function (x, ...) {
  aperm(simplify2array(lapply(seq_along(para(x, "all")), function (i) cbind(x$para[[i]], x$latent0[[i]], x$latent[[i]], x$tau[[i]], x$beta[[i]])), higher = TRUE), c(1, 3, 2))
}

#' @method as.array svpredict
#' @export
as.array.svpredict <- function (x, ...) {
  aperm(simplify2array(lapply(seq_along(x$y), function (i) cbind(x$y[[i]], x$h[[i]], x$vol[[i]])), higher = TRUE), c(1, 3, 2))
}

contains_beta <- function(x) {
  NROW(x$beta) > 0
}

contains_tau <- function(x) {
  NROW(x$tau) > 0
}

contains_indicators <- function(x) {
  NROW(x$indicators) > 0
}

join_mcmclist <- function (x) {
  do.call(rbind, x)
}

# x is svdraws with a single chain
# output: svdraws-like object but with mcmc objects instead of the mcmc.list objects
flatten <- function (x) {
  res <- x
  mcmcnames <- c("para", "latent", "latent0", "beta", "tau")
  for (mcmcname in mcmcnames) {
    if (!is.null(res[[mcmcname]])) {
      res[[mcmcname]] <- res[[mcmcname]][[1]]
    }
  }
  class(res) <- NULL
  res
}

#' Updating the Summary of MCMC Draws
#'
#' Creates or updates a summary of an \code{svdraws} object.
#'
#' \code{updatesummary} will always calculate the posterior mean and the
#' posterior standard deviation of the raw draws and some common
#' transformations thereof. Moroever, the posterior quantiles, specified by the
#' argument \code{quantiles}, are computed. If \code{esspara} and/or
#' \code{esslatent} are \code{TRUE}, the corresponding effective sample size
#' (ESS) will also be included.
#'
#' @param x \code{svdraws} object.
#' @param quantiles numeric vector of posterior quantiles to be computed. The
#' default is \code{c(0.05, 0.5, 0.95)}.
#' @param esspara logical value which indicates whether the effective sample
#' size (ESS) should be calculated for the \emph{parameter draws}. This is
#' achieved by calling \code{\link[coda]{effectiveSize}} from the \code{coda}
#' package. The default is \code{TRUE}.
#' @param esslatent logical value which indicates whether the effective sample
#' size (ESS) should be calculated for the \emph{latent log-volatility} draws.
#' This is achieved by calling \code{\link[coda]{effectiveSize}} from the
#' \code{coda} package. The default is \code{FALSE}, because this can be quite
#' time-consuming when many latent variables are present.
#' @return The value returned is an updated list object of class \code{svdraws}
#' holding \item{para}{\code{mcmc} object containing the \emph{parameter} draws
#' from the posterior distribution.} \item{latent}{\code{mcmc} object
#' containing the \emph{latent instantaneous log-volatility} draws from the
#' posterior distribution.} \item{latent0}{\code{mcmc} object containing the
#' \emph{latent initial log-volatility} draws from the posterior distribution.}
#' \item{y}{argument \code{y}.} \item{runtime}{\code{"proc_time"} object
#' containing the run time of the sampler.} \item{priors}{\code{list}
#' containing the parameter values of the prior distribution, i.e. the
#' arguments \code{priormu}, \code{priorphi}, \code{priorsigma} (and
#' potentially \code{nu}).} \item{thinning}{\code{list} containing the thinning
#' parameters, i.e. the arguments \code{thinpara}, \code{thinlatent} and
#' \code{keeptime}.} \item{summary}{\code{list} containing a collection of
#' summary statistics of the posterior draws for \code{para}, \code{latent},
#' and \code{latent0}.}
#'
#' To display the output, use \code{print}, \code{summary} and \code{plot}. The
#' \code{print} method simply prints the posterior draws (which is very likely
#' a lot of output); the \code{summary} method displays the summary statistics
#' currently stored in the object; the \code{plot} method gives a graphical
#' overview of the posterior distribution by calling \code{\link{volplot}},
#' \code{\link{traceplot}} and \code{\link{densplot}} and displaying the
#' results on a single page.
#' @note \code{updatesummary} does not actually overwrite the object's current
#' summary, but in fact creates a new object with an updated summary. Thus,
#' don't forget to overwrite the old object if this is want you intend to do.
#' See the examples below for more details.
#' @seealso \code{\link{svsample}}
#' @keywords utilities
#' @example inst/examples/updatesummary.R
#' @export
updatesummary <- function(x, quantiles = c(.05, .5, .95), esspara = TRUE, esslatent = FALSE) {
  sampled_para <- sampled_parameters(x)

  matrixsummary <- function (x, quants = quantiles) {
    cbind(mean = colMeans(x), sd = apply(x, 2, sd), t(apply(x, 2, quantile, prob = quants)))
  }

  # x is an mcmc.list object containing matrices
  # transformation is a singleton named list with a vectorized function of interest for the variable(s)
  summaryfunction <- function(x, quants = quantiles, ess = TRUE, transformation = NULL) {
    stopifnot(inherits(x, "mcmc.list"))
    x_join <- join_mcmclist(x)
    res <- matrixsummary(x_join, quants = quants)
    if (ess) {
      res <- cbind(res, ESS = as.numeric(coda::effectiveSize(x)))  # computes the sum of effective sample sizes over the chains
    }
    if (!is.null(transformation)) {
      stopifnot(length(transformation) == 1)
      name <- names(transformation)
      stopifnot(length(name) == 1)
      x_trans <- transformation[[name]](x_join)
      toadd <- colMeans(x_trans)
      toadd <- cbind(toadd, apply(x_trans, 2, sd))
      colnames(toadd) <- paste0(c("mean", "sd"), "(", name, ")")
      res <- cbind(res, toadd)
    }
    res
  }

  res <- list()

  res$para <- summaryfunction(x$para[, sampled_para, drop=FALSE], ess = esspara)
  if ("mu" %in% sampled_para) {
    toadd <- matrixsummary(exp(join_mcmclist(x$para[, "mu", drop=FALSE])/2))
    if (esspara) {
      toadd <- cbind(toadd, ESS = res$para["mu", "ESS"])
    }
    rownames(toadd) <- "exp(mu/2)"
    res$para <- rbind(res$para, toadd)
  }
  if ("sigma" %in% sampled_para) {
    toadd <- matrixsummary(join_mcmclist(x$para[, "sigma", drop=FALSE])^2)
    if (esspara) {
      toadd <- cbind(toadd, ESS = res$para["sigma", "ESS"])
    }
    rownames(toadd) <- "sigma^2"
    res$para <- rbind(res$para, toadd)
  }

  res$latent0 <- summaryfunction(x$latent0, ess = esslatent, transformation = list("exp(h_t/2)" = function (x) exp(x/2)))
  res$latent <- summaryfunction(x$latent, ess = esslatent, transformation = list("exp(h_t/2)" = function (x) exp(x/2)))

  vol <- vola(x, "all")
  res$sd <- summaryfunction(vol, ess = esslatent)
  rownames(res$sd) <- gsub("h", "sd", rownames(vol))

  if (contains_beta(x)) {
    res$beta <- summaryfunction(svbeta(x, "all"), ess = esspara)
  }

  x$summary <- res
  x
}

#' @export
summary.svdraws <- function (object, showpara = TRUE, showlatent = FALSE, ...) {
  ret <- vector("list")
  class(ret) <- "summary.svdraws"
  ret$mcp <- mcpar(para(object, 1))
  ret$mcl <- mcpar(latent(object, 1))
  ret$priors <- priors(object)
  if (isTRUE(showpara)) {
    ret$para <- object$summary$para
    ret$beta <- object$summary$beta
  }
  if (isTRUE(showlatent)) {
    ret$latent <- object$summary$latent
    ret$latent <- rbind("h_0" = object$summary$latent0, ret$latent)
  }
  if (isTRUE(object$resampled)) {
    ret$resampled <- list(para = list(max_entropy = log(length(object$correction_weight_para[[1]])),
                                      entropy = min(sapply(object$correction_weight_para, function (w) -sum(w * log(w))))),
                          latent = list(max_entropy = log(length(object$correction_weight_latent[[1]])),
                                        entropy = min(sapply(object$correction_weight_latent, function (w) -sum(w * log(w))))),
                          same_resampling = object$thinning$para == object$thinning$latent)
  }
  ret
}

#' @export
print.svdraws <- function (x, showpara = TRUE, showlatent = FALSE, ...) {
  print(summary(x, showpara = showpara, showlatent = showlatent), ...)
  invisible(x)
}

#' @method print summary.svdraws
#' @export
print.summary.svdraws <- function (x, showpriorbeta = !is.null(x$beta), ...) {
  cat("\nSummary of 'svdraws' object\n\n")
  print(x$priors, showbeta = showpriorbeta)
  cat("\nStored ", (x$mcp[2] - x$mcp[1] + 2*x$mcp[3] - 1) %/% x$mcp[3], " MCMC draws after a burn-in of ", x$mcp[1]-x$mcp[3], ".\n", sep="")
  if (x$mcp[3] == 1) {
    cat("No thinning.\n")
  } else {
    cat("Thinning: ", x$mcp[3], ".\n", sep="")
  }

  if (exists("para", x)) {
    cat("\nPosterior draws of SV parameters (thinning = ", x$mcp[3], "):\n", sep='')
    print(x$para, digits=2, ...)
  }

  if (exists("beta", x)) {
    cat("\nPosterior draws of regression coefficients (thinning = ", x$mcp[3], "):\n", sep='')
    print(x$beta, digits=2, ...)
  }

  if (exists("latent", x)) {
    cat("\nPosterior draws of initial and contemporaneous latents (thinning = ", x$mcl[3], "):\n", sep='')
    print(x$latent, digits=2, ...)
  }

  if (exists("resampled", x)) {
    cat("\nPosterior draws from the auxiliary SV model were re-sampled after the MCMC\nprocedure ended to correct for model mis-specification.\n")
    print_resampling <- function (entropy_list, text) {
      if (entropy_list$entropy >= entropy_list$max_entropy*0.99) {
        cat("  Re-sampling of ", text, " was of little practical importance:\n", sep = "")
      } else {
        cat("  Re-sampling of ", text, " might have made practical a difference:\n", sep = "")
      }
      cat("    - max reachable entropy for this sample size: ", entropy_list$max_entropy, ",\n    - entropy of the re-sampling weight distribution: ", entropy_list$entropy, "\n", sep = "")
    }
    if (x$resampled$same_resampling) {
      print_resampling(x$resampled$para, "parameters and latents")
    } else {
      print_resampling(x$resampled$para, "parameters")
      print_resampling(x$resampled$latent, "latents")
    }
  }
  cat("\n")
  invisible(x)
}

#' @export
residuals.svdraws <- function(object, type = "mean", ...) {
  if (!inherits(object, "svdraws")) stop("This function expects an 'svdraws' object.")
  if (coda::nchain(para(object, "all")) > 1) stop("Multiple chains in the svdraws object: not yet implemented")
  if (!type %in% c("mean", "median")) stop("Argument 'type' must currently be either 'mean' or 'median'.")

  if (thinning(object)$time != 'all') stop("Not every point in time has been stored ('keeptime' was not set to 'all' during sampling), thus residuals cannot be extracted.")

  y <- as.vector(object$y)
  if (contains_beta(object)) {
    y <- y - object$designmatrix %*% t(svbeta(object))
  }

  if (type == "mean") {
    res <- rowMeans(y[seq(1, length(y))] / exp(t(latent(object)) / 2))
  }

  if (type == "median") {
    res <- apply(y[seq(1, length(y))] / exp(t(latent(object)) / 2), 1, median)
  }

  names(res) <- sub("h", "r", colnames(latent(object)))
  class(res) <- "svresid"
  attr(res, "type") <- type

  # Also return posterior mean/median of df parameter if terr = TRUE
  if ("nu" %in% sampled_parameters(object)) {
    attr(res, "nu") <- get(type)(para(object)[, "nu"])
  }

  res
}


#' Prediction of Future Returns and Log-Volatilities
#'
#' Simulates draws from the predictive density of the returns and the latent log-volatility
#' process. The same mean model is used for prediction as was used for fitting, which is
#' either a) no mean parameter, b) constant mean, c) AR(k) structure, or d) general
#' Bayesian regression. In the last case, new regressors need to be provided for prediction.
#'
#'
#' @param object \code{svdraws} or \code{svldraws} object.
#' @param steps \emph{optional} single number, coercible to integer. Denotes the number of
#' steps to forecast.
#' @param newdata \emph{only in case d) of the description} corresponds to input
#' parameter \code{designmatrix} in \code{\link{svsample}}.
#' A matrix of regressors with number of rows equal to parameter \code{steps}.
#' @param \dots currently ignored.
#' @return Returns an object of class \code{svpredict}, a list containing
#' three elements:
#' \item{vol}{\code{mcmc.list} object of simulations from the predictive density of the standard deviations \code{sd_(n+1),...,sd_(n+steps)}}
#' \item{h}{\code{mcmc.list} object of simulations from the predictive density of \code{h_(n+1),...,h_(n+steps)}}
#' \item{y}{\code{mcmc.list} object of simulations from the predictive density of \code{y_(n+1),...,y_(n+steps)}}
#' @note You can use the resulting object within \code{\link{plot.svdraws}} (see example below), or use
#' the list items in the usual \code{coda} methods for \code{mcmc} objects to
#' print, plot, or summarize the predictions.
#' @seealso \code{\link{plot.svdraws}}, \code{\link{volplot}}.
#' @keywords ts
#' @example inst/examples/predict.R
#' @export
predict.svdraws <- function(object, steps = 1L, newdata = NULL, ...) {
  if (!(inherits(object, "svdraws"))) stop("Argument 'object' must be of class 'svdraws'.")

  if (isTRUE(object$svlm)) {
    terms <- stats::delete.response(object$model_terms)
    m <- stats::model.frame(terms, data = newdata, xlev = object$xlevels)
    newdata <- stats::model.matrix(terms, m)
  }

  steps <- as.integer(steps)
  if (steps < 1) stop("Argument 'steps' must be greater or equal to 1.")
  # Error checking for the mean model
  arorder <- 0  # AR(0) means either a constant mean or a no mean model here
  if (object$meanmodel == "none") {
    if (!is.null(newdata)) warning("No regression coefficients were used when estimating the model. Omitting 'newdata'.")
    regressors <- function (y, newdata, stepind) 0
  } else if (object$meanmodel == "matrix") {
    if (is.null(newdata)) stop("Regressors needed for prediction. Please provide regressors through parameter 'newdata'.")
    newdata <- as.matrix(newdata)
    if (is.null(object$beta) || coda::nvar(object$beta[[1]]) != NCOL(newdata)) {
      if (isTRUE(object$svlm)) stop("Mismatch of covariates. The formula interface of 'svlm' was used for fitting the data. Now at prediction, please use the same structure of data for argument 'newdata' as the one used for argument 'data' when calling 'svlm'.")
      else stop(paste0("The number of fitted regression coefficients (", coda::nvar(object$beta[[1]]), ") does not equal the number of given regressors (", NCOL(newdata), ")."))
    }
    if (NROW(newdata) != steps) {
      warning("The size of the design matrix (", NROW(newdata), " rows) does not match the number of steps to predict (", steps, "). Argument 'steps' is ignored.")
      steps <- NROW(newdata)
    }
    regressors <- function (y, newdata, stepind) matrix(newdata[stepind, ], nrow = NROW(y), ncol = NCOL(newdata), byrow = TRUE)  # matches the format in the ar* case
  } else if (object$meanmodel == "constant") {
    if (!is.null(newdata)) warning("Constant mean was assumed when estimating the model. Omitting 'newdata'.")
    regressors <- function (y, newdata, stepind) 1
  } else if (any(grep("ar[1-9][0-9]*", object$meanmodel))) {
    arorder <- as.integer(gsub("ar", "", object$meanmodel))
    if (!is.null(newdata)) warning(paste0("An AR(", arorder, ") mean was assumed when estimating the model. Omitting 'newdata'."))
    regressors <- function (y, newdata, stepind) cbind(1, y[,seq.int(stepind, stepind-1+arorder),drop=FALSE])
  } else {
    stop("Unknown mean model. Please contact the developer.")
  }

  thinlatent <- object$thinning$latent
  thinpara <- object$thinning$para
  if (thinpara != thinlatent) {
    warning("Thinning of parameters is different from thinning of latent variables. Trying to sort this out.")  # TODO use lowest common multiple
    if (thinpara %% thinlatent == 0) {
      usepara <- seq_len(NROW(para(object, 1)))
      uselatent <- seq(thinpara %/% thinlatent, NROW(latent(object, 1)), by = thinpara %/% thinlatent)
    } else if (thinlatent %% thinpara == 0) {
      uselatent <- seq_len(NROW(latent(object, 1)))
      usepara <- seq(thinlatent %/% thinpara, NROW(para(object, 1)), by = thinlatent %/% thinpara)
    } else stop("Incompatible thinning parameters. Prediction currently not implemented.")
  } else {
    usepara <- uselatent <- seq.int(NROW(para(object, 1)))
  }

  ret <- list(y = list(), h = list(), vol = list())
  for (chain in seq_len(coda::nchain(para(object, "all")))) {
    params <- para(object, chain)
    mu <- params[usepara,"mu"]
    phi <- params[usepara,"phi"]
    sigma <- params[usepara,"sigma"]
    rho <- params[usepara,"rho"]
    nu <- params[usepara,"nu"]
    heavy_tailed <- is.finite(tail(nu, 1))
    hlast <- latent(object, chain)[uselatent,NCOL(latent(object, chain))]
    taulast <- if (heavy_tailed) svtau(object, chain)[uselatent,NCOL(svtau(object, chain))] else 1
    ylast <- as.vector(tail(object$y, 1))

    mythin <- max(thinpara, thinlatent)
    len <- length(usepara)
    volpred <- matrix(as.numeric(NA), nrow=len, ncol=steps)
    hpred <- matrix(as.numeric(NA), nrow=len, ncol=steps)
    ypred <- matrix(as.numeric(NA), nrow=len, ncol=steps+arorder)

    ypred[,seq_len(arorder)] <- matrix(as.vector(tail(object$y, arorder)), nrow=len, ncol=arorder, byrow=TRUE)  # AR(x) helper columns

    betacoeff <- if (contains_beta(object)) {
      if (arorder > 0) {
        svbeta(object, chain)[usepara,
                              c(1, rev(seq_len(NCOL(svbeta(object, chain))-1))+1),
                              drop=FALSE]
      } else {
        svbeta(object, chain)[usepara, , drop=FALSE]
      }
    } else {
      matrix(0)
    }

    resilast <- if (object$meanmodel == "none") {  # last fitted residual
      ylast*exp(-hlast/2)/sqrt(taulast)
    } else {  # if mean regression
      (ylast - colSums(object$designmatrix[NROW(object$designmatrix),]*t(betacoeff)))*exp(-hlast/2)/sqrt(taulast)  # recycles the last row of the design matrix
    }

    hpred[,1] <- mu+phi*(hlast-mu) + sigma*(rho*resilast + sqrt(1-rho^2)*rnorm(len))
    tau <- if (heavy_tailed) 1/rgamma(len, shape=.5*nu, rate=.5*(nu-2)) else 1
    volpred[,1] <- sqrt(tau) * exp(hpred[,1]/2)
    if (steps > 1) {
      for (i in seq.int(from=2, to=steps)) {
        resi <- rnorm(len)
        incr <- rho*resi + sqrt(1-rho^2)*rnorm(len)
        ypred[,i-1+arorder] <- rowSums(regressors(ypred, newdata, i-1) * betacoeff) + volpred[,i-1]*resi
        hpred[,i] <- mu + phi*(hpred[,i-1] - mu) + sigma*incr
        tau <- if (heavy_tailed) 1/rgamma(len, shape=.5*nu, rate=.5*(nu-2)) else 1
        volpred[,i] <- sqrt(tau) * exp(hpred[,i]/2)
      }
    }
    ypred[,steps+arorder] <- rowSums(regressors(ypred, newdata, steps) * betacoeff) + volpred[,steps]*rnorm(len)

    ypred <- ypred[, setdiff(seq_len(NCOL(ypred)), seq_len(arorder)), drop=FALSE]  # remove temporary AR(x) helper columns
    lastname <- tail(coda::varnames(object$latent), 1)
    lastnumber <- as.integer(gsub("h_", "", lastname))
    colnames(volpred) <- paste0("sd_", seq(lastnumber + 1, lastnumber + steps))
    colnames(hpred) <- paste0("h_", seq(lastnumber + 1, lastnumber + steps))
    colnames(ypred) <- paste0("y_", seq(lastnumber + 1, lastnumber + steps))
    ret$vol[[chain]] <- coda::mcmc(volpred, start=mythin, end=len*mythin, thin=mythin)
    ret$h[[chain]] <- coda::mcmc(hpred, start=mythin, end=len*mythin, thin=mythin)
    ret$y[[chain]] <- coda::mcmc(ypred, start=mythin, end=len*mythin, thin=mythin)
  }

  ret$vol <- coda::mcmc.list(ret$vol)
  ret$h <- coda::mcmc.list(ret$h)
  ret$y <- coda::mcmc.list(ret$y)
  class(ret) <- c("svpredict")
  ret
}

#' @export
print.svpredict <- function (x, ...) {
  cat("'svpredict' object containing predicted values for variables:\n")
  cat("  - predicted observations:        ", paste(coda::varnames(x$y[[1]]), collapse = ", "), "\n", sep = "")
  cat("  - predicted standard deviations: ", paste(coda::varnames(x$vol[[1]]), collapse = ", "), "\n", sep = "")
  cat("  - predicted latent variables:    ", paste(coda::varnames(x$h[[1]]), collapse = ", "), "\n", sep = "")
  invisible(x)
}

#' @export
print.svdraws_roll <- function (x, ...) {
  cat("'svdraws_roll' object containing results over ", length(x), " time windows\n", sep="")
}

#' @export
summary.svdraws_roll <- function (object, ...) {
  cat("\nSummary of 'svdraws_roll' object\n\n")
  have_predlik <- !is.null(object[[1]]$predictive_likelihood)
  have_predquant <- length(object[[1]]$quantiles) > 0
  if (have_predlik) {
    cat(" Summary of predictive likelihoods\n")
    print(summary(sapply(object, function (xx) xx$predictive_likelihood)))
    cat("\n")
  }
  if (have_predquant) {
    cat(" Summary of predicted quantiles\n")
    print(summary(t(sapply(object, function (xx) xx$predicted_quantile))))
    cat("\n\n")
  }
}