JSS Journal of Statistical Software

October 2019, Volume 91, Issue 4. doi: 10.18637/jss.v091.i04

Markov-Switching GARCH Models in R:

The MSGARCH Package

David Ardia Keven Bluteau Kris Boudt

HEC Montréal University of Neuchâtel Ghent University
Vrije Universiteit Brussel Vrije Universiteit Brussel
Vrije Universiteit Amsterdam

Leopoldo Catania Denis-Alexandre Trottier

Aarhus University Laval University
CREATES

Abstract
We describe the package MSGARCH, which implements Markov-switching GARCH
(generalized autoregressive conditional heteroscedasticity) models in R with efficient C++
object-oriented programming. Markov-switching GARCH models have become popular
methods to account for regime changes in the conditional variance dynamics of time se-
ries. The package MSGARCH allows the user to perform simulations as well as maximum
likelihood and Bayesian Markov chain Monte Carlo estimations of a very large class of
Markov-switching GARCH-type models. The package also provides methods to make
single-step and multi-step ahead forecasts of the complete conditional density of the vari-
able of interest. Risk management tools to estimate conditional volatility, value-at-risk,
and expected-shortfall are also available. We illustrate the broad functionality of the
MSGARCH package using exchange rate and stock market return data.

Keywords: GARCH, MSGARCH, Markov-switching, conditional volatility, forecasting, R soft-

ware.

1. Introduction
In 2003, Robert Engle received the Nobel Prize for his “contribution to methods of analyzing
economic time series with time-varying volatility (ARCH)” (Nobel Media 2003). The seminal
paper introducing the orginal autoregressive conditional heteroscedasticity (ARCH) model is
2 MSGARCH: Markov-Switching GARCH Models in R

Engle (1982), while its generalization to GARCH was introduced by Bollerslev (1986). Since
then, multiple extensions of the GARCH scedastic function have been proposed to capture
additional stylized facts observed in financial and economic time series, such as nonlineari-
ties, asymmetries, and long-memory properties; see Teräsvirta (2009) for a review. According
to the Time Series Analysis (Hyndman 2019) and Empirical Finance (Eddelbuettel 2019)
task views at https://CRAN.R-project.org/web/views, the following implementations of
univariate GARCH-type models are available in the R (R Core Team 2018) programming
language: bayesGARCH (Ardia and Hoogerheide 2010), fGarch (Wuertz, Chalabi, Miklovic,
Boudt, and Chausse 2016), GAS (Ardia, Boudt, and Catania 2019b), gets (Pretis, Reade, and
Sucarrat 2018), GEVStableGarch (Sousa, Otiniano, Lopes, and Diethelm 2015), lgarch (Su-
carrat 2015), rugarch (Ghalanos 2017) and tseries (Trapletti and Hornik 2017). In GARCH-
type models, the conditional volatility is driven by shocks in the observed time series. An
alternative approach is to assume that volatility is driven by volatility-specific shocks. This
is the case in stochastic volatility models, as available in the R package stochvol (Kastner
2016).
Recent studies show that volatility predictions by GARCH-type models may fail to capture the
true variation in volatility in the case of regime changes in the volatility dynamics (see, e.g.,
Lamoureux and Lastrapes 1990; Bauwens, Backer, and Dufays 2014). A solution to this
problem is to allow the parameters of the GARCH model to vary over time according to
a latent discrete Markov process. This approach is called the Markov-switching GARCH
(MSGARCH) model, which leads to volatility forecasts that can quickly adapt to variations
in the unconditional volatility level.
MSGARCH models are mainly used in finance. Other fields of applications include the anal-
ysis of business cycles in economics (see, e.g., Kim and Nelson 1999) and the forecasting of
windpower (see, e.g., Zhang, Wang, and Wang 2014). Their use is not widespread in industry
applications mainly because their implementation is tedious and their estimation is not trivial.
We fill this gap with the R package MSGARCH (Ardia, Bluteau, Boudt, Catania, Ghalanos,
Peterson, and Trottier 2019a), and provide researchers and practitioners with a ready-to-use
program within the R environment. The R package MSGARCH implements the specifica-
tion of Haas, Mittnik, and Paolella (2004a), which, compared to the MSGARCH model in
the seminal work of Gray (1996), has the advantage that the estimation does not face the
“path-dependency problem”.1 In their model, the authors define K separate GARCH-type
processes, one for each regime of the unobserved Markov chain. In addition to its appealing
computational aspects, the MSGARCH model of Haas et al. (2004a) has conceptual advan-
tages. Indeed, one reason for specifying Markov-switching models that allow for different
GARCH behavior in each regime is to capture the difference in the variance dynamics of low-
and high-volatility periods.
The R package MSGARCH provides a comprehensive set of methods for estimating, sim-
ulating, and forecasting with MSGARCH models. It includes the possibility of specifying
different GARCH processes and conditional distributions for each state. Evaluation of risk
management metrics such as the value-at-risk and the expected-shortfall is also available. The
1
Literature on computationally feasible implementations of MSGARCH models started with the seminal
work of Gray (1996), Dueker (1997), and Klaassen (2002), who proposed different solutions to the “path-
dependency problem” raised by Cai (1994) and Hamilton and Susmel (1994). It means that, for a sample of
size T and a MSGARCH model with K regimes, the evaluation of K T different volatility paths is required,
rendering the estimation intractable for medium and large sample sizes.
 Journal of Statistical Software 3

R package MSGARCH relies on object-oriented programming techniques in C++ via the R

packages Rcpp (Eddelbuettel, François, Allaire, Ushey, Kou, Bates, and Chambers 2018; Ed-
delbuettel and François 2011) and RcppArmadillo (Eddelbuettel, François, and Bates 2017;
Eddelbuettel and Sanderson 2014). Efficient implementation is key, especially when backtest-
ing models in risk management.
The package’s usage and the available methods are similar to those of the R package rugarch
(Ghalanos 2017). First, the user creates a model specification. Second, she/he estimates
the model. Third, she/he can perform predictions and risk forecasts. Plotting, summary
and model discrimination methods can be applied to the fitted model. The R package MS-
GARCH is available from the Comprehensive R Archive Network (CRAN) repository at
https://CRAN.R-project.org/package=MSGARCH and the development version is available
on the MSGARCH website at http://keblu.github.io/MSGARCH/.
The outline of the paper is as follows. Section 2 introduces the models. Section 3 describes
the R implementation and serves as a brief user manual. Section 4 illustrates the package’s
usage on an application using Swiss market index data. Section 5 concludes.

2. Markov-switching GARCH models

Denote the variable of interest at time t by yt . We assume that yt has zero mean and is
not serially correlated, that is, the following moment conditions are assumed: E[yt ] = 0 and
E[yt yt−l ] = 0 for l 6= 0 and all t > 0. This assumption is realistic for high frequency returns
for which the (conditional) mean is often assumed to be zero; see, for instance, McNeil, Frey,
and Embrechts (2015). In other applications, the assumption that the MSGARCH process
has conditional mean zero often implies applying the MSGARCH model to a de-meaned time
series.2 In the most simple case of a constant mean, this is the series in excess of the sample
mean. When the series has dynamics in the conditional mean, the de-meaned time series are
the residuals of a time series model, like an ARFIMAX, as in the R package rugarch (Ghalanos
2017). Since the MSGARCH package decouples the mean and volatility estimation, GARCH-
in-mean models and joint approaches to mean-variance switching as in Kim and Nelson (1999)
are not possible.
We allow for regime-switching in the conditional variance process. Denote by It−1 the informa-
tion set observed up to time t − 1, that is, It−1 ≡ {yt−i , i > 0}. The general Markov-switching
GARCH specification can then be expressed as:
yt | (st = k, It−1 ) ∼ D(0, hk,t , ξ k ) , (1)
where D(0, hk,t , ξ k ) is a continuous distribution with zero mean, time-varying variance hk,t ,
and additional shape parameters gathered in the vector ξ k .3 The integer-valued stochastic
variable st , defined on the discrete space {1, . . . , K}, characterizes the Markov-switching
1/2 iid
GARCH model. We define the standardized innovations as ηk,t ≡ yt /hk,t ∼ D(0, 1, ξ k ).
2
Furthermore, assuming a non zero state dependent mean would generally imply E[yt yt−1 ] 6= 0, which is
against the model we specified.
3
As explained below, the parametric formulation of the conditional distribution D(0, hk,t , ξ k ) can be different
across regimes. In this case, the notation Dk (0, hk,t , ξ k ) would be more appropriate. The same applies for the
h(·) function in (2). We keep the simpler notation to improve readability. Also, for t = 1, we initialize the
regime probabilities and the conditional variances at their unconditional levels. To simplify exposition, we use
henceforth for t = 1 the same notation as for general t, since there is no confusion possible.
4 MSGARCH: Markov-Switching GARCH Models in R

2.1. State dynamics

The R package MSGARCH package implements two approaches to the dynamics of the
state variable, namely the assumption of a first-order ergodic homogeneous Markov chain
which characterizes the Markov-switching GARCH model of Haas et al. (2004a), and the
assumption of independent draws from a multinomial distribution which characterizes the
mixture of GARCH models of Haas, Mittnik, and Paolella (2004b).

First-order Markov chain

We assume that st evolves according to an unobserved first-order ergodic homogeneous
Markov chain with K × K transition probability matrix P:

p1,1 ... p1,K

 
 .. .. ..  ,
P≡ . . . 
pK,1 ... pK,K

where pi,j ≡ P[st = j | st−1 = i] is the probability of a transition from state st−1 = i to
state st = j. Obviously, the following constraints hold: 0 < pi,j < 1 ∀i, j ∈ {1, . . . , K},
and K 2
j=1 pi,j = 1, ∀i ∈ {1, . . . , K}. Given the parametrization of D(·), we have E[yt | st =
P

k, It−1 , ] = hk,t , that is, hk,t is the variance of yt conditional on the realization of st = k.
In the MSGARCH model of Haas et al. (2004a), the conditional variances hk,t for k = 1, . . . , K
are assumed to follow K separate GARCH-type processes which evolve in parallel.

Independent states
A related specification has been also introduced by Haas et al. (2004b) and is referred to
as the mixture of GARCH. In this case, we assume that st is sampled independently over
time from a Multinomial distribution with support {1, . . . , K} and vector of probabilities
ω = (ω1 , . . . , ωK )> , that is P[st = k] = ωk . We have that the same parametric formulation
of (1) with K separate GARCH-type models defined for each component of the mixture.4

2.2. Conditional variance dynamics

As in Haas et al. (2004a), the conditional variance of yt is assumed to follow a GARCH-
type model. Hence, conditionally on regime st = k, hk,t is available as a function of the
past observation, yt−1 , past variance hk,t−1 , and the additional regime-dependent vector of
parameters θ k :
hk,t ≡ h(yt−1 , hk,t−1 , θ k ) , (2)
where h(·) is a It−1 -measurable function that defines the filter for the conditional variance
and also ensures its positiveness. In the MSGARCH package, the initial value of the variance
recursions, that is hk,1 (k = 1, . . . , K), are set equal to the unconditional variance in regime k.
Depending on the form of h(·), we obtain different scedastic specifications. In the R package
MSGARCH, we follow this specification in order to reduce model complexity. Finally, when
K = 1, we recover single-regime GARCH-type models identified by the form of h(·).
4
The mixture of GARCH model presented in Haas et al. (2004b) allows for interactions between the mixture
component variances. Here, we report the case referred to as “Diagonal” by Haas et al. (2004b).
 Journal of Statistical Software 5

Below we briefly present the scedastic specifications available in the R package MSGARCH.
Each of them is identified with a label used in the code for defining a model specification.
Similarly, model coefficients are also identified with labels.

ARCH model
The ARCH model of Engle (1982) is given by:
2
hk,t ≡ α0,k + α1,k yt−1 ,

for k = 1, . . . , K. In this case, we have θ k = (α0,k , α1,k )> . To ensure positivity, we require
that α0,k > 0 and α1,k ≥ 0. Covariance-stationarity in each regime is obtained by requiring
that α1,k < 1. The ARCH specification is identified with the label "sARCH".

GARCH model
The GARCH model of Bollerslev (1986) is given by:
2
hk,t ≡ α0,k + α1,k yt−1 + βk hk,t−1 ,

for k = 1, . . . , K. In this case, we have θ k = (α0,k , α1,k , βk )> . To ensure positivity, we require
that α0,k > 0, α1,k > 0, and βk ≥ 0. Covariance-stationarity in each regime is obtained by
requiring that α1,k + βk < 1. The GARCH specification is identified with the label "sGARCH".

EGARCH model
The Exponential GARCH (EGARCH) of Nelson (1991) is given by:


ln(hk,t ) ≡ α0,k + α1,k |ηk,t−1 | − E [|ηk,t−1 |] + α2,k ηk,t−1 + βk ln(hk,t−1 ) ,

for k = 1, . . . , K, where the expectation E[|ηk,t−1 |] is taken with respect to the distribution
conditional on regime k. In this case, we have θ k = (α0,k , α1,k , α2,k , βk )> . This specification
takes into consideration the so-called leverage effect where past negative observations have
a larger influence on the conditional volatility than past positive observations of the same
magnitude (Black 1976; Christie 1982). Positivity is automatically ensured by the model
specification. Covariance-stationarity in each regime is obtained by requiring that βk < 1.
The EGARCH specification is identified with the label "eGARCH".

GJR model
The GJR model of Glosten, Jagannathan, and Runkle (1993) is also able to capture the
asymmetry in the conditional volatility process. This model is given by:
2
hk,t ≡ α0,k + (α1,k + α2,k I{yt−1 < 0})yt−1 + βk hk,t−1 ,

for k = 1, . . . , K, where I{·} is the indicator function taking value one if the condition holds,
and zero otherwise. In this case, we have θ k = (α0,k , α1,k , α2,k , βk )> . The parameter α2,k
controls the degree of asymmetry in the conditional volatility response to the past shock
in regime k. To ensure positivity, we require that α0,k > 0, α1,k > 0, α2,k ≥ 0, βk ≥ 0.
Covariance-stationarity in each regime is obtained by requiring that α1,k + α2,k E[ηk,t 2 I{η
k,t <
0}] + βk < 1. The GJR specification is identified with the label "gjrGARCH".
6 MSGARCH: Markov-Switching GARCH Models in R

TGARCH model
Zakoian (1994) introduces the TGARCH specification where the conditional volatility is the
dependent variable instead of the conditional variance. This model is given by:
1/2 1/2
hk,t ≡ α0,k + (α1,k I{yt−1 ≥ 0} − α2,k I{yt−1 < 0})yt−1 + βk hk,t−1 ,

for k = 1, . . . , K. In this case, we have θ k = (α0,k , α1,k , α2,k , βk )> . To ensure positivity,
we require that α0,k > 0, α1,k > 0, α2,k > 0 and βk ≥ 0. Covariance-stationarity in each
2 + β 2 − 2β (α 2
regime is obtained by requiring that α1,k k k 1,k + α2,k )E[ηt,k I{ηt,k < 0}] − (α1,k −
2 2
α2,k )E[ηk,t I{ηk,t < 0}] < 1 (see Francq and Zakoian 2011, Section 10.2). The TGARCH
specification is identified with the label "tGARCH".
2 I{η
The quantities E[ηt,k I{ηt,k < 0}] and E[ηt,k t,k < 0}] required for the covariance-stationarity
conditions in the GJR and TGARCH models, and the quantity E [|ηk,t−1 |] required in the
conditional variance equation of the EGARCH model, are implemented following Trottier
and Ardia (2016).

2.3. Conditional distribution

Model specification is completed by the definition of the conditional distribution of the stan-
dardized innovations ηt,k in each regime of the Markov chain. Here, we present the condi-
tional distributions available in the R package MSGARCH. The most common distributions
employed to model financial log-returns are implemented; additional distributions might be
included in a future release of the R package MSGARCH if required by the users. Each
distribution is standardized to have a zero mean and a unit variance. As for the conditional
variance specification, distributions are identified with labels. Here we drop the time and
regime indices for notational purposes, but the shape parameters can be conditional on the
regime.

Normal distribution
The probability density function (PDF) of the standard normal distribution is given by:
1 1 2
fN (η) ≡ √ e− 2 η , η ∈ R.
2π
This distribution is identified with the label "norm".

Student-t distribution
The PDF of the standardized Student-t distribution is given by:

ν+1
 !− ν+1
Γ 2 η2 2
fS (η; ν) ≡ p 1 + , η ∈ R,
(ν − 2)π Γ ν2 (ν − 2)

where Γ(·) is the Gamma function. The constraint ν > 2 is imposed to ensure that the second
order moment exists. The kurtosis of this distribution is higher for lower ν.5 This distribution
is identified with the label "std".
5
For ν = ∞, the Student-t distribution is equivalent to the normal distribution.
 Journal of Statistical Software 7

GED distribution
The PDF of the standardized generalized error distribution (GED) is given by:
1 ν 1/2
νe− 2 |η/λ| Γ(1/ν)

fGED (η; ν) ≡ (1+1/ν) , λ≡ 1/ν
, η ∈ R,
λ2 Γ(1/ν) 4 Γ(3/ν)

where ν > 0 is the shape parameter.6 This distribution is identified with the label "ged".

Skewed distributions
Fernández and Steel (1998) provide a simple way to introduce skewness into any unimodal
standardized distribution, via the additional parameter ξ > 0; if ξ = 1 the distribution turns
out to be symmetric. Trottier and Ardia (2016) derive the moments of the standardized
Fernandez-Steel skewed distributions which are needed in the estimation of the EGARCH,
GJR, and TGARCH models. We refer the reader to that publication for details. The skewed
version of the normal, Student-t, and GED distributions are identified with the labels "snorm",
"sstd", and "sged", respectively.
In Panels A and B of Table 1, we provide a summary of the scedastic functions and conditional
distributions available in the package.

2.4. Model estimation

Estimation of MSGARCH and mixture of GARCH models can be done either by maximum
likelihood (ML) or by Bayesian Markov chain Monte Carlo (MCMC) techniques. Both ap-
proaches require the evaluation of the likelihood function.
Let Ψ ≡ (θ 1 , ξ 1 , . . . , θ K , ξ K , P) be the vector of model parameters. The likelihood function
is:
T
Y
L(Ψ | IT ) ≡ f (yt | Ψ, It−1 ) , (3)
t=1

where f (yt | Ψ, It−1 ) denotes the density of yt given past observations, It−1 , and model pa-
rameters Ψ. For MSGARCH, the conditional density of yt is:
K X
X K
f (yt | Ψ, It−1 ) ≡ pi,j zi,t−1 fD (yt | st = j, Ψ, It−1 ) , (4)
i=1 j=1

where zi,t−1 ≡ P[st−1 = i | Ψ, It−1 ] represents the filtered probability of state i at time t − 1
obtained via Hamilton’s filter; see Hamilton (1989) and Hamilton (1994, Chapter 22) for
details. Similarly, for the mixture of GARCH model, the conditional density of yt is:
K
X
f (yt | Ψ, It−1 ) ≡ ωj fD (yt | st = j, Ψ, It−1 ) . (5)
j=1

In both (4) and (5), the conditional density of yt in state/component st = k given Ψ and
It−1 is denoted by fD (yt | st = k, Ψ, It−1 ).
6
Special cases of this distribution are obtained for ν = 1 (Laplace distribution) and ν = 2 (normal distri-
bution). The uniform distribution is obtained in the limit ν → ∞.
8 MSGARCH: Markov-Switching GARCH Models in R

Panel A: Conditional volatility models

Model Label Equation
ARCH "sARCH" ht ≡ α0 + α1 yt−12

GARCH "sGARCH" ht ≡ α0 + α1 yt−12 + βht−1

EGARCH "eGARCH" ln(ht ) ≡ α0 + α1 |yt−1 | − E[|yt−1 |] + α2 yt−1 + β ln(ht−1 )
"gjrGARCH" ht ≡ α0 + α1 yt−1 2 2 I{y
+ α2 yt−1
GJR t−1 < 0} + βht−1
1/2
TGARCH "tGARCH" ht ≡ α0 + α1 yt−1 I{yt−1 ≥ 0}+
1/2
α2 yt−1 I{yt−1 < 0} + βht−1
Panel B: Conditional distributions
Model Label Equation
1 2
Normal "norm" fN (η) ≡ √12π e− 2 η , η ∈ R
Γ( ν+1
− ν+1
2 ) η2

2
Student-t "std" fS (η; ν) ≡ √ 1 + (ν−2) ,η ∈ R
(ν−2)π Γ( ν2 )
− 1 |η/λ|ν
 1/2
νe 2 Γ(1/ν)
GED "ged" fGED (η; ν) ≡ λ2(1+1/ν) Γ(1/ν)
, λ ≡ 41/ν Γ(3/ν)
,η ∈ R
Skewed normal "snorm" See Trottier and Ardia (2016, Equation 1)
Skewed Student-t "sstd" See Trottier and Ardia (2016, Equation 1)
Skewed GED "sged" See Trottier and Ardia (2016, Equation 1)
Panel C: Methods
Function Description
CreateSpec Create a model specification
DIC Compute the deviance information criterion
ExtractStateFit Create a fitted object for each extracted regime
FitMCMC Fit the model by Markov chain Monte Carlo
FitML Fit the model by maximum likelihood
PIT Compute the probability integral transform
predict Compute the conditional volatility forecasts and the density forecasts
PredPdf Compute the predictive density (PDF)
Risk Compute the value-at-risk and expected-shortfall risk measures
simulate Simulate a MSGARCH process
State Compute the in-sample state probabilities
TransMat Compute the transition matrix
UncVol Compute the unconditional volatility
Volatility Compute the in-sample conditional volatilities

Table 1: Specifications and methods available in the R package MSGARCH. Panel A reports
the conditional volatility models. Panel B reports the PDF of the conditional distributions.
Panel C reports the methods.

b is obtained by maximizing the logarithm of (3).7 In the case of MCMC

The ML estimator Ψ
7
Starting values are chosen in the following way: 1) estimate using an expectation-maximization algo-
rithm the static version of the model, that is, with hk,t = h̄k ; 2) assign each observation to a regime of the
Markov chain using the Viterbi algorithm (see Viterbi 1967) and stack all the series in K vectors, one for each
regime; 3) estimate via Quasi-maximum likelihood a volatility model for each vector of decoded observations;
4) estimate via ML the shape parameters of the conditional distribution of the standardized decoded observa-
tions. Positivity and covariance-stationarity constraints are guaranteed through specific parameter-mapping
functions.
 Journal of Statistical Software 9

estimation, we follow Ardia (2008), by combining the likelihood with a diffuse (truncated)
prior f (Ψ) to build the kernel of the posterior distribution f (Ψ | IT ). As the posterior is of
an unknown form (the normalizing constant is numerically intractable), it must be approxi-
mated by simulation techniques. In the R package MSGARCH, draws from the posterior are
generated with the adaptive random-walk Metropolis sampler of Vihola (2012). For both ML
and MCMC estimations, positivity and covariance-stationarity constraints of the conditional
variance in each regime are ensured during the estimation.8
As detailed in Frühwirth-Schnatter (2006), mixture (and therefore Markov-switching) models
are identified up to a relabeling of the coefficients. For MSGARCH and mixture of GARCH
models, the same applies when the model specification in each regime is the same, and the
prior is symmetric. In the R package MSGARCH, by default the identification is carried out
by ordering the states according to the unconditional variance of each GARCH-type process,
from lower to higher values.
In the case of the Bayesian estimation, the likelihood function is combined with a prior f (Ψ) to
build the kernel of the posterior distribution f (Ψ | IT ). We build our prior from independent
diffuse priors as follows:

f (Ψ) = f (θ 1 , ξ 1 ) · · · f (θ K , ξ K )f (P)
f (θ k , ξ k ) ∝ f (θ k )f (ξ k ) I{(θ k , ξ k ) ∈ CSC k } (k = 1, . . . , K)
f (θ k ) ∝ fN (θ k ; µθ k , diag(σ 2θ )) I{θ k ∈ PC k } (k = 1, . . . , K)
k

f (ξ k ) ∝ fN (ξ k ; µξ , diag(σ 2ξ )) I{ξk,1 > 0, ξk,2 > 2} (k = 1, . . . , K) (6)

k k
 
K
Y K
Y
f (P) ∝  pi,j  I{0 < pi,i < 1} ,
i=1 j=1

where CSC k denotes the covariance-stationarity condition and PC k the positivity condition
in regime k; see Trottier and Ardia (2016). ξk,1 is the asymmetry parameter and ξk,2 the
tail parameter of the skewed Student-t distribution in regime k. fN (•; µ, Σ) denotes the
multivariate normal density with mean vector µ and covariance matrix Σ. Finally, µ• and
σ 2• are vectors of prior means and variances (of appropriate sizes), whose entries are set by
default to zero and to 1,000, respectively.

3. The R package MSGARCH

3.1. Model specification

Model specification in the R package MSGARCH is performed via the CreateSpec() function.
CreateSpec() accepts several arguments and returns an S3 object of class ‘MSGARCH_SPEC’
for which methods such as print() and summary() (among others) are available; we re-
fer the reader to help("CreateSpec") for the complete documentation. The arguments of
CreateSpec() are:
8
Note that these constraints are stronger than those derived by Haas et al. (2004a). However, they allow
us to introduce a wider range of conditional variance specifications and distributional assumptions in the
MSGARCH package.
10 MSGARCH: Markov-Switching GARCH Models in R

• variance.spec: A list with the element model, a character vector (of size K, that
is, the number of regimes/components) containing the conditional variance specifica-
tions. Valid models are "sARCH", "sGARCH", "eGARCH", "gjrGARCH" and "tGARCH" (see
Section 2.2 for details). By default, variance.spec = list(model = c("sGARCH",
"sGARCH")).

• distribution.spec: A list with element distribution, a character vector (of size K)

containing the conditional distribution specifications. Valid distributions are "norm",
"snorm", "std", "sstd", "ged", and "sged" (see Section 2.3 for details). The length of
distribution has to be the same as the length of model, otherwise an error is reported.
By default, distribution.spec = list(distribution = c("norm", "norm")).

• switch.spec: A list with elements do.mix and K. The first element, do.mix, is a logical
indicating if the model specification is Markov-switching or mixture of GARCH. If
do.mix = TRUE, a mixture of GARCH (Haas et al. 2004b) is specified, while if do.mix =
FALSE, a MSGARCH (Haas et al. 2004a) is specified. By default do.mix = FALSE. The
second element, K, is an integer which controls for the number of regime/components.
If the length of model in variance.spec and distribution in distribution.spec is
one, then the chosen specification is the same across the K regime/components. By
default, K = NULL, that is, the number of components is determined from the length of
the vector model in variance.spec.

• constraint.spec: A list with elements fixed and regime.const. The first element,
fixed, is a list with numeric entries and named elements. This argument controls
for fixed parameters set by the user. The names of the fixed entries have to co-
incide with the labels associated to the model parameters. For instance, if fixed =
list(beta_1 = 0)), beta_1 will be fixed to 0 during the optimization. The second
element, regime.const, is a character vector. It controls for the parameters which
are set equal across regimes. The names of the entries in the list have to coincide
with the names of the model parameters minus the regime indicator. For instance, if
constraint.spec = list(regime.const = "beta"), all the parameters named beta
will be the same across regimes. Note that both types of constraint cannot be defined
contemporaneously, that is, either fixed or regime.const can be specified. By default,
fixed and regime.const are set to NULL.

• prior: A list with elements mean and sd. The elements mean and sd are both lists with
numeric named elements which allow the user to adjust the prior mean and standard
deviation of the independent (symmetric) truncated normal priors in (6). The names of
the entries in the lists have to coincide with the labels associated to model parameters.
For instance, if prior = list(mean = list(beta_1 = 0.7), sd = list(beta_1 =
0.1)), the prior mean of beta_1 will be set to 0.7 while the prior standard devia-
tion will be set to 0.1.

As an illustration, let us consider the MSGARCH model of Haas et al. (2004a) with K = 2
regimes. This model assumes that conditionally on each regime of the Markov chain, returns
are normally distributed with GARCH(1, 1) variances. It can be easily specified with the
following lines of code:
 Journal of Statistical Software 11

R> library("MSGARCH")
R> spec <- CreateSpec()

The relevant information is summarized with the summary method:

R> summary(spec)

Specification type: Markov-switching

Specification name: sGARCH_norm sGARCH_norm
Number of parameters in each variance model: 3 3
Number of parameters in each distribution: 0 0
------------------------------------------
Fixed parameters:
None
------------------------------------------
Across regime constrained parameters:
None
------------------------------------------

The output printed in the console provides information regarding: (i) the model, such as
whether a Markov-switching or mixture of GARCH has been selected, (ii) the GARCH-type
specification within each regime, (iii) the number of variance parameters, and (iv) the number
of shape parameters in each regime. The presence of fixed parameters or equal parameters
across regimes set by fixed.pars and regime.const.pars is also displayed.
We now report examples of various MSGARCH models that can be specified in the R pack-
age MSGARCH. We refer the reader to the documentation for additional examples; see
help("CreateSpec").

Example 1: A single-regime model

The R package MSGARCH also supports single-regime models, as they are the building
blocks of Markov-switching models. A simple specification is the GARCH model with normal
conditional distribution:

R> spec <- CreateSpec(variance.spec = list(model = "sGARCH"),

+ distribution.spec = list(distribution = "norm"))

Example 2: A model with heterogeneous regimes

The R package MSGARCH has a modular approach to define the conditional variance and
density in each regime, the state dynamics and the number of regimes. Thanks to this flexi-
bility, the user can choose many combinations of specifications. Here we report an example of
a three-state MSGARCH model, where each regime is characterized by a different conditional
volatility and a different conditional distribution:

R> spec <- CreateSpec(

+ variance.spec = list(model = c("sGARCH", "tGARCH", "eGARCH")),
+ distribution.spec = list(distribution = c("snorm", "std", "sged")))
12 MSGARCH: Markov-Switching GARCH Models in R

Example 3: A model with non-switching shape parameters

The user can also choose to constrain the parameters to be the same across regimes. In the
R package MSGARCH package, the shape parameters ν and ξ reported in Section 2.3 are
identified with the labels "nu" and "xi", respectively. Here is an example of a two-state
MSGARCH model where the conditional distributions’ shape parameters of both regimes are
constrained to be the same:

R> spec <- CreateSpec(variance.spec = list(model = c("sGARCH", "sGARCH")),

+ distribution.spec = list(distribution = c("sstd", "sstd")),
+ constraint.spec = list(regime.const = c("nu", "xi")))

In this model, the only element that switches according to the Markov chain is the conditional
variance.

3.2. Model estimation

In the R package MSGARCH, model estimation can be either achieved by maximum likelihood
(ML) using the function FitML(), or via Markov chain Monte Carlo (MCMC) simulation
using the function FitMCMC().9 These functions accept three common arguments which are:
(i) spec, (ii) data, and (iii) ctr. The first argument, spec, is an ‘MSGARCH_SPEC’ object
created with the CreateSpec() function detailed in Section 3.1. The second argument, data,
is a numeric vector of T observations. The last argument, ctr, is a list of control parameters
for model estimation.
Control parameters are different between the FitML() and FitMCMC() functions. Regarding
the function FitML(), the following controls may be defined:

• par0: A numeric vector of starting parameters that overwrites the default starting
parameter scheme (see footnote 7 on page 8). The starting parameters should follow
the order of the default parameter of the ‘MSGARCH_SPEC’ object. By default par0 =
NULL.

• do.se: A logical indicating if standard errors are computed. In the case where the stan-
dard errors are not needed, setting do.se = FALSE will speed up estimation. Default
is do.se = TRUE.

• OptimFUN: A custom optimization function set by the user. By default, OptimFUN is set
such that optimization is done via the well-known Broyden-Fletcher-Goldfarb-Shanno
(BFGS) algorithm using the optim() function with method = "BFGS". For a custom
optimizer, we refer to the examples in help("FitML").

The function FitML() outputs an object of class ‘MSGARCH_ML_FIT’ which can be used with
several MSGARCH functionalities (see Section 3.5). We refer to help("FitML") for more
details.
Regarding the function FitMCMC(), the following control parameters may be defined:
9
We follow Ardia (2008) and use non-informative truncated normal priors.
 Journal of Statistical Software 13

• par0: A numeric vector of starting parameters, as for FitML().

• nburn: An integer indicating the number of discarded draws (i.e., the burn-in phase).
Default is nburn = 500L.

• nmcmc: An integer indicating the number of MCMC draws. Default is nmcmc =

10000L.

• nthin: An integer indicating the thinning factor (every nthin draws are kept in the
posterior sample). Default is nthin = 10L.

• SamplerFUN: A custom MCMC sampler. By default, SamplerFUN is set up such that

the estimation relies on the adaptive Metropolis-Hastings sampler described in Vihola
(2012). A custom MCMC sampler can be set up by the user using the SamplerFUN
element of ctr. We refer to the example section of help("FitMCMC") for an illustration.

• do.sort: A logical indicating if the MCMC draws are post-processed and sorted fol-
lowing Geweke (2007). By default, do.sort = TRUE, such that the MCMC draws are
ordered to ensure that the unconditional variance is an increasing function of the state
value. If the user sets do.sort = FALSE, no sorting is imposed, and label switching
can occur (for a given model specification). In this case, the user can extract the
MCMC chain and investigate a suitable identification constraint with the R package
label.switching (Papastamoulis 2016).10

The main purpose of nthin is to diminish the autocorrelation in the MCMC chain. The
argument nburn also serves as pre-optimization step. The total length of the chain is:
nmcmc/nthin. Finally, the chain is converted to a coda object, meaning that all functions for
MCMC analysis available in the R package coda (Plummer, Best, Cowles, and Vines 2006)
are available.

Illustration of FitML()
As an example, we estimate the MSGARCH model of Haas et al. (2004a) on daily observations
of the Deutschmark vs. British Pound (DEM/GBP) foreign exchange log-returns available in
the MSGARCH package. The sample period is from January 3, 1985, to December 31, 1991,
for a total of 1,974 observations. This data set has been promoted as a standard benchmark
for GARCH time series software validation (Brooks, Burke, and Persand 2001). It can be
loaded in the workspace by running:

R> data("dem2gbp", package = "MSGARCH")

We then fit the model to the dem2gbp data set using:

R> ms2.garch.n <- CreateSpec(variance.spec = list(model = "sGARCH"),

+ distribution.spec = list(distribution = "norm"),
+ switch.spec = list(K = 2))
R> fit.ml <- FitML(spec = ms2.garch.n, data = dem2gbp)
R> summary(fit.ml)
10
Since our MCMC sampler does not generate draws for the state variable, only the methods aic, pra,
stephens, and sjw in the R package label.switching can be used.
14 MSGARCH: Markov-Switching GARCH Models in R

Specification type: Markov-switching

Specification name: sGARCH_norm sGARCH_norm
Number of parameters in each variance model: 3 3
Number of parameters in each distribution: 0 0
------------------------------------------
Fixed parameters:
None
------------------------------------------
Across regime constrained parameters:
None
------------------------------------------
Fitted parameters:
Estimate Std. Error t value Pr(>|t|)
alpha0_1 0.0007 0.0003 2.307 1.053e-02
alpha1_1 0.0515 0.0184 2.796 2.590e-03
beta_1 0.9178 0.0061 149.819 <1e-16
alpha0_2 0.2813 0.1495 1.882 2.994e-02
alpha1_2 0.4805 0.4764 1.008 1.566e-01
beta_2 0.3996 0.2202 1.815 3.479e-02
P_1_1 0.9109 0.0680 13.394 <1e-16
P_2_1 0.5947 0.0369 16.127 <1e-16
------------------------------------------
Transition matrix:
t+1|k=1 t+1|k=2
t|k=1 0.9109 0.0891
t|k=2 0.5947 0.4053
------------------------------------------
Stable probabilities:
State 1 State 2
0.8697 0.1303
------------------------------------------
LL: -971.911
AIC: 1959.822
BIC: 2004.5245
------------------------------------------

The output of summary() reports various information regarding model estimation. Esti-
mated parameters are displayed along with significance levels according to their asymptotic
Gaussian distribution. The summary also returns the unconditional distribution (Stable
probabilities) of the Markov chain, π, such that Pπ = π. Finally, the log-likelihood
evaluated at its optimum together with Akaike and Bayesian information criteria are also
reported. Note that the ML estimation results are ordered with respect to the unconditional
variance in each regime, from lower to higher values, when regimes have the same model
specification.
 Journal of Statistical Software 15

Illustration of FitMCMC()
The function FitMCMC() outputs an object of class ‘MSGARCH_MCMC_FIT’ which, like the class
‘MSGARCH_ML_FIT’, can be used with several MSGARCH functionalities. We refer the reader
to the documentation for details; see help("FitMCMC").
Bayesian estimation for the MSGARCH model previously specified in the object ms2.garch.n
is easily performed with:

R> set.seed(1234)
R> fit.mcmc <- FitMCMC(spec = ms2.garch.n, data = dem2gbp)

The adaptive MCMC estimation procedure generates draws from the posterior distribution.11
These draws, referred to as the posterior sample, are used to characterize the distribution of
the parameters. As for FitML(), the summary() method is defined for an object of class
‘MSGARCH_MCMC_FIT’ delivered by FitMCMC():

R> summary(fit.mcmc)

Specification type: Markov-switching

Specification name: sGARCH_norm sGARCH_norm
Number of parameters in each variance model: 3 3
Number of parameters in each distribution: 0 0
------------------------------------------
Fixed parameters:
None
------------------------------------------
Across regime constrained parameters:
None
------------------------------------------
Posterior sample (size: 1000)
Mean SD SE TSSE RNE
alpha0_1 0.0011 0.0003 0.0000 0.0000 0.5000
alpha1_1 0.0315 0.0074 0.0002 0.0003 0.4944
beta_1 0.9401 0.0118 0.0004 0.0005 0.4751
alpha0_2 0.6341 0.1200 0.0038 0.0061 0.3807
alpha1_2 0.0756 0.0475 0.0015 0.0034 0.1900
beta_2 0.0431 0.0680 0.0022 0.0043 0.2535
P_1_1 0.9565 0.0095 0.0003 0.0004 0.5135
P_2_1 0.1785 0.0286 0.0009 0.0018 0.2500
------------------------------------------
Posterior mean transition matrix:
t+1|k=1 t+1|k=2
t|k=1 0.9565 0.0435
t|k=2 0.1785 0.8215
11
The adaptive MCMC sampler requires Cholesky decomposition and eigenvalue calculation. The obtained
results, therefore, depend on the linear algebra library used. We refer the reader to the computational details
at the end of the paper for more discussion about this.
16 MSGARCH: Markov-Switching GARCH Models in R

------------------------------------------
Posterior mean stable probabilities:
State 1 State 2
0.8042 0.1958
------------------------------------------
Acceptance rate MCMC sampler: 27.6%
nmcmc: 10000
nburn: 5000
nthin: 10
------------------------------------------
DIC: 1986.655
------------------------------------------

Among the various information reported, part of the output is generated by calling the
summary() method of the R package coda (Plummer et al. 2006). This part coincides with:
(i) the posterior mean (Mean), (ii) standard deviation (SD), (iii) naive standard error of the
mean (i.e., ignoring the autocorrelation in the MCMC chain) (SE), (iv) time series standard
error based on an estimate of the spectral density at zero (TSSE), and (v) the relative numer-
ical efficiency (RNE), defined as (SE/TSSE)2 . Other useful statistics are the acceptance rate
of the MCMC sampler and the deviance information criterion (DIC) of Spiegelhalter, Best,
Carlin, and Van der Linde (2002).

3.3. Forecasting
The R package MSGARCH implements functionalities to perform forecasting of the under-
lying time series for which the model has been estimated. Suppose an MSGARCH model is
estimated using the first T observations (y1 , . . . , yT ). Then, the user may be interested in: (i)
forecasting the volatility at time T + h (h > 0), that is, the standard deviation of the random
variable yT +h |IT ; or (ii) forecasting the shape of the distribution of yT +h |IT . Both types of
forecasts can be easily obtained using the function predict() in the R package MSGARCH
starting from an object of class ‘MSGARCH_MLE_FIT’ or ‘MSGARCH_BAY_FIT’ obtained via the
functions FitML() and FitMCMC(), respectively.
Volatility and density forecasts are performed via the predict() method defined for an
object of class ‘MSGARCH_MLE_FIT’ or ‘MSGARCH_BAY_FIT’. The horizon of the prediction, h,
is determined with the integer argument nahead (default: nahead = 1L). For h = 1, the
predictive volatility and density are available in closed form; in other circumstances (if h > 1)
forecasts are obtained by simulation, drawing iteratively a new observation from the (one-step
ahead) predictive distribution and updating the K conditional variance processes accordingly.
The number of draws controlling for the accuracy of the simulations are defined in the ctr
argument. Specifically, the entry nsim in ctr defines the number of draws such that, ctr =
list(nsim = 1e4L) indicates that 10,000 simulations are performed.
If an object of class ‘MSGARCH_BAY_FIT’ is provided, the number of simulations used to draw
from the predictive distribution is equal to the number of draws in the posterior distribu-
tions.12 However, if nsim is specified, then for each posterior draw, nsim observations are
sampled.
12
This is specified in the ctr argument of the FitMCMC() function via the nmcmc argument.
 Journal of Statistical Software 17

The predict() method returns an object of class ‘MSGARCH_FORECAST’ with two elements:
vol and draw. The first element, vol, is a numeric vector of length h, containing the standard
deviations of the distributions yT +j |IT for j = 1, . . . , h. The second element, draw, is a matrix
of dimension h × nsim of the simulated MSGARCH process. If we set do.return.draw =
FALSE in the function call, then draw is NULL.
For instance, the following code returns the five-step ahead conditional volatilities and draws
for the dem2gbp log-returns data using the MSGARCH model of Haas et al. (2004a) estimated
in Section 3.2:

R> pred <- predict(fit.ml, nahead = 5, do.return.draw = TRUE)

R> pred$vol

h=1 h=2 h=3 h=4 h=5

0.4006 0.3967 0.3836 0.3938 0.4004

R> pred$draw[, 1:4]

Sim #1 Sim #2 Sim #3 Sim #4

h=1 0.27169 0.39881 0.3893 -0.185686
h=2 -0.04223 -0.08693 -0.1556 -0.083054
h=3 -0.03076 -0.55994 0.2884 -0.060956
h=4 0.00159 0.27208 -0.1725 0.086989
h=5 -0.18180 -0.06522 -1.6092 0.004029

The predict() method also accepts the additional numeric argument newdata which allows
us to use a previous fit and to augment the data with new observations before forecasting.
Thus, if the vector of new observations is of size T ∗ , the prediction is made for the random
variable yT +T ∗+h |IT +T ∗ , using the parameter estimates obtained with observations up to
time T .

3.4. Quantitative risk management

One of the main applications of MSGARCH models is in quantitative finance where investors
want to allocate their wealth among a series of risky investment opportunities. In this area,
observations usually coincide with log-returns of financial assets, and quantities of interest
include the quantile of their future distribution at a specific risk level α ∈ (0, 1) as well as their
expected value below this level. These two quantities are referred to as value-at-risk (VaR)
and expected-shortfall (ES), respectively (see, e.g., McNeil et al. 2015). The VaR measures
the threshold value such that the probability of observing a loss larger or equal to it in a
given time horizon is equal to α. The ES measures the expected loss below the VaR level.
Formally, the VaR forecast in T + 1 at risk level α (given the information set up to at time
T ) is defined as:
VaRαT +1 ≡ inf {yT +1 ∈ R | F (yT +1 | IT ) = α} , (7)
where F (y|IT ) is the one-step ahead cumulative density function (CDF) evaluated in y. The
ES is obtained as:
ESαT +1 ≡ E[yT +1 | yT +1 ≤ VaRαT +1 , IT ] , (8)
18 MSGARCH: Markov-Switching GARCH Models in R

It is straightforward to estimate these two metrics with the R package MSGARCH, starting
from a fitted object of class ‘MSGARCH_ML_FIT’ or ‘MSGARCH_MCMC_FIT’. Evaluation of VaR and
ES can be performed both in-sample and out-of-sample using the Risk() function. In-sample
evaluation means that we want to investigate the evolution of the two risk measures over the
estimation period, whereas out-of-sample refers to prediction. The α level is controlled via
the alpha argument in Risk(). The argument alpha is a numeric vector such that different
α levels can be evaluated contemporaneously. By default alpha = c(0.01, 0.05). Other
arguments are:

• do.es: A logical indicating if expected-shortfall is also calculated. By default, do.es =

TRUE, such that both VaR and ES are reported. (Note that do.es = FALSE speeds up
the computations.)

• do.its: A logical indicating if in-sample risk measures are computed. Default is do.its
= FALSE such that out-of-sample risk measures are computed.

• newdata: A numeric vector of new observations (see Section 3.3). By default, newdata
= NULL.

• nahead: An integer indicating the forecast horizon. By default, nahead = 1L.

• do.cumulative: A logical indicating if the risk measure should be computed on the

distribution of simulated cumulative values. This is useful in the case of log-returns,
to compute the risk measures for the distribution of aggregated returns. By default,
do.cumulative = FALSE.

• ctr: A list of control parameters; see help("Risk").

The function Risk() creates an object of class ‘MSGARCH_RISK’ with elements VaR and ES.
Both VaR and ES are a matrix of dimension nahead × length(alpha) containing the risk
measures, VaR and ES, respectively. If do.its = TRUE, these matrices contain the in-sample
levels such that their dimension is T × length(alpha).13
For instance, using the fit.ml object created in Section 3.3, the five-step ahead VaR and ES
measures at the 1% and 5% risk levels are computed as follows:

R> risk <- Risk(fit.ml, alpha = c(0.01, 0.05), nahead = 5)

R> risk$VaR

0.01 0.05
h=1 -1.211 -0.5851
h=2 -1.153 -0.5403
h=3 -1.101 -0.5469
h=4 -1.166 -0.5641
h=5 -1.085 -0.5614

R> risk$ES
13
If newdata is provided, the dimension is (T + T *) × length(alpha), where, as before, T * =
length(newdata).
 Journal of Statistical Software 19

0.01 0.05
h=1 -1.549 -0.9403
h=2 -1.556 -0.9024
h=3 -1.588 -0.9075
h=4 -1.587 -0.9265
h=5 -1.551 -0.9068

Note that, if model estimation is performed via MCMC, and hence an object of class
‘MSGARCH_MCMC_FIT’ is provided to Risk(), the VaR and ES measures integrate the parameter
uncertainty; see Hoogerheide and Van Dijk (2010).

3.5. Other functionalities

Several other functionalities are available in the R package MSGARCH, which allow the user
to extract the in-sample conditional volatility and latent states (functions Volatility() and
State(), respectively), to simulate (function simulate()), to compute the predictive density
(function PredPdf()) and the probability integral transform (function PIT()). We refer the
reader to the documentation manual for details; see help("MSGARCH").
In all cases, objects of classes ‘MSGARCH_ML_FIT’ and ‘MSGARCH_MCMC_FIT’ can be used as an
input. In the case of the MCMC estimation, all functions return the aggregated value (using
the mean) over MCMC draws, thus integrating the parameter uncertainty.
For instance, we can easily simulate from the previous estimated model with FitML(). Below,
we generate two paths of length four with a burn-in phase of length 500.

R> simulate(fit.ml, nsim = 2, nahead = 4, nburn = 500)

$draw
Sim #1 Sim #2
t=1 -0.1322 0.1943
t=2 -0.5581 0.2910
t=3 -0.5703 0.6642
t=4 -0.3568 0.2495

$state
Sim #1 Sim #2
t=1 1 2
t=2 1 2
t=3 2 2
t=4 2 1

$CondVol
, , k=1

Sim #1 Sim #2
t=1 0.2962 0.4212
t=2 0.2866 0.4067
t=3 0.3035 0.3961
20 MSGARCH: Markov-Switching GARCH Models in R

t=4 0.3193 0.4091

, , k=2

Sim #1 Sim #2
t=1 0.7095 0.7025
t=2 0.7006 0.7047
t=3 0.7919 0.7214
t=4 0.8295 0.8374

The function outputs the simulated MSGARCH process together with the simulated regimes
and the conditional volatilities in each state.
It is worth emphasizing the difference between predict() and simulate(). The function
predict() is used for forecasting the conditional volatility and the predictive distribution
(using the argument do.return.draw = TRUE) while the function simulate() aims at gen-
erating simulation paths for a given MSGARCH model.
To perform in-sample model selection, information criteria such as the Akaike information
criterion (Akaike 1974), the Bayesian information criterion (Schwarz 1978), and the deviance
information criterion (Spiegelhalter et al. 2002) are available. These are all measures of the
parsimony of statistical models in describing a given set of data, where lower values indicate a
better performance in terms of goodness-of-fit. They are estimated with the functions AIC()
and BIC() from the stats package (R Core Team 2018), and DIC() from the MSGARCH
package, respectively.
For instance, the BIC can be computed from the model fit:

R> BIC(fit.ml)

[1] 2005

Finally, the function ExtractStateFit() allows the user to extract single-regime model re-
sults from the fitted object ‘MSGARCH_ML_FIT’ or ‘MSGARCH_MCMC_FIT’. Hence, if the user
estimates a model with K regimes, he can retrieve the information of the K single-regime
layers as a list of length K. For instance, still considering the object fit.ml, if we want to
predict the five-step ahead VaR conditionally of being in the two regimes, we simply write:

R> sr.fit <- ExtractStateFit(fit.ml)

R> risk1 <- Risk(sr.fit[[1]], alpha = 0.05, nahead = 5)
R> risk2 <- Risk(sr.fit[[2]], alpha = 0.05, nahead = 5)
R> VaR <- cbind(risk1$VaR, risk2$VaR)
R> colnames(VaR) <- c("State 1", "State 2")
R> VaR

State 1 State 2
h=1 -0.4411 -1.293
h=2 -0.4314 -1.515
h=3 -0.4244 -1.633
h=4 -0.4322 -1.724
h=5 -0.4206 -1.742
 Journal of Statistical Software 21

Hence, we are able to evaluate the risk exposure of an investment conditionally on differ-
ent regimes of the market. All the other package functionalities, like Volatility() and
predict(), can be applied to the single-regime layers extracted via ExtractStateFit().
This way, scenario analyses can be easily performed.
In Panel C of Table 1, we provide a summary of the various methods available in the package.

4. Empirical illustration
In this section, we illustrate how the R package MSGARCH can be used for model comparison,
state/regime smoothing, and volatility filtering. Estimation via MCMC is also discussed. Our
illustration focuses on an in-sample analysis of the daily log-returns of the major equity index
for the Swiss market, namely the Swiss market index (SMI). MSGARCH models have been
shown to be superior than single-regime counterparts on out-of-sample backtesting results;
see Ardia, Bluteau, Boudt, and Catania (2018) for a large-scale empirical study of their
performance. An example of a backtest implementation is provided at the end of this section.
The dataset can be loaded in the workspace by running data("SMI", package = "MSGARCH")
in the console. The plot of the time series is presented in Figure 1. Well-known stylized facts
observed in financial time series, such as volatility clustering and presence of outliers, are
evident from Figure 1 (see, e.g., McNeil et al. 2015). Furthermore, we also note that large
(absolute) returns are more frequent at the start (1990–1993) and at the end (1997–2000) of

SMI log−returns (%)

−5

1992 1994 1996 1998 2000

Date (year)

Figure 1: Percentage daily log-returns of the Swiss market index for a period ranging from
November 12, 1990, to October 20, 2000, for a total of 2,500 observations.
22 MSGARCH: Markov-Switching GARCH Models in R

the sample, than in the middle (1993–1997). This suggests that the conditional variance is
time-varying according to a regime-switching specification. For the sake of illustration, we
consider the asymmetric two-state MSGARCH model implemented by Ardia (2008, Chap-
ter 7) and Mullen, Ardia, Gil, Windover, and Cline (2011, Section 5). This is an extension of
the MSGARCH model introduced in Haas et al. (2004a), where a GJR variance specification
with a Student-t distribution is assumed in each regime. The model may be written as:

yt | (st = k, It−1 ) ∼ S(0, hk,t , ν) ,

2
(9)
hk,t ≡ α0,k + (α1,k + α2,k I{yt−1 < 0})yt−1 + βk hk,t−1 ,

where k ∈ {1, 2}. Note that the degree of freedom parameter of the Student-t distribution is
fixed across the regimes. Model specification and ML estimation of this model is performed
with the following code:

R> ms2.gjr.s <- CreateSpec(variance.spec = list(model = "gjrGARCH"),

+ distribution.spec = list(distribution = "std"),
+ switch.spec = list(K = 2),
+ constraint.spec = list(regime.const = "nu"))
R> fit.ml <- FitML(ms2.gjr.s, data = SMI)

The summary of the estimation is obtained as follows:

R> summary(fit.ml)

Specification type: Markov-switching

Specification name: gjrGARCH_std gjrGARCH_std
Number of parameters in each variance model: 4 4
Number of parameters in each distribution: 1 1
------------------------------------------
Fixed parameters:
None
------------------------------------------
Across regime constrained parameters:
nu
------------------------------------------
Fitted parameters:
Estimate Std. Error t value Pr(>|t|)
alpha0_1 0.2071 0.0488 4.2432 1.102e-05
alpha1_1 0.0005 0.0088 0.0569 4.773e-01
alpha2_1 0.2137 0.0619 3.4505 2.798e-04
beta_1 0.5264 0.0995 5.2921 6.045e-08
nu_1 9.2468 1.3292 6.9569 1.739e-12
alpha0_2 0.0922 0.0349 2.6397 4.149e-03
alpha1_2 0.0052 0.0169 0.3050 3.802e-01
alpha2_2 0.1516 0.0381 3.9771 3.488e-05
beta_2 0.8716 0.0354 24.6380 <1e-16
P_1_1 0.9977 0.0014 701.8429 <1e-16
 Journal of Statistical Software 23

P_2_1 0.0027 0.0017 1.6006 5.473e-02

------------------------------------------
Transition matrix:
t+1|k=1 t+1|k=2
t|k=1 0.9977 0.0023
t|k=2 0.0027 0.9973
------------------------------------------
Stable probabilities:
State 1 State 2
0.5407 0.4593
------------------------------------------
LL: -3350.8467
AIC: 6723.6934
BIC: 6787.7579
------------------------------------------

Parameter estimates indicate that the evolution of the volatility process is heterogeneous
across the two regimes. Indeed, we first note that the two regimes report different uncondi-
tional volatility levels:

R> set.seed(1234)
R> sqrt(250) * sapply(ExtractStateFit(fit.ml), UncVol)

[1] 11.87 22.04

as well as a different reactions to past negative returns: α2,1 ≈ 0.21 vs. α2,2 ≈ 0.15. Also the
volatility persistence in the two regimes is different. The first regime reports α1,1 + 12 α2,1 +β1 ≈
0.63 while the second regime reports α1,2 + 12 α2,2 + β2 ≈ 0.95.14 In summary, the first regime
is characterized by: (i) low unconditional volatility, (ii) strong volatility reaction to past
negative returns, and (iii) low persistence of the volatility process. Differently, the second
regime is characterized by: (i) high unconditional volatility, (ii) weak volatility reaction to
past negative returns, and (iii) high persistence of the volatility process. Clearly, regime
one would be perceived by market participants as “tranquil market conditions” with low
volatility levels, low persistence and high reaction to past negative returns, while regime two
as “turbulent market conditions” with high volatility levels and strong persistence.
b It ], P[St = k|Ψ,
Filtered, predicted, and smoothed probabilities (i.e., P[St = k|Ψ, b It−1 ] and
P[St = k|Ψ, IT ], respectively) can be computed starting from estimated objects using the
b
State() function. State() reports a list of four elements: (i) FiltProb, (ii) PredProb,
(iii) SmoothProb, and (iv) Viterbi. The first three elements are array objects of dimension
T × 1 × K containing the filtered, predicted, and smoothed probabilities at each time t,
respectively. The last element, Viterbi, is a matrix of dimension T ×1 of indices representing
decoded states according to the Viterbi algorithm detailed in Viterbi (1967).15 For instance,
smoothed probabilities of being in the second regime (k = 2) can be computed with the
following code:
14
Recall that E[η 2 I{η < 0}] = 21 when η is symmetrically distributed.
15
The reason why State() outputs arrays of dimension T ×1×K is because when it is evaluated for an object
of class ‘MSGARCH_MCMC_FIT’, the second dimension corresponds to the number of posterior draws. Analogously
for the Viterbi matrix.
24 MSGARCH: Markov-Switching GARCH Models in R

Smoothed probabilities
1.0 ●

● ●

0.8 ● ● ●
●
● ● ● ●
● ● ● ●●
● ● ● ● ● ●●● ● ●●
● ●
● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ●● ● ●
● ● ● ● ● ●
●● ● ●● ● ● ●● ●●● ● ● ● ● ●● ●● ● ● ●
● ●● ●● ● ● ●● ●
● ●● ● ● ●● ● ● ●●
● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ●
● ●
● ● ● ●● ●●● ●● ● ● ● ● ● ● ● ● ●
● ● ● ●●● ●●●● ●● ● ● ●
● ● ● ●● ●●● ● ● ● ● ●
● ● ●● ●
●
● ●●●●● ● ●● ● ● ● ●● ●
●●●
●● ● ● ●
●● ● ●
●● ● ● ●
● ● ●● ●●● ●
● ●●● ● ● ● ●● ●●●●● ●● ● ●
●● ●● ● ● ● ● ● ●●●●● ●● ● ● ● ● ●●● ● ● ●● ● ● ●●● ● ● ● ● ● ●● ●
0.6 ● ●
●
●●
●
● ●●
● ●●
●●● ●
● ●●●●
●
●● ● ●
● ●●
●● ●● ●●●●●●● ●
● ●
●
●●
●●
●●
●●
●●
●
●
● ●●●● ●
●
● ●
●
●●
● ●●
●
●●
●
●●● ●●●●●
● ●●
●●● ●
● ●
●
●●
●●●●● ● ●●●
● ●● ● ●
●
●●
●
●
●
●●●●●● ●
● ●● ●● ●
●
●● ● ● ● ●
●● ● ●● ●
● ●
●
●●
●●● ●●●
● ●●
● ●
●●
●●●● ● ●● ●
●● ●
●●●●
●● ●●
●
● ●
●
●● ●
●
●
●
●●●
●●●
●
●
●●● ●● ●●
● ●
●
●
●
●●
●
●● ●●
● ●
●
●
● ●●●● ●
●
● ● ● ●● ●
● ● ●
●
●● ●
●●●●●
●●
●●●●●
●●●●●●● ● ●
● ●●●●●●●●●● ●
●●● ●
●●●●● ● ●●
●
●● ●●● ●
●
● ●● ●● ●●● ●●●●●
●
●
●● ● ●●●●●●●●●● ● ●● ●●● ●●●●
●● ●● ● ●
●●●●●
●
●●
●●
●●● ●
●● ●●●● ●
● ●●●
●
● ● ●
● ●
●
●● ● ●●●● ●●●● ● ●●●● ● ●● ● ● ● ●● ●●●●
● ●
●● ●●● ● ●●●●● ●●●●●● ●●●● ●● ●●●●
● ●
●
●●●●
●
●● ●
●
● ●●● ●●● ● ●
●●● ● ● ●● ● ●●● ● ●● ●●●●● ●● ●●●●
●●● ● ●● ● ●●●●
●● ●●●● ● ●● ●● ● ●● ● ●
●● ●● ● ● ● ●● ● ● ●● ●● ● ● ●● ●●● ●
●●● ●● ● ●
● ●●
●●● ●●● ●●●
●●●
●
●● ●● ●●● ●●●●●● ●● ●●●
●
●● ●●●●
●
● ●● ● ●●●●
● ●
●●
● ●●●● ● ●● ●●
●● ● ●● ●● ●
●●
●
●●●●● ●
●
●●● ● ●●●●
●●● ● ●●●● ●● ● ●● ●●● ●●●●●● ●●
●● ●
● ●●●●● ●● ●●●
●●●●●●● ● ●● ●●●●●● ●●
●● ●
● ●●●●●● ●●● ●●●●●●●●●●
● ●●●●
●●
● ●● ●●● ● ● ●● ●
● ● ●●
● ●●
● ●● ● ●●● ●●●●● ● ●●●● ●
● ●●
●● ●●● ●●●
●●● ●●●
● ●
●● ● ●
●●●
● ●● ●●●●
●●
●●●●
●● ● ● ●●●● ●●
●●●●
● ●●●●● ●●●
●● ●● ●● ● ●●● ●●●● ●
●●●●●●
●●
●
●●●●●
●●●●● ●●●●●●
●●●
●●●●●
● ● ●●● ● ● ●
●●
● ●● ●
●
● ●●● ●●●●● ●●
● ●●●
●
●
●●
●● ●
●●
●●
●● ●●● ●●●●●●● ●● ●● ●●●●● ● ● ●
● ●●● ●●● ● ●● ●●● ●
●● ●●
● ●● ● ●●
●●
●
● ●● ●
●● ●●● ● ●
●
●
●●● ●●● ●●●
● ●●●● ●
●
●●
●
● ●●● ●
●● ●●● ● ●●●●● ● ●
● ●● ●
● ●●
●
●●●
●●●●●●●●●
● ●●
●● ●
●
●
●● ●● ● ●●● ● ● ●●
●●●●● ●●
●●
●
●● ●●● ●●●●
●● ●●● ●
●●●●
●
●
●●
●
●
● ●●
●●●●●●●●● ●●● ● ●●●●●●●
● ●●●● ● ●● ●●●●●●●● ●●
●●●●
●● ●● ●● ●●● ●● ●● ●● ●●
●
●●
●●●●
●●● ●●●● ● ●●
●●●●●●●●
●● ●●
●●●●● ●
● ●● ●●● ●● ●
● ●● ●
●●●●●●●●
●
●●● ●●●
●●● ●● ● ● ●● ●●●
● ● ●
●●
● ●
●●● ●
● ●
●
● ●●
●
●●
●● ●
● ●
● ●●
●
●
● ● ● ● ●
● ● ●●
● ●
●
● ●● ● ● ● ●● ● ●
●● ● ●●
● ●
●
●● ●
●
●
●
●●
●●
●
●●●
●●●●●●
● ●●●● ●●●●●● ● ●
● ●●
● ●
●●
●● ● ●●
●●
● ●●● ● ●
● ● ●
● ●●
●● ● ●
● ●
●
● ●● ●
●
●● ● ● ●● ● ●● ● ●●
● ●● ●
●
● ● ●
● ●● ●●
● ● ● ●
●●
●● ●● ● ● ● ●●● ●●● ● ●●● ●● ●
● ●●●● ●● ●●●
●● ● ●● ● ●
● ●
●●●●● ●
●
●● ● ● ● ●●● ●●
● ●● ●●●●●● ●● ● ●● ●● ●●●● ● ● ●●●● ● ●●● ●● ● ●●● ●●● ●● ● ● ●
● ● ● ● ● ● ● ●●●● ● ● ● ●
● ●●● ● ●●●●●●
●● ●● ●● ● ●
●
●● ●●●●● ●● ●
●●●
●●●●
●● ●● ● ●●
● ● ●
● ●
●●●
●● ●●●●●● ● ●●●● ●● ●●●
●●●●
●● ●● ●● ● ●●
● ●● ●● ●●●
● ●●●●● ●● ●● ● ●● ● ●●● ● ●● ● ● ● ● ●● ●●● ●●● ● ●● ● ●● ●● ●● ●
●●● ●●● ● ● ●●●●●
● ●●●●●●
● ●●
● ●● ●● ●●● ●
●
●
●●
● ●
●●●
●
●● ●●● ● ●●●●● ● ● ●● ●
● ● ● ●● ●
● ● ●●● ● ●● ● ●● ●● ● ● ●●●●● ● ●●● ●
●● ●●● ● ● ● ●
● ●● ● ●● ●● ● ● ● ● ● ●● ● ● ●
● ● ●● ●● ● ● ●● ● ● ●●● ● ● ● ● ●● ●● ● ●
● ●
● ● ● ● ● ● ● ● ● ●● ●● ● ●
● ● ●
●●●
●●●●●● ● ● ● ●●
●●●●
● ● ● ●●
●● ● ●● ● ●●●●●●● ●●● ● ●● ●
● ●●● ● ●
●● ●● ●● ● ●● ● ● ● ● ● ●● ●
● ●●●●●●●● ●● ● ●● ● ● ●● ●
●
●● ●
●● ● ● ● ●● ●
● ● ●● ●
● ●● ● ●● ● ●● ●● ● ● ●● ●
● ● ● ●● ●●●● ●
●●● ●●●●● ●● ● ● ●●●●● ●● ● ●● ● ● ●
●● ●
●
● ● ●● ● ●● ● ● ● ●● ● ●● ●●
● ● ● ● ●● ● ●● ● ●● ● ●●● ● ●● ●
●●
●● ● ● ●●● ● ● ●
● ●● ● ● ● ● ● ● ● ●●● ●● ● ● ●
● ● ● ●● ● ●
●● ● ● ● ●●● ● ●
● ● ● ● ● ● ● ● ●● ● ● ●
0.4 ● ● ●
●
●
●
●
●
●
●
●●
●
●
●
●
●
● ● ● ●●●●
●
● ●
● ● ●●
● ● ● ●●
● ● ●
● ●
● ●● ● ●
●● ● ●
● ●

● ● ● ●
●
● ● ●● ● ●
● ● ●
● ● ●
● ●●
●
● ●
● ● ●
●●
● ●●
0.2 ●●

0.0 ●

1992 1994 1996 1998 2000

Volatility
Date (%)
60

50

40

30

20

10
1992 1994 1996 1998 2000
Date

Figure 2: Top: Estimated smoothed probabilities of the second regime, P[st = 2|Ψ, b IT ], for
t = 1, . . . , T . The thin black line depicts the SMI log-returns. Bottom: Filtered conditional
volatilities.

R> smooth.prob <- State(fit.ml)$SmoothProb[, 1, 2, drop = TRUE]

Figure 2 displays the smoothed probabilities of being in regime two (high unconditional volatil-
ity regime), P[St = 2|IT ] for t = 1, . . . , T , superimposed on the SMI log-returns (top graph)
as well as the filtered volatility of the overall process (bottom graph). Filtered (annualized)
volatilities are extracted from estimated object using the function Volatility():

R> vol <- sqrt(250) * Volatility(fit.ml)

We refer the reader to help("Volatility") for further details. As expected, when the
smoothed probabilities of regime two are near one, the filtered volatility of the process sharply
increases. Interestingly, we further note that the Markov chain evolves persistently over
time and that, in the limit, as reported by the summary() (Stable probabilities), the
probabilities of being in the two states are about 54% and 46%. As documented by Ardia
(2008) and Mullen et al. (2011), ML estimation can be difficult for MSGARCH-type models.
Fortunately, MCMC procedures can be used to explore the joint posterior distribution of
the model parameters, thus avoiding convergence to local maxima commonly encountered
via ML estimation. The Bayesian approach offers additional advantages. For instance, exact
 Journal of Statistical Software 25

distributions of nonlinear functions of the model parameters can be obtained at low cost by
simulating from the joint posterior distribution, and parameter uncertainty can be integrated
in the forecasts through the predictive distribution.
To estimate MSGARCH-type models from a Bayesian viewpoint, the R package MSGARCH
relies on the adaptive MCMC sampler of Vihola (2012).16 Specifically, the estimation method
is a random-walk Metropolis-Hastings algorithm with coerced acceptance rate. We observed
excellent performance in the context of (identified) mixture models. Using the ML parameter
estimates as starting values, we can estimate the model by MCMC as follows:

R> nmcmc <- 12500

R> nburn <- 5000
R> nthin <- 5
R> ctr <- list(nmcmc = nmcmc, nburn = nburn, nthin = nthin,
+ par0 = fit.ml$par)
R> fit.mcmc <- FitMCMC(ms2.gjr.s, data = SMI, ctr = ctr)
R> summary(fit.mcmc)

Specification type: Markov-switching

Specification name: gjrGARCH_std gjrGARCH_std
Number of parameters in each variance model: 4 4
Number of parameters in each distribution: 1 1
------------------------------------------
Fixed parameters:
None
------------------------------------------
Across regime constrained parameters:
nu
------------------------------------------
Posterior sample (size: 2500)
Mean SD SE TSSE RNE
alpha0_1 0.2101 0.0323 0.0006 0.0015 0.1826
alpha1_1 0.0007 0.0002 0.0000 0.0000 0.2079
alpha2_1 0.2110 0.0408 0.0008 0.0017 0.2251
beta_1 0.5216 0.0526 0.0011 0.0026 0.1668
nu_1 9.6161 1.4718 0.0294 0.0753 0.1528
alpha0_2 0.1417 0.0454 0.0009 0.0021 0.1803
alpha1_2 0.0062 0.0024 0.0000 0.0001 0.1839
alpha2_2 0.1839 0.0425 0.0009 0.0021 0.1671
beta_2 0.8316 0.0359 0.0007 0.0017 0.1749
P_1_1 0.9971 0.0013 0.0000 0.0001 0.2216
P_2_1 0.0045 0.0017 0.0000 0.0001 0.1765
------------------------------------------
Posterior mean transition matrix:
t+1|k=1 t+1|k=2
t|k=1 0.9971 0.0029
16
However, as reported in Section 3.2, the user is free to implement his own sampler.
26 MSGARCH: Markov-Switching GARCH Models in R

t|k=2 0.0045 0.9955

------------------------------------------
Posterior mean stable probabilities:
State 1 State 2
0.6137 0.3863
------------------------------------------
Acceptance rate MCMC sampler: 28.3%
nmcmc: 12500
nburn: 5000
nthin: 5
------------------------------------------
DIC: 6721.4909
------------------------------------------

The posterior distribution of mixture and Markov-switching models often exhibits non-ellipti-
cal shapes which lead to non-reliable estimation of the uncertainty of model parameters
(see, e.g., Ardia, Hoogerheide, and Van Dijk 2009). This invalidates the use of the Gaus-
sian asymptotic distribution for inferential purposes in finite samples. Our results display
this characteristic as shown in Figure 3 where we plot 2,500 draws of the posterior sample
for the parameters α1,1 and α1,2 . The blue square reports the posterior mean while the red
triangle reports the ML estimate. An interesting aspect of the Bayesian estimation is that
we can make distributional (probabilistic) statements on any (possibly nonlinear) function
of the model parameters. This is achieved by simulation. For instance, for each draw in
the posterior sample we can compute the unconditional volatility in each regime, to get its
posterior distribution. Figure 4 displays the posterior distributions of the unconditional an-
nualized volatility in each regime. In the low-volatility regime, the distribution is centered
around 8.9% per annum. For the high-volatility regime, the distribution is centered around
32.7% per annum. The 95% confidence bands given by the Bayesian approach are [7.9%,
10.3%] and [25.0%, 46.6%], respectively. Notice that both distributions exhibit positive skew-
ness. Hence, relying on the asymptotic normal approximation with the delta method would
yield erroneous estimates of the 95% confidence band of the unconditional volatility in each
regime. We show now how to include parameter uncertainty in the one-step ahead predictive
density of MSGARCH models. The predictive density is computed via the PredPdf() func-
tion defined for ‘MSGARCH_ML_FIT’ and ‘MSGARCH_MCMC_FIT’ objects. The arguments of the
PredPdf() function are:

• object: An object of class ‘MSGARCH_ML_FIT’, ‘MSGARCH_MCMC_FIT’, or ‘MSGARCH_SPEC’.

• x: A numeric vector of mesh points on the domain for which the predictive distribution
is calculated.
• par: A numeric vector of model parameters if object is of class ‘MSGARCH_SPEC’. By
default par = NULL.
• data: A numeric vector of observations if object is of class ‘MSGARCH_SPEC’. By default
data = NULL.
• log: A logical indicating if the logarithm of the density is returned. By default, log =
FALSE.
 Journal of Statistical Software 27

0.35 ●●

● ●
● ●
● ● ●●
●
● ● ● ●●
● ● ● ●
● ● ●
●●
●●
● ● ● ● ●
● ● ● ●
● ●● ● ● ●
0.30 ● ●
●
● ●● ● ●●●
● ●
● ● ●●
●●● ● ●
●●● ● ● ●● ● ● ●● ●
● ● ● ●
● ● ●● ●
● ● ● ●
● ● ● ● ●
●● ● ●● ● ●●● ● ●
●● ●
●
●●
● ● ●● ● ●
● ●● ●
● ●● ●
● ● ● ● ● ● ●
●● ● ●● ●●
● ● ●●● ●● ●● ● ● ●
● ● ●
● ● ●● ●● ●● ●● ●
●● ● ● ●
●●●
● ●●● ● ● ●●●● ● ●● ●
●● ● ●
●●
● ●● ●●
● ●
● ● ●●● ● ●● ● ●● ● ● ●● ● ●
● ●● ● ●●● ● ● ●● ● ●
● ●● ●●
●●
●●●● ● ●●●● ● ●● ●
●
● ●● ●●●● ●● ● ● ● ● ●
● ●● ●
●● ● ● ●● ●● ● ●● ● ● ●●●●
● ●
● ●● ● ● ● ● ● ●
● ● ●●●● ● ● ● ●
● ●● ●● ● ●● ● ●
0.25 ● ●● ● ● ● ●● ●● ● ●●●●● ● ●
alpha2_1

● ●●●
● ● ●● ● ●● ●
● ● ● ●●● ● ●● ●● ● ●● ●● ● ● ●
● ● ●● ●● ●●●●
● ● ● ● ●●● ● ● ●● ● ●● ●● ● ● ●
● ● ● ● ● ●● ● ● ●●●●● ●●
●●● ● ●● ● ●
●●●
●● ● ●●● ● ●● ●
●
●
●
●
●●●● ●●●
●
●
●●● ● ● ●● ●●●● ● ● ● ●● ●
● ● ● ● ● ● ● ●
●
●● ●● ●● ●● ●●
● ● ● ●●
●●● ●●● ● ●●●●
●● ●● ●●●●● ●
● ●●
● ●●
● ●● ● ● ●
●● ●● ● ● ●
● ● ●● ● ●●● ●
● ●
● ● ● ●
●●
●
●●● ●●●●●●●
●●●
●
●● ●●
● ●●● ● ●● ●● ● ● ● ● ●
●●●
●●● ● ●● ● ● ● ●●
●
● ●● ●● ● ● ●
●● ●● ● ●●● ● ● ●● ●●● ●●● ● ● ●●
● ● ●●●● ●
●
●●●●●
●
●●●●●● ●
● ●●
● ●● ●● ● ● ●●●
●●●
● ● ●●●● ●●● ●● ●●● ●
●
● ● ●● ●●● ● ● ● ● ●
● ● ●
●
● ● ●● ● ●●●● ●●
●● ●●● ●●●●
● ● ●● ●● ●●● ●
●●● ●● ●
● ● ●
● ●●
●● ●● ●●●●●
●●●●●●●●● ● ● ● ● ● ●
● ●
● ●●
●●
● ●● ●● ● ●●●●● ● ● ● ● ● ● ●● ●●
● ●
● ● ● ●● ● ● ●● ●●●
●● ●●●●● ● ● ●● ●
●● ●● ● ● ●●●●● ● ●●
●● ● ●●●● ●●● ●
●●●● ● ●●
●●
●
● ●●● ●
●
● ● ●
●● ●●
● ● ●●●●● ●●●●●●
● ●● ● ●●● ●
●●
● ●● ●●● ● ●●● ●●● ● ●●
●● ● ●●● ● ● ● ● ● ● ● ● ● ● ●
● ●● ● ●● ●
● ●●
●
●●●● ●
● ●●● ●●● ●
●● ●●●
●
●● ●●
● ●● ● ●
● ● ● ●●
●● ●
● ●● ●● ●● ●●●●
●●●● ● ●●●●●●● ● ● ● ●● ● ● ●● ● ● ● ●
● ●● ●●●● ●●● ● ●● ● ● ● ● ● ●● ●●
●
●●●● ●●
●
●●
●● ● ● ● ● ● ●
0.20 ● ● ● ●● ●● ● ● ●● ●● ●
●●●●● ●● ●●● ●● ● ●● ● ●● ●
● ● ●●● ●●●●● ●
● ●● ● ●●●
● ●● ●● ●●● ●
●●●●●
●●●● ● ● ●
● ● ●
●●
●● ● ●
● ●● ●●●
● ●●
● ●●●● ●● ●● ● ●●● ● ●
●
●●●
● ●●
●
●
●
●
●●
● ●
●
● ●●● ● ●
● ●
● ● ● ● ● ●●● ●●●●● ● ● ● ●● ● ● ● ●
● ●● ●●●●
● ●● ●● ●● ●●● ●● ●
●●●
●
● ●●●● ●● ●● ●● ● ●
●● ● ● ●● ●● ●●●●● ● ●●
● ●●●
●
●●● ●●●● ●● ●● ●●
● ●●● ● ● ● ● ●●●●
● ●
●●● ●● ●●●
●●●●●● ●
●● ● ●● ● ● ● ●●● ● ● ●● ● ●
●
● ● ●●●● ●●● ● ●●●
● ●●● ●●● ●● ●●● ● ● ●●●
● ● ●● ●
● ● ●
●● ●●
●●● ●●
● ●● ● ●
●
●
●●
●●● ● ● ●●●● ●
●●●
●●● ●●●●●●●● ●●●● ●● ●
● ●● ● ●
● ●
● ● ● ● ●●
●● ●● ● ●● ●● ●● ●
●
● ●●● ●●●●● ●● ●
●
● ● ● ● ●
● ● ●● ● ● ●
●
●
● ● ●● ● ● ● ●
●●●●
●●● ● ●●●● ● ● ● ●
● ●
● ● ● ● ●
● ● ● ● ● ●● ●●●● ●● ● ● ●● ●●● ●●● ● ●●
● ●●● ● ● ●●●● ● ●
● ● ● ●●● ●● ● ● ● ●●
●●●● ●● ● ●●●●● ●●● ●●● ● ●● ● ●●
●● ●● ● ●● ● ●
● ●● ● ● ●● ● ●●●● ●
●
●●● ● ● ● ●● ● ● ● ● ●● ●●● ●
● ● ●● ● ● ●● ●● ●● ● ●● ● ●● ●● ●● ●●● ●● ●●
● ● ● ●
● ● ●
●
● ●● ● ● ●● ●
● ● ●●● ●● ●●●●
● ● ●● ●● ● ● ●●● ●
●
● ● ●● ● ● ●●●●● ● ● ● ● ●
● ●● ● ●
●
● ●● ●● ●●● ● ●
● ●
●
●● ● ●● ●● ●●
● ●●●●●
● ●●
● ●
●●
●
● ● ● ● ●● ●● ●● ● ● ● ● ●
0.15 ●
●●● ● ●
●● ●
● ● ● ●● ● ●
● ● ● ● ●
●● ● ● ● ● ●
●● ● ● ●● ● ●
● ● ●
● ●● ●
●● ● ● ● ●● ● ●
● ● ●● ● ● ● ●● ● ● ● ● ● ● ●● ●●
● ● ● ● ● ●
● ● ● ●
●
●● ●
●
●
● ● ● ● ●● ●
● ●● ● ●
● ●● ●
● ● ●● ● ● ● ●
● ● ●● ●
● ● ● ● ●
● ● ● ●●
● ● ●
● ●
● ● ●●
● ● ●

0.10
0.0005 0.0010 0.0015
alpha1_1

Figure 3: Scatter plot of posterior draws from the marginal distribution of (α1,1 , α1,2 )> ob-
tained with the adaptive random walk strategy. The blue square reports the posterior mean,
and the red triangle reports the ML estimate. The graph is based on 2,500 draws from the
joint posterior sample.

• do.cumulative: A logical indicating if the predictive density should be computed on

the distribution of simulated cumulative values. This is useful in the case of log-returns,
to compute the predictive density of multi-period returns. By default, do.cumulative
= FALSE.

Additional arguments are: nahead, do.its, and ctr; see help("PredPdf"). As for other
functions available in the R package MSGARCH, when an object of class ‘MSGARCH_MCMC_FIT’
is used, the output is computed accounting for parameter uncertainty. For instance, if we
want to evaluate the one-step ahead predictive density in the range of values from −5 to 0,
we run:

R> nmesh <- 1000

R> x <- seq(from = -5, to = 0, length.out = nmesh)
R> pred.mle <- as.vector(PredPdf(fit.ml, x = x, nahead = 1))
R> pred.bay <- as.vector(PredPdf(fit.mcmc, x = x, nahead = 1))

where pred.ml and pred.mcmc are numeric vectors of length 1,000. We stress that PredPdf()
28 MSGARCH: Markov-Switching GARCH Models in R

Regime 1 Regime 2

0.8

0.08

0.6

0.06

0.4

0.04

0.2
0.02

0.0 0.00

8 9 10 11 20 30 40 50 60 70 80
Volatility (%) Volatility (%)

Figure 4: Histograms of the posterior distribution for the unconditional volatility in each
regime. Both graphs are based on 2,500 draws from the joint posterior sample. The blue
square reports the posterior mean, and the red triangle reports the ML estimate.

is also defined for objects of class ‘MSGARCH_SPEC’ such that, after providing the vector of
parameters par, and a series of observations data, the predictive distribution can be evaluated
also for non-estimated models.17 Another possibility is to extract the posterior draws from
an estimated object of class ‘MSGARCH_MCMC_FIT’ and to compute the predictive density for
each of them. This way, we can investigate the impact of the parameter uncertainty on
the conditional distribution as, for example, in Ardia, Kolly, and Trottier (2017). This is
performed with the following commands:

R> draws <- fit.mcmc$par

R> pred.draws <- matrix(data = NA, nrow = nrow(draws), ncol = nmesh)
R> for (i in 1:nrow(draws)) {
+ tmp <- PredPdf(ms2.gjr.s, par = draws[i, ], x = x, data = SMI,
+ nahead = 1)
+ pred.draws[i, ] <- as.vector(tmp)
+ }
17
This is also the case for the functions Risk(), PIT(), Volatility() and predict(); see the related R
documentation.
 Journal of Statistical Software 29

Figure 5: Left tail of the one-step ahead predictive distribution for the SMI log-returns. The
solid blue line reports the predictive Bayesian density (integrating the parameter uncertainty),
the dashed red line reports the ML conditional density, and the solid light-blue lines the
conditional densities obtained for each of the 2,500 draws in the posterior sample.

In Figure 5, we display the left tail of the one-step ahead distribution for the log-returns. We
see why integrating parameter uncertainty can be useful for left-tail prediction. The Bayesian
predictive density (solid blue line) is a particular average of the predictive densities that can
be formed with individual posterior MCMC draws (thin solid blue lines). It is generally more
conservative than the predictive density with plugged ML estimates (dashed red line) and
offers additional flexibility by accounting for all likely scenarios within the model structure.
As a last illustration in our empirical section, we provide the code for a backtest analysis of
the ms2.gjr.s model against its single-regime counterpart. First we create the single-regime
model and gather both specifications in a list.

R> gjr.s <- CreateSpec(variance.spec = list(model = "gjrGARCH"),

+ distribution.spec = list(distribution = "std"),
+ switch.spec = list(K = 1))
R> models <- list(gjr.s, ms2.gjr.s)

Then we define the backtest settings. We decide to test the performance of the models over
1,000 out-of-sample observations and focus on the one-step ahead value-at-risk forecasts at
30 MSGARCH: Markov-Switching GARCH Models in R

Backtesing VaR at 5% risk level

VaR 5% − GJR−std ●

VaR 5% − MS2−GJR−std

6
●

4 ● ●
●
●

●
● ●
● ●
● ●

● ●
● ● ● ● ●
● ●
● ● ● ● ● ●●
● ● ● ● ● ●
● ● ● ● ●
● ●● ● ●
2 ● ● ●
●
● ●
● ●
●
● ●
● ●
● ●
● ● ● ● ●
● ● ● ●
● ● ● ● ● ● ●●●
●
● ● ● ●
●
● ● ● ● ● ● ●● ●
● ● ● ● ●
● ● ● ●● ● ● ●● ● ● ● ●● ● ●
● ●● ● ● ● ●●● ● ● ● ● ● ● ● ● ●
● ● ● ●
●
● ● ●● ●● ● ● ●
● ● ● ●
● ● ● ● ● ● ● ●
● ● ● ● ● ● ● ● ● ● ●
● ● ● ● ●●
● ● ●● ● ● ●● ●●● ● ●
●
● ●
●
● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●
● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●
● ●● ● ● ● ● ● ● ●●● ●
●● ● ●●● ● ●●● ● ● ● ●● ● ● ● ●●● ● ●●
● ● ● ●
● ● ● ● ● ● ● ●
●
● ● ● ●● ● ● ● ● ●● ●● ● ●●
● ●●
●● ● ● ● ● ● ● ● ●● ● ● ● ● ●
● ● ● ● ● ●●
● ●
● ● ● ● ●
●
● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●
● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ●● ●● ●
●●
● ●● ●● ● ● ●●
● ● ●● ●● ●
●
● ● ●● ●
●
●●
● ● ● ● ● ● ●
● ●● ●
● ● ●● ●
● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●
● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●
●●●● ●● ● ●● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ●●●
● ●●
●● ●● ● ● ● ● ● ● ●● ● ●● ● ● ●●
0 ●● ●●
●● ● ●● ●
●
●
●●
● ●
●●
●
●●● ● ●
●
● ● ●● ●●
●
●
● ● ●● ●
● ●
● ● ●
● ●
● ●●
●
●
● ●
●●
● ● ●● ●
● ● ●●
●●
● ● ●
● ●
●
●
●
●● ●
● ●
● ●●
●
●
●● ●
●● ●
●
●● ●●
● ● ● ●●● ● ● ●● ●● ● ●
●●
●
● ● ●●
● ● ● ● ●● ● ●
●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●
●● ● ● ● ● ● ●●● ● ● ● ● ●
● ●
● ● ●
● ●● ● ● ● ● ● ●● ● ● ●● ● ● ●● ● ● ●● ● ●●● ● ●
● ● ● ● ● ●
● ●● ●● ● ●●● ● ●●● ●● ● ● ● ● ● ● ● ●
● ● ●●● ●●● ● ● ● ● ● ● ● ●●●
●● ● ● ●● ●● ● ●
● ● ●●
● ● ● ● ● ● ●
● ● ●● ● ● ● ●● ●● ● ● ●● ● ● ●
● ● ● ● ● ● ● ● ● ● ●
● ● ● ● ● ●
● ● ● ●
●
● ● ●● ● ● ●● ● ● ● ● ● ●
●
● ● ● ● ●
● ● ● ● ● ●● ● ● ● ●
●
● ● ● ● ● ●
●● ● ● ● ● ● ●
● ●
● ●
● ● ● ● ●● ●
●● ● ● ● ● ●● ● ● ● ●
● ● ● ● ● ●
● ● ● ●● ● ● ● ●
● ● ● ● ● ●
● ● ●
● ● ● ● ●
● ● ● ●
● ● ● ●
●
−2 ● ● ●
●●
● ●
●
● ● ● ●
●
● ●
● ●●
● ● ●
● ● ● ●
● ●
● ● ● ● ●
● ● ●●
●
● ● ● ●
●
● ●
●
● ● ●
● ●
●
−4 ●
●
● ●
●
●
●
●
●

● ●

1997 1998 1999 2000

Date (year)

Figure 6: One-day ahead VaR forecasts at the 5% risk level provided by the Markov-switching
(solid blue line) and single-regime (dashed red line) specifications together with the realized
returns (black dots).

the 5% risk level. Forecasts are based on rolling windows of 1,500 observations, and models
are re-estimated every 100 observations for speed of illustration purposes.

R> n.ots <- 1000

R> n.its <- 1500
R> alpha <- 0.05
R> k.update <- 100

We initialize the vector of out-of-sample returns and the matrix of VaR forecasts. Then, we
loop over the observations in the out-of-sample window. For each new observation, we use
the last 1,500 data to predict the one-step ahead VaR at the 5% level, and store it. Models
are estimated by ML every 100 observations in the out-of-sample window.

R> VaR <- matrix(NA, nrow = n.ots, ncol = length(models))

R> y.ots <- matrix(NA, nrow = n.ots, ncol = 1)
R> model.fit <- vector(mode = "list", length = length(models))
R> for (i in 1:n.ots) {
+ cat("Backtest - Iteration: ", i, "\n")
+ y.its <- SMI[i:(n.its + i - 1)]
 Journal of Statistical Software 31

+ y.ots[i] <- SMI[n.its + i]

+ for (j in 1:length(models)) {
+ if (k.update == 1 || i %% k.update == 1) {
+ cat("Model", j, "is reestimated\n")
+ model.fit[[j]] <- FitML(spec = models[[j]], data = y.its,
+ ctr = list(do.se = FALSE))
+ }
+ VaR[i, j] <- Risk(model.fit[[j]]$spec, par = model.fit[[j]]$par,
+ data = y.its, n.ahead = 1, alpha = alpha, do.es = FALSE,
+ do.its = FALSE)$VaR
+ }
+ }

In Figure 6 we display the resulting VaR forecasts of the two models together with the realized
returns. We notice the discrepancy between the forecasts in 1998. The economically relevant
question is then to evaluate which of the VaR forecasts are most accurate in terms of correctly
predicting the α-quantile loss such that we expect to have a proportion α of exceedances. We
use the R package GAS (Ardia et al. 2019b) to compute the p values of two backtesting
hypothesis tests of correct conditional coverage of the VaR: the conditional coverage (CC)
test by Christoffersen (1998) and the dynamic quantile (DQ) test by Engle and Manganelli
(2004). Both tests aim at determining if the VaR forecasts achieve correct unconditional
coverage and if the violations of the VaR are independent over time. Both should be fulfilled
under correct model specification.

R> library("GAS")
R> CC.pval <- DQ.pval <- vector("double", length(models))
R> for (j in 1:length(models)) {
+ test <- GAS::BacktestVaR(data = y.ots, VaR = VaR[, j],
+ alpha = alpha)
+ CC.pval[j] <- test$LRcc[2]
+ DQ.pval[j] <- test$DQ$pvalue
+ }
R> names(CC.pval) <- names(DQ.pval) <- c("GJR-std", "MS2-GJR-std")
R> print(CC.pval)

For both tests, we notice the best performance of the Markov-switching specification.

GJR-std MS2-GJR-std
0.02092 0.08402

R> print(DQ.pval)

GJR-std MS2-GJR-std
0.03478 0.14137
32 MSGARCH: Markov-Switching GARCH Models in R

5. Conclusion
We introduced the R package MSGARCH which allows users to estimate, simulate, and
perform forecasts with Markov-switching GARCH models in the R statistical software. We
detailed how to create various single-regime and regime-switching specifications with different
scedastic functions and conditional distributions. We documented how to perform maximum
likelihood and Bayesian estimations of these models. In an empirical illustration with financial
returns, we showed how to estimate, select, and forecast with this class of models.
The R language has become an important vehicle for knowledge transfer in time series analysis
over the last years. We hope the R package MSGARCH will be useful for academics and
practitioners to improve their conditional volatility and density forecasts in the broad range
of applications involving regime changes in volatility dynamics.
Finally, if you use the R package MSGARCH, please cite the software in publications using
citation(package = "MSGARCH").

Computational details
The results in this paper were obtained using R 3.5.0 (R Core Team 2018) with the pack-
ages coda version 0.19-1 (Plummer et al. 2006), GAS version 0.2.6 (Ardia et al. 2019b),
numDeriv version 2016.8-1 (Gilbert and Varadhan 2016), MSGARCH version 2.3 (Ardia
et al. 2019a), Rcpp version 0.12.16 (Eddelbuettel et al. 2018; Eddelbuettel and François
2011), RcppArmadillo version 0.8.500.0 (Eddelbuettel et al. 2017; Eddelbuettel and Sanderson
2014), and zoo version 1.8-1 (Zeileis and Grothendieck 2005). Computations were performed
on Windows 10 x86 64-w64-mingw32/x64 (64-bit) with Intel(R) Xeon(R) CPU E5-2650 2x
2.30 GHz. Code for the computations is available in the R script v91i04.R, available in
the supplementary materials and also in the Examples folder of the GitHub repository at
https://github.com/keblu/MSGARCH.
The results in this paper rely on simulations from an adaptive MCMC scheme. Therefore,
results depend on the set.seed value and on the linear algebra library used, as the adaptive
scheme requires Cholesky decomposition and eigenvalue calculation. While this would have
a negligible impact for a single draw, the difference accumulates over the MCMC iterations,
and can lead to different MCMC posterior results than the ones reported in this paper. An
upper bound on the magnitude of those differences is difficult to evaluate. In the comparison
between Windows 10 and Ubuntu 18.04.3, as reported in the supplementary material and in
the folder Examples on the GitHub repository, we find differences of up to the third digit
for the MCMC results of Section 3.2 and up to the second digit for the MCMC results of
Section 4.
In the script v91i04-timing.R, we provide a timing comparison between the ML and MCMC
estimation strategies. The estimation of a two-regime MSGARCH model with normal inno-
vations on the SMI dataset (2,500 observations) is around ten times faster with the ML
approach than with the MCMC approach (2.8 seconds instead of 26.7 seconds). We also per-
form a comparison between the package MSGARCH and the R package bayesGARCH (Ardia
and Hoogerheide 2010) in the case of a single-regime GARCH(1, 1) model with normal in-
novations. We observe that MSGARCH is not only competitive in terms of computational
time but also for accuracy (measured with the effective number of draws). With the default
settings (5,000 burn-in draws, 10,000 MCMC draws and a thinning of 10, thus a posterior
 Journal of Statistical Software 33

sample of size 1,000), the estimation with MSGARCH is about ten times faster (6.5 against
69.1 seconds) and the effective number of draws is 3.1 times larger.
R (R Core Team 2018) itself and all packages used are available from the Comprehensive R
Archive Network (CRAN) at http://CRAN.R-project.org/. The MSGARCH version under
development is available on the website at http://keblu.github.io/MSGARCH/.

Acknowledgments
The authors are grateful to the Associate Editor (Rob J. Hyndman) and two anonymous
referees for useful comments. They also thank Carol Alexander, Samuel Borms, Muriel Buri,
Peter Carl, Dirk Eddelbuettel, Laurent Fastnacht, Félix-Antoine Fortin, Alexios Ghalanos,
Richard Gerlach, Lennart Hoogerheide, Eliane Maalouf, Brian Peterson, Enrico Schumann,
Tobias Setz, Max Tchirikov, conference participants at R/Finance 2017 (Chicago), the 37th
International Symposium on Forecasting (Cairns), useR! 2017 (Brussels), Quant Insights 2017
(London), MAFE 2018 (Madrid), eRum 2018 (Budapest), and seminar participants at HEC
Liège, Paris-Dauphine, and IAE-AMSE Aix-Marseille. We acknowledge Industrielle-Alliance,
International Institute of Forecasters (https://forecasters.org), Google Summer of Code
(https://summerofcode.withgoogle.com), FQRSC (http://www.frqsc.gouv.qc.ca, grant
2015-NP-179931), swissuniversities (https://www.swissuniversities.ch), IVADO (https:
//ivado.ca), and the Swiss National Science Foundation (http://www.snf.ch, grant 179281)
for their financial support, and Calcul Québec (https://www.calculquebec.ca) for compu-
tational support.

References

Akaike H (1974). “A New Look at the Statistical Model Identification.” IEEE Transactions
on Automatic Control, 19(6), 716–723. doi:10.1007/978-1-4612-1694-0_16.

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

Ardia D, Bluteau K, Boudt K, Catania L (2018). “Forecasting Risk with Markov-Switching

GARCH Models: A Large-Scale Performance Study.” International Journal of Forecasting,
34(4), 733–747. doi:10.1016/j.ijforecast.2018.05.004.

Ardia D, Bluteau K, Boudt K, Catania L, Ghalanos A, Peterson B, Trottier DA (2019a).

MSGARCH: Markov-Switching GARCH Models in R. R package version 2.31, URL https:
//CRAN.R-project.org/package=MSGARCH.

Ardia D, Boudt K, Catania L (2019b). “Generalized Autoregressive Score Models in R: The

GAS Package.” Journal of Statistical Software, 88(6), 1–28. doi:10.18637/jss.v088.i06.

Ardia D, Hoogerheide LF (2010). “Bayesian Estimation of the GARCH(1, 1) Model with

Student-t Innovations.” The R Journal, 2(2), 41–47. doi:10.32614/rj-2010-014.
34 MSGARCH: Markov-Switching GARCH Models in R

Ardia D, Hoogerheide LF, Van Dijk HK (2009). “Adaptive Mixture of Student-t Distributions
as a Flexible Candidate Distribution for Efficient Simulation: The R Package AdMit.”
Journal of Statistical Software, 29(3), 1–32. doi:10.18637/jss.v029.i03.

Ardia D, Kolly J, Trottier DA (2017). “The Impact of Parameter and Model Uncertainty
on Market Risk Predictions from GARCH-Type Models.” Journal of Forecasting, 36(7),
808–823. doi:10.1002/for.2472.

Bauwens L, Backer B, Dufays A (2014). “A Bayesian Method of Change-Point Estimation

with Recurrent Regimes: Application to GARCH Models.” Journal of Empirical Finance,
29, 207–229. doi:10.1016/j.jempfin.2014.06.008.

Black F (1976). “Studies of Stock Price Volatility Changes.” In Proceedings of the 1976 Meet-
ings of the Business and Economics Statistics Section, pp. 177–181. American Statistical
Association.

Bollerslev T (1986). “Generalized Autoregressive Conditional Heteroskedasticity.” Journal of

Econometrics, 31(3), 307–327. doi:10.1016/0304-4076(86)90063-1.

Brooks C, Burke SP, Persand G (2001). “Benchmarks and the Accuracy of GARCH
Model Estimation.” International Journal of Forecasting, 17(1), 45–56. doi:10.1016/
s0169-2070(00)00070-4.

Cai J (1994). “A Markov Model of Switching-Regime ARCH.” Journal of Business & Eco-
nomic Statistics, 12(3), 309–316. doi:10.2307/1392087.

Christie AA (1982). “The Stochastic Behavior of Common Stock Variances: Value, Leverage
and Interest Rate Effects.” Journal of Financial Economics, 10(4), 407–432. doi:10.1016/
0304-405x(82)90018-6.

Christoffersen PF (1998). “Evaluating Interval Forecasts.” International Economic Review,

39(4), 841–862. doi:10.2307/2527341.

Dueker MJ (1997). “Markov Switching in GARCH Processes and Mean-Reverting Stock-

Market Volatility.” Journal of Business & Economic Statistics, 15(1), 26–34. doi:10.
2307/1392070.

Eddelbuettel D (2019). CRAN Task View: Empirical Finance. Version 2019-03-07, URL
https://CRAN.R-project.org/view=Finance.

Eddelbuettel D, François R (2011). “Rcpp: Seamless R and C++ Integration.” Journal of

Statistical Software, 40(8), 1–18. doi:10.18637/jss.v040.i08.

Eddelbuettel D, François R, Allaire JJ, Ushey K, Kou Q, Bates D, Chambers J (2018).

Rcpp: Seamless R and C++ Integration. R package version 0.12.16, URL https://CRAN.
R-project.org/package=Rcpp.

Eddelbuettel D, François R, Bates D (2017). RcppArmadillo: Rcpp Integration for the

Armadillo Templated Linear Algebra Library. R package version 0.8.500.0, URL https:
//CRAN.R-project.org/package=RcppArmadillo.
 Journal of Statistical Software 35

Eddelbuettel D, Sanderson C (2014). “RcppArmadillo: Accelerating R with High-Performance

C++ Linear Algebra.” Computational Statistics & Data Analysis, 71, 1054–1063. doi:
10.1016/j.csda.2013.02.005.
Engle RF (1982). “Autoregressive Conditional Heteroscedasticity with Estimates of the Vari-
ance of United Kingdom Inflation.” Econometrica, 50(4), 987–1007. doi:10.2307/1912773.
Engle RF, Manganelli S (2004). “CAViaR: Conditional Autoregressive Value at Risk by
Regression Quantiles.” Journal of Business & Economic Statistics, 22(4), 367–381. doi:
10.1198/073500104000000370.
Fernández C, Steel MFJ (1998). “On Bayesian Modeling of Fat Tails and Skewness.” Journal
of the American Statistical Association, 93(441), 359–371. doi:10.1080/01621459.1998.
10474117.
Francq C, Zakoian JM (2011). GARCH Models: Structure, Statistical Inference and Financial
Applications. John Wiley & Sons. doi:10.1002/9780470670057.
Frühwirth-Schnatter S (2006). Finite Mixture and Markov Switching Models. 1st edition.
Springer-Verlag, New York. doi:10.1007/978-0-387-35768-3.
Geweke J (2007). “Interpretation and Inference in Mixture Models: Simple MCMC Works.”
Computational Statistics & Data Analysis, 51(7), 3529–3550. doi:10.1016/j.csda.2006.
11.026.
Ghalanos A (2017). rugarch: Univariate GARCH Models. R package version 1.3-8, URL
https://CRAN.R-project.org/package=rugarch.
Gilbert P, Varadhan R (2016). numDeriv: Accurate Numerical Derivatives. R package
version 2016.8-1, URL https://CRAN.R-project.org/package=numDeriv.
Glosten LR, Jagannathan R, Runkle DE (1993). “On the Relation Between the Expected
Value and the Volatility of the Nominal Excess Return on Stocks.” Journal of Finance,
48(5), 1779–1801. doi:10.1111/j.1540-6261.1993.tb05128.x.
Gray SF (1996). “Modeling the Conditional Distribution of Interest Rates as a Regime-
Switching Process.” Journal of Financial Economics, 42(1), 27–62. doi:10.1016/
0304-405x(96)00875-6.
Haas M, Mittnik S, Paolella MS (2004a). “A New Approach to Markov-Switching GARCH
Models.” Journal of Financial Econometrics, 2(4), 493–530.
Haas M, Mittnik S, Paolella MS (2004b). “Mixed Normal Conditional Heteroskedasticity.”
Journal of Financial Econometrics, 2(2), 211–250.
Hamilton JD (1989). “A New Approach to the Economic Analysis of Nonstationary Time
Series and the Business Cycle.” Econometrica, 57(2), 357–384. doi:10.2307/1912559.
Hamilton JD (1994). Time Series Analysis. Princeton University Press, Princeton.
Hamilton JD, Susmel R (1994). “Autoregressive Conditional Heteroskedasticity and Changes
in Regime.” Journal of Econometrics, 64(1–2), 307–333. doi:10.1016/0304-4076(94)
90067-1.
36 MSGARCH: Markov-Switching GARCH Models in R

Hoogerheide L, Van Dijk HK (2010). “Bayesian Forecasting of Value at Risk and Expected
Shortfall Using Adaptive Importance Sampling.” International Journal of Forecasting,
26(2), 231–247. doi:10.1016/j.ijforecast.2010.01.007.

Hyndman RJ (2019). CRAN Task View: Time Series Analysis. Version 2019-08-25, URL
https://CRAN.R-project.org/view=TimeSeries.

Kastner G (2016). “Dealing with Stochastic Volatility in Time Series Using the R Package
stochvol.” Journal of Statistical Software, 69(5), 1–30. doi:10.18637/jss.v069.i05.

Kim CJ, Nelson CR (1999). “Has the U.S. Economy Become More Stable? A Bayesian
Approach Based on a Markov-Switching Model of the Business Cycle.” Review of Economics
and Statistics, 81(4), 608–616. doi:10.1162/003465399558472.

Klaassen F (2002). “Improving GARCH Volatility Forecasts with Regime-Switching

GARCH.” In Advances in Markov-Switching Models, pp. 223–254. Springer-Verlag, Berlin
Heidelberg. doi:10.1007/978-3-642-51182-0_10.

Lamoureux CG, Lastrapes WD (1990). “Persistence in Variance, Structural Change, and

the GARCH Model.” Journal of Business & Economic Statistics, 8(2), 225–234. doi:
10.2307/1391985.

McNeil AJ, Frey R, Embrechts P (2015). Quantitative Risk Management: Concepts, Tech-
niques and Tools. 2nd edition. Princeton University Press.

Mullen KM, Ardia D, Gil DL, Windover D, Cline J (2011). “DEoptim: An R Package for
Global Optimization by Differential Evolution.” Journal of Statistical Software, 40(6), 1–26.
doi:10.18637/jss.v040.i06.

Nelson DB (1991). “Conditional Heteroskedasticity in Asset Returns: A New Approach.”

Econometrica, 59(2), 347–370. doi:10.2307/2938260.

Nobel Media (2003). “The Prize in Economic Sciences 2003 – Press Release.” URL
https://www.nobelprize.org/nobel_prizes/economic-sciences/laureates/2003/
press.html.

Papastamoulis P (2016). “label.switching: An R Package for Dealing with the Label Switching
Problem in MCMC Outputs.” Journal of Statistical Software, 69(1), 1–24. doi:10.18637/
jss.v069.c01.

Plummer M, Best N, Cowles K, Vines K (2006). “coda: Convergence Diagnosis and Output
Analysis for MCMC.” R News, 6(1), 7–11. URL https://www.R-project.org/doc/Rnews/
Rnews_2006-1.pdf.

Pretis F, Reade JJ, Sucarrat G (2018). Automated General-to-Specific (GETS) Regression

Modeling and Indicator Saturation for Outliers and Structural Breaks. doi:10.18637/jss.
v086.i03.

R Core Team (2018). R: A Language and Environment for Statistical Computing. R

Foundation for Statistical Computing, Vienna, Austria. R version 3.5.0, URL https:
//www.R-project.org/.
 Journal of Statistical Software 37

Schwarz G (1978). “Estimating the Dimension of a Model.” The Annals of Statistics, 6(2),
461–464.

Sousa T, Otiniano C, Lopes S, Diethelm W (2015). GEVStableGarch: ARMA-

GARCH/APARCH Models with GEV and Stable Distributions. R package version 1.1,
URL https://CRAN.R-project.org/package=GEVStableGarch.

Spiegelhalter DJ, Best NG, Carlin BP, Van der Linde A (2002). “Bayesian Measures of
Model Complexity and Fit.” Journal of the Royal Statistical Society B, 64(4), 583–639.
doi:10.1111/1467-9868.00353.

Sucarrat G (2015). lgarch: Simulation and Estimation of Log-GARCH Models. R package

version 0.6-2, URL https://CRAN.R-project.org/package=lgarch.

Teräsvirta T (2009). “An Introduction to Univariate GARCH Models.” In Handbook of

Financial Time Series, pp. 17–42. Springer-Verlag, Berlin Heidelberg. doi:10.1007/
978-3-540-71297-8_1.

Trapletti A, Hornik K (2017). tseries: Time Series Analysis and Computational Finance.
R package version 0.10-42, URL https://CRAN.R-project.org/package=tseries.

Trottier DA, Ardia D (2016). “Moments of Standardized Fernandez-Steel Skewed Distribu-

tions: Applications to the Estimation of GARCH-Type Models.” Finance Research Letters,
18, 311–316. doi:10.1016/j.frl.2016.05.006.

Vihola M (2012). “Robust Adaptive Metropolis Algorithm with Coerced Acceptance Rate.”
Statistics and Computing, 22(5), 997–1008. doi:10.1007/s11222-011-9269-5.

Viterbi A (1967). “Error Bounds for Convolutional Codes and an Asymptotically Optimum
Decoding Algorithm.” IEEE Transactions on Information Theory, 13(2), 260–269. doi:
10.1109/tit.1967.1054010.

Wuertz D, Chalabi Y, Miklovic M, Boudt K, Chausse P (2016). fGarch: Rmetrics – Au-

toregressive Conditional Heteroskedastic Modelling. R package version 3010.82.1, URL
https://CRAN.R-project.org/package=fGarch.

Zakoian JM (1994). “Threshold Heteroskedastic Models.” Journal of Economic Dynamics &

Control, 18(5), 931–955. doi:10.1016/0165-1889(94)90039-6.

Zeileis A, Grothendieck G (2005). “zoo: S3 Infrastructure for Regular and Irregular Time
Series.” Journal of Statistical Software, 14(6), 1–27. doi:10.18637/jss.v014.i06.

Zhang Y, Wang J, Wang X (2014). “Review on Probabilistic Forecasting of Wind Power

Generation.” Renewable and Sustainable Energy Reviews, 32, 255–270. doi:10.1016/j.
rser.2014.01.033.
38 MSGARCH: Markov-Switching GARCH Models in R

Affiliation:
David Ardia (corresponding author)
Department of Decision Sciences
HEC Montréal
3000, Ch. de la Côte-Sainte-Catherine, Montréal (Québec) H3T 2A7, Canada
E-mail: [email protected]

Keven Bluteau
Institute of Financial Analysis
University of Neuchâtel
Rue A.-L. Breguet 2, 2000 Neuchâtel, Switzerland
and
Solvay Business School
Vrije Universiteit Brussel, Belgium
E-mail: [email protected]

Kris Boudt
Department of Economics
Ghent University
Sint-Pietersplein 5, 9000 Gent, Belgium
and
Solvay Business School
Vrije Universiteit Brussel, Belgium and
Econometrics and Finance Department
Vrije Universiteit Amsterdam, The Netherlands
E-mail: [email protected]

Leopoldo Catania
Department of Economics and Business Economics
Aarhus University & CREATES
Fuglesangs Allé 4, 8210, Aarhus V, Denmark
E-mail: [email protected]

Denis-Alexandre Trottier
Faculty of Business Administration
Laval University
Pavillon Palasis-Prince
2325, rue de la Terrasse, Québec, G1V 0A6, Canada
E-mail: [email protected]

Journal of Statistical Software http://www.jstatsoft.org/

published by the Foundation for Open Access Statistics http://www.foastat.org/
October 2019, Volume 91, Issue 4 Submitted: 2017-10-30
doi:10.18637/jss.v091.i04 Accepted: 2018-06-23