Package {bayesGARCH}

Version: 2.2.0
Date: 2026-08-19
Title: Bayesian Estimation of the GARCH(1,1) Model with Student-t Innovations
Maintainer: David Ardia <david.ardia.ch@gmail.com>
Imports: coda
Suggests: testthat (≥ 3.0.0), mvtnorm
Description: Provides the bayesGARCH() function which performs the Bayesian estimation of the GARCH(1,1) model with Student's t innovations as described in Ardia (2008) <doi:10.1007/978-3-540-78657-3>.
BugReports: https://github.com/ArdiaD/bayesGARCH/issues
URL: https://github.com/ArdiaD/bayesGARCH
License: GPL-2 | GPL-3 [expanded from: GPL (≥ 2)]
Encoding: UTF-8
Config/testthat/edition: 3
NeedsCompilation: yes
Packaged: 2026-08-21 13:46:33 UTC; ardiad
Author: David Ardia ORCID iD [aut, cre, cph, fnd]
Repository: CRAN
Date/Publication: 2026-08-21 15:50:02 UTC

Bayesian Estimation of the GARCH(1,1) Model with Student-t Innovations

Description

Performs the Bayesian estimation of the GARCH(1,1) model with Student-t innovations.

Usage


  bayesGARCH(y, mu.alpha = c(0,0), Sigma.alpha = 1000 * diag(1,2), 
             mu.beta = 0, Sigma.beta = 1000,
             lambda = 0.01, delta = 2, control = list())

Arguments

y

vector of observations of size T. NA values are not allowed.

mu.alpha

hyper-parameter \mu_\alpha (prior mean) for the truncated Normal prior on parameter \alpha := (\alpha_0 \ \alpha_1)'. Default: a 2 \times 1 vector of zeros.

Sigma.alpha

hyper-parameter \Sigma_\alpha (prior covariance matrix) for the truncated Normal prior on parameter \alpha. Default: a 2 \times 2 diagonal matrix whose variances are set to 1'000, i.e., a diffuse prior. Note that the matrix must be symmetric positive definite.

mu.beta

hyper-parameter \mu_\beta (prior mean) for the truncated Normal prior on parameter \beta. Default: zero.

Sigma.beta

hyper-parameter \Sigma_\beta > 0 (prior variance) for the truncated Normal prior on parameter \beta. Default: 1'000, i.e., a diffuse prior.

lambda

hyper-parameter \lambda > 0 for the translated Exponential distribution on parameter \nu. Default: 0.01.

delta

hyper-parameter \delta \ge 2 for the translated Exponential distribution on parameter \nu. Default: 2 (to ensure the existence of the conditional variance).

control

list of control parameters (See *Details*).

Details

The function bayesGARCH performs the Bayesian estimation of the GARCH(1,1) model with Student-t innovations. The underlying algorithm is based on Nakatsuma (1998, 2000) for generating the parameters of the GARCH(1,1) scedastic function \alpha := (\alpha_0 \ \alpha_1)' and \beta and on Geweke (1993) and Deschamps (2006) for the generating the degrees of freedom parameter \nu. Further details and examples can be found in Ardia (2008) and Ardia and Hoogerheide (2010). Finally, we refer to Ardia (2009) for an extension of the algorithm to Markov-switching GARCH models.

The series is rescaled internally to unit root mean square before sampling and the draws are mapped back, so that results do not depend on the units of y. Only \alpha_0 carries that scale: under y \to y/s one has \alpha_0 \to \alpha_0/s^2 while \alpha_1, \beta and \nu are unchanged, and the Normal prior on \alpha transforms exactly under that map, so the posterior explored is the same one. The factor is a power of two, which makes the rescaling and its inverse exact in binary floating point. Note that the prior itself is stated in the units of y; rescaling y without rescaling Sigma.alpha changes the posterior, because a prior that is diffuse on one scale is informative on another.

The control argument is a list that can supply any of the following components:

n.chain

number of MCMC chain(s) to be generated. Default: n.chain=1. When start.val is a matrix holding one row per chain, n.chain may be omitted, in which case it is taken to be the number of rows.

l.chain

length of each MCMC chain. Default: l.chain=10000.

start.val

vector of starting values of chain(s). Default: c(0.01, 0.1, 0.7, delta + 1/lambda), whose last entry is the prior mean of \nu and therefore lies in the prior support for any delta, and whose first entry applies to the internally rescaled series (see *Details*), making it independent of the units of y. A value supplied by the user is read in the units of y. A matrix of size n \times 4 containing starting values in rows can also be provided. This will generate n chains starting at the different row values. A vector, or a matrix with a single row, is replicated across the n.chain chains.

Every row must lie in the support of the posterior: \alpha_0, \alpha_1, \beta > 0, \nu > \delta, and, when addPriorConditions is used, it must satisfy that condition as well. Rows that do not are rejected with an error naming the row and the condition it fails. See *Note*.

addPriorConditions

function which allows the user to add constraints on the model parameters. Default: NULL, i.e. not additional constraints are imposed (see below). It must return a single non-missing TRUE or FALSE for every parameter vector it is given, both for the starting values and for the states proposed during sampling.

refresh

frequency of reports. Default: refresh=10 iterations.

digits

number of printed digits in the reports. Default: digits=4.

Value

A list of class mcmc.list (R package coda), with one mcmc object per chain.

Each chain carries a "move.rates" attribute holding, for every parameter, the proportion of iterations at which its value changed. This is a diagnostic only: a high rate is not by itself evidence of good mixing, since the accepted moves are typically small. Use effectiveSize to judge how many independent draws the chain actually delivers.

Note

By using bayesGARCH you agree to the following rules:

The GARCH(1,1) model with Student-t innovations may be written as follows:

y_t = \epsilon_t (\varrho h_t)^{1/2}

for t=1,\ldots,T, where the conditional variance equation is defined as:

h_t := \alpha_0 + \alpha_1 y_{t-1}^2 + \beta h_{t-1}

where \alpha_0 > 0, \alpha_1 \ge 0, \beta \ge 0 to ensure a positive conditional variance. We set the initial variance and the initial observation to h_0 := 0 and y_0 := 0 for convenience, so that h_1 = \alpha_0. The parameter \varrho := (\nu-2)/\nu is a scaling factor which ensures the conditional variance of y_t to be h_t. Finally, \epsilon_t follows a Student-t distribution with \nu degrees of freedom.

The prior distributions on \alpha is a bivariate truncated Normal distribution:

p(\alpha) \propto N_2(\alpha \mid \mu_\alpha, \Sigma_\alpha) I_{[\alpha > 0]}

where \mu_\alpha is the prior mean vector, \Sigma_\alpha is the prior covariance matrix and I_{[\bullet]} is the indicator function.

The prior distribution on \beta is a univariate truncated Normal distribution:

p(\beta) \propto N(\beta \mid \mu_\beta, \Sigma_\beta) I_{[\beta > 0]}

where \mu_\beta is the prior mean and \Sigma_\beta is the prior variance.

The prior distribution on \nu is a translated Exponential distribution:

p(\nu) = \lambda \exp [ - \lambda (\nu - \delta) ] I_{[\nu > \delta]}

where \lambda > 0 and \delta \ge 2. The prior mean for \nu is \delta + 1/\lambda.

The joint prior on parameter \psi := (\alpha, \beta, \nu) is obtained by assuming prior independence:

p(\psi) = p(\alpha) p(\beta) p(\nu).

The default hyperparameters \mu_\alpha, \Sigma_\alpha, \mu_\beta, \Sigma_\beta and \lambda define a rather vague prior. The hyper-parameter \delta \geq 2 ensures the existence of the conditional variance. The kth conditional moment for \epsilon_t is guaranteed by setting \delta \geq k.

The Bayesian estimation of the GARCH(1,1) model with Normal innovations is obtained as a special case by setting lambda=100 and delta=500. In this case, the generated values for \nu are centered around 500 which ensure approximate Normality for the innovations.

The function addPriorConditions allows to add prior conditions on the model parameters \psi := (\alpha_0 \ \alpha_1 \ \beta \ \nu)'. The function must return TRUE if the constraint holds and FALSE otherwise.

By default, the function is: 

    addPriorConditions <- function(psi)
    {
      TRUE
    }

and therefore does not add any other constraint than the positivity of the parameters which are obtained through the prior distribution for \psi.

You simply need to modify addPriorConditions in order to add constraints on the model parameters \psi. For instance, to impose the covariance-stationary conditions to hold, i.e. \alpha_1 + \beta < 1, just define the function addPriorConditions as follows: 

    addPriorConditions <- function(psi)
    {
      psi[2] + psi[3] < 1
    }

Note that adding prior constraints on the model parameters can diminish the acceptance rate and therefore lead to a very inefficient sampler. This would however indicate that the condition is not supported by the data.

The chain must start in the constrained region. Every step only ever accepts a move into it, so a chain started outside can never enter: the starting value would be returned unchanged at every iteration and the whole sample would consist of draws the posterior gives zero mass. This is easy to trigger by accident, since the advice below to start from the parameter values of a previous estimation step will supply an unconstrained estimate that need not satisfy the constraint. Starting values are therefore validated against addPriorConditions before any chain is generated.

The constraint enters the acceptance test of each Metropolis-Hastings step, so that the sampler leaves the constrained posterior p(\psi \mid y) I_{[\psi \in A]} invariant. Up to version 2.1.10 the constraint was instead applied once to the whole block of parameters, accepting or rejecting the entire sweep on the final value of \psi. That scheme does not preserve the constrained posterior — the block is a composition of reversible steps, which is invariant but not itself reversible — and the chain converged to a different distribution. Results obtained with addPriorConditions under earlier versions should be re-computed; see the ‘NEWS’ file.

The estimation strategy implemented in bayesGARCH is fully automatic and does not require any tuning of the MCMC sampler. The cost per iteration is modest on current hardware, but the draws are strongly autocorrelated, so long chains are needed: on the dem2gbp series (T = 1974) the default l.chain=10000 delivers an effective sample size of only a few hundred draws per parameter. Always inspect the output before using it — effectiveSize and gelman.diag (R package coda) apply directly to the value returned by bayesGARCH — and raise l.chain until the effective sample size is adequate for the quantity of interest. Estimating the model over several datasets on a daily basis can therefore take a significant amount of time. In this case, the algorithm can be easily parallelized, by running a single chain on several processors. Also, when the estimation is repeated over updated time series (i.e. time series with more recent observations), it is wise to start the algorithm using the posterior mean or median of the parameters obtained at the previous estimation step. The impact of the starting values (burn-in phase) is likely to be smaller and thus the convergence faster.

Finally, note that as any MH algorithm, the sampler can get stuck to a given value, so that the chain does not move anymore. However, the sampler uses Taylor-made candidate densities that are especially ‘constructed’ at each step, so it is almost impossible for this MCMC sampler to get stuck at a given value for many subsequent draws. A high proportion of accepted moves should not be read as good mixing, however: the accepted moves are typically small, and the resulting chains are strongly autocorrelated, as noted above. In the unlikely case that such ill behavior would occur, one could run the algorithm with different initial values or a different random seed. Rescaling the data is no longer necessary: it is done internally, as described in *Details*. Before version 2.2.0 a series whose scale was far from unity — log-returns in decimals rather than in percent was already enough — could have every \alpha and \beta move rejected, leaving a chain that never moved and eventually failed in the \nu update.

Author(s)

David Ardia david.ardia.ch@gmail.com

References

Ardia, D. (2009) Bayesian Estimation of a Markov-Switching Threshold Asymmetric GARCH Model with Student-t Innovations. Econometrics Journal 12(1), pp. 105-126. doi:10.1111/j.1368-423X.2008.00253.x

Ardia, D., Hoogerheide, L.F. (2010) Bayesian Estimation of the GARCH(1,1) Model with Student-t Innovations. R Journal 2(2), pp.41-47. doi:10.32614/RJ-2010-014

Ardia, D. (2008) Financial Risk Management with Bayesian Estimation of GARCH Models. Lecture Notes in Economics and Mathematical Systems 612. Springer-Verlag, Berlin, Germany. ISBN 978-3-540-78656-6, e-ISBN 978-3-540-78657-3, doi:10.1007/978-3-540-78657-3

Deschamps, P.J. (2006) A Flexible Prior Distribution for Markov Switching Autoregressions with Student-t Errors. Journal of Econometrics 133, pp.153-190.

Geweke, J.F. (1993) Bayesian Treatment of the Independent Student-t Linear Model. Journal of Applied Econometrics 8, pp.19-40.

Nakatsuma, T. (2000) Bayesian Analysis of ARMA-GARCH Models: A Markov Chain Sampling Approach. Journal of Econometrics 95(1), pp.57-69.

Nakatsuma, T. (1998) A Markov-Chain Sampling Algorithm for GARCH Models. Studies in Nonlinear Dynamics and Econometrics 3(2), pp.107-117.

See Also

garchFit (R package fGarch) for the classical Maximum Likelihood estimation of GARCH models.

Examples


  ## !!! INCREASE THE NUMBER OF MCMC ITERATIONS !!!

  ## LOAD DATA
  data(dem2gbp)
  y <- dem2gbp[1:750]

  ## RUN THE SAMPLER (2 chains)
  MCMC <- bayesGARCH(y, control = list(n.chain = 2, l.chain = 200))

  ## MCMC ANALYSIS (using coda)
  plot(MCMC)
  
  ## FORM THE POSTERIOR SAMPLE
  smpl <- formSmpl(MCMC, l.bi = 50)

  ## POSTERIOR STATISTICS
  summary(smpl)
  smpl <- as.matrix(smpl)
  pairs(smpl)

  ## GARCH(1,1) WITH NORMAL INNOVATIONS
  MCMC <- bayesGARCH(y, lambda = 100, delta = 500,
                     control = list(n.chain = 2, l.chain = 200))

  ## GARCH(1,1) WITH NORMAL INNOVATIONS AND 
  ## WITH COVARIANCE STATIONARITY CONDITION
  addPriorConditions <- function(psi){psi[2] + psi[3] < 1}
  MCMC <- bayesGARCH(y, lambda = 100, delta = 500,
                     control = list(n.chain = 2, l.chain = 200, 
                     addPriorConditions = addPriorConditions))

DEM/GBP exchange rate log-returns

Description

The vector dem2gbp contains daily observations of the Deutschmark vs British Pound foreign exchange rate log-returns. This data set has been promoted as an informal benchmark for GARCH time-series software validation. See McCullough and Renfro (1999), and Brooks, Burke, and Persand (2001) for details. The nominal returns are expressed in percent as in Bollerslev and Ghysels (1996). The sample period is from January 3, 1984, to December 31, 1991, for a total of 1974 observations.

Usage

data(dem2gbp)

Format

A vector of size 1974.

Source

Journal of Business and Economic Statistics

References

Bollerslev T., Ghysels, E. (1996) Periodic Autoregressive Conditional Heteroscedasticity. Journal of Business and Economic Statistics 14(2), pp.139–151.

Brooks C., Burke S. P., Persand G. (2001) Benchmarks and the Accuracy of GARCH Model Estimation. International Journal of Forecasting 17(1), pp.45–57.

McCullough B. D., Renfro C. G. (1999) Benchmarks and Software Standards: A Case Study of GARCH Procedures. Journal of Economic and Social Measurement 25(2), pp.59–71.

Form the Posterior Sample

Description

Form the joint posterior sampler from the MCMC output.

Usage


  formSmpl(MCMC, l.bi = 0, batch.size = 1)

Arguments

MCMC

object of the class mcmc.list (R package coda) or a list of matrices or a matrix. When several chains are given they must share the same number of rows and columns and the same column names.

l.bi

length of the burn-in phase, a whole number smaller than the chain length.

batch.size

batching size used to diminish the autocorrelation within the chains, a whole number greater or equal than one.

Value

The joint posterior sample as an mcmc object (R package coda).

Note

Please cite the package in publications. Use citation("bayesGARCH").

See Also

bayesGARCH for the Bayesian estimation of the GARCH(1,1) model with Student-t innovations.

Examples

  ## !!! INCREASE THE NUMBER OF MCMC ITERATIONS !!!

  ## LOAD DATA SET
  data(dem2gbp)
  y <- dem2gbp[1:750]

  ## RUN THE ESTIMATION
  MCMC <- bayesGARCH(y, control = list(n.chain = 2, l.chain = 100))

  ## FORM THE SAMPLE FROM THE MCMC OUTPUT
  smpl <- formSmpl(MCMC, l.bi = 50, batch.size = 2)

  ## POSTERIOR STATISTICS
  summary(smpl)