Heavy-Tailed Innovations in the R Package stochvol

Gregor Kastner
WU Vienna University of Economics and Business

First version: May 15, 2015

This version: Aug 24, 2016

Abstract
We document how sampling from a conditional Student’s t distribution is implemented
in stochvol. Moreover, a simple example using EUR/CHF exchange rates illustrates how
to use the augmented sampler. We conclude with results and implications.

Keywords: Student’s t distribution, data augmentation, EUR/CHF exchange rates.

Preface
This note serves as a preliminary add-on to the more elaborate article “Dealing with Stochastic
Volatility in Time Series using the R package stochvol” (Kastner 2016a). It discusses and
relaxes the restriction to conditionally normal errors in the vanilla stochastic volatility (SV)
model.

1. The SV model with Student’s t errors

Several authors have suggested to use non-normal conditional innovation distributions for
stochastic volatility modeling. Examples include the Student’s t distribution (Harvey, Ruiz,
and Shephard 1994), the extended Generalized Inverse Gaussian (Silva, Lopes, and Migon
2006), (semi-)parametric innovations (Jensen and Maheu 2010; Delatola and Griffin 2011),
or the GH skew Student’s t distribution (Nakajima and Omori 2012). In the following, we
describe how the estimation of the SV model with Student’s t errors is implemented in the R
(R Core Team 2016) package stochvol (Kastner 2016b).
Let y = (y1 , y2 , . . . , yn )> be a vector of returns with mean zero. The SV model with Student’s
t errors (in short SV-t) is given through

yt |ht , ν ∼ tν (0, exp ht ) , (1)

µ), ση2


ht |ht−1 , µ, φ, ση ∼ N µ + φ(ht−1 − , (2)
µ, ση2 /(1 2


h0 |µ, φ, ση ∼ N −φ ) , (3)

i.e., conditionally on ht , the data is assumed to follow a zero-mean non-standardized Student’s

t distribution with ν degrees of freedom and variance (ν exp ht )/(ν − 2) for ν > 2. Following
Chib, Nardari, and Shephard (2002), we assume that a priori the degrees of freedom parameter
 2 Heavy-Tailed Innovations in the R Package stochvol

ν ∼ U(a, b), i.e., follows a uniform distribution with support on the real interval (a, b). All
other prior components are chosen as in Kastner (2016a).

2. Usage
Estimating a stochastic volatility model with conditional t errors via stochvol is very similar to
estimating a model with standard Gaussian errors, differing only through specifying a non-NA
argument priornu. This triggers the sampler specified in Section 3. To provide an example,
we investigate the historical daily EUR/CHF exchange rates and display these in Figure 1.

R> library(stochvol)
R> data(exrates)
R> par(mfrow = c(2, 1), mar = c(1.7, 1.7, 1.7, 0.1), mgp = c(1.6, 0.6, 0))
R> plot(exrates$date, exrates$CHF, type = 'l', main = 'Price of 1 EUR in CHF')
R> dat <- logret(exrates$CHF, demean = TRUE)
R> plot(exrates$date[-1], dat, type = 'l', main = 'Demeaned log returns')

Price of 1 EUR in CHF

1.71.5
exrates$CHF
1.3 1.1

2000 2002 2004 2006 2008 2010 2012

exrates$date
Demeaned log returns
0.06
dat
0.02 −0.02

2000 2002 2004 2006 2008 2010 2012

Figure 1: Levels and demeaned log returns of EUR in CHF.

By specifying the argument priornu (a two-element vector containing the lower and upper
bounds of the uniform prior for ν), we can trigger the sampler to allow for heavy-tailed
conditional innovations.

R> rest <- svsample(dat, priormu = c(-12, 1), priorphi = c(20, 1.1),
+ priorsigma = 0.1, priornu = c(2, 100), burnin = 2000)
R> plot(rest, showobs = FALSE)
 Gregor Kastner 3

Results are displayed in Figure 2, containing the output from the SV-t model. Row 1 depicts
p t /2) for t ∈ {1, . . . , n}; row 2 shows the time varying standard deviations given through
exp(h
ν/(ν − 2) exp(ht /2) for t ∈ {1, . . . , n}; row 3 portrays trace plots and row 4 outlines the
corresponding smoothed kernel density estimates for the four parameters µ, φ, σ, and ν. It
is worth noting that ν is estimated to lie between 6 and 18 with high posterior probability,
indicating evidence for the presence of heavy tails even after catering for stochastic volatility.
The extra flexibility of the SV-t sampler seems to allow for increased persistence φ and smaller
variance of log-volatility σ 2 , resulting in smoother time-varying volatility estimates.

Estimated conditional volatilities in percent (5% / 50% / 95% posterior quantiles)

1.5
1.0
0.5
0.0

0 500 1000 1500 2000 2500 3000

Estimated volatilities in percent (5% / 50% / 95% posterior quantiles)

2.0
1.5
1.0
0.5
0.0

0 500 1000 1500 2000 2500 3000

Trace of mu (thinning = 1) Trace of phi (thinning = 1) Trace of sigma (thinning = 1) Trace of nu (thinning = 1)
1.000
−10

16
0.995
−11

0.18

14
0.990

12
−12

10
0.14
0.985
−13

8
−14

0.980

6
0.10

2000 4000 6000 8000 10000 12000 2000 4000 6000 8000 10000 12000 2000 4000 6000 8000 10000 12000 2000 4000 6000 8000 10000 12000

Density of mu Density of phi Density of sigma Density of nu

25
1.2

120

0.20
1.0

20

0.15
0.8

15
80
0.6

0.10
60

10
0.4

40

0.05
5
0.2

20

0.00
0.0

−14 −13 −12 −11 −10 0.980 0.985 0.990 0.995 1.000 0.10 0.12 0.14 0.16 0.18 0.20 0.22 4 6 8 10 12 14 16 18

Figure 2: Standard output of the plot method when applied to an svdraws object containing
posterior draws from an SV model with Student’s t errors.

We investigate one-day-ahead out-of-sample predictive performance of an AR(1) model with

(a) homoskedastic, (b) SV, (c) SV-t, (d) GARCH(1,1) errors, applied to the raw exchange
rate data. Details about this procedure are provided in Chapter 5 of Kastner (2016a). The
results, summarized in Figure 3, speak in favor of the SV-t model for this dataset.
4 Heavy-Tailed Innovations in the R Package stochvol

Observed and predicted residuals

● Observed residuals
●
1% and 99% one−day−ahead predictive quantiles: SV
1% and 99% one−day−ahead predictive quantiles: SVt
1% and 99% one−day−ahead predictive quantiles: GARCH
1% and 99% one−day−ahead predictive quantiles: homoskedastic

0.05
●

●●
●

● ●
● ●● ●
● ●● ● ●
●
●● ● ● ● ● ●● ●
● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ●
● ●● ●●● ● ● ●
● ● ●● ● ●
●
●● ●
●●● ● ● ●● ●● ● ●
●
● ●● ●
● ● ● ● ●●● ●●●●● ● ● ● ●●●
●
● ● ●● ● ● ● ● ●● ●● ●● ●● ● ● ●● ● ● ●
●
●● ● ●●●● ● ● ●● ● ●● ● ● ●● ●●● ●● ● ●●● ● ● ●●●
●●● ●● ● ●
●
● ●●● ● ●● ● ● ●●●●●
●
●●●
● ●
●●● ●●● ●● ●
●● ● ●● ●●
● ● ●
● ● ● ● ● ● ●●● ● ● ● ●● ●● ●● ● ● ● ●●●● ● ●● ● ● ● ●
●● ●● ●● ● ●
● ●● ● ● ●● ●
● ● ●●●● ● ●●●● ● ●
●● ●●
●● ● ● ●● ●●●● ●●● ●
●●
● ●● ●●●●●●● ●
●●●●
●●
● ●● ●
●●●●
●● ●● ●●
● ●● ●
● ●● ● ● ●●● ●
● ●
●● ●●●●●● ●
●● ● ●
●● ●●● ● ● ● ●● ●●●● ● ●●● ● ● ● ●● ● ● ● ●●
●● ●●
●
●
●●
●
●
●
●●
● ●●
●●●●●● ●●●●●●
●
●●●●
●●
●● ●
●● ●
●●
●●
●●
●●
●●●●
● ●●
●
●
●
●
●
●
●●●
●● ●●
●
●
●
●
●
● ●●
●●
●
●
●
●●
●
●
●
●●
●
●●●●
● ●●●●
●●
●●●
●●
●● ●
●●●●●●
●●
●
●
●●
●
●
●●
●
●●●●
●●● ●
●
●
●●
●
●
●●
●●
●●
●●
●
●●
●●●●●●●
●●●
●
●
●
●
●●●
●
●
●
●
●
●
●●●
● ●
●
●
●
●● ●
●●●
●
●
●
●●
●
●●
●●●●●●
●
●●
●●●● ●
●
●●●
● ●●●
●●
●●●●
●● ●●●● ●
●●
● ●●
●●● ●
●●●●●
●●●
●
● ●●
●
●● ●● ● ●●
● ●● ● ●●● ●●● ●●●●
●●●●●
●●●●●●● ●●
●
●● ● ●
●● ●
●● ●● ●
●●●●● ● ●●
●●●
●●
●
●●
●
●
●● ●● ● ● ●●● ●●
●● ●● ●●● ●●●● ●● ●●●
● ●●●
●
●
●● ●
0.00 ●●●
● ●
●●●● ●●●●
●●
●●●
● ●● ●●●
● ●
● ●
●●●●
● ●●
● ●●
● ●
●●●
●●
●●●●●
●●●
●●●●●●●
●
● ●●
●●●
●●● ●
●
●
●●●
●●
●● ●●●●●●
●●●●● ●
●●
●
●●●
●●● ●
●●● ●
●
●●●
●●
●●● ●
●●
●
●
●●●●●●● ●
● ●● ●
●●●●
●● ●●
●●●●●●●
●●
●●●● ● ●●●●●
●●● ●● ● ●
●● ●●
● ●●
●●●●●●●●●●●● ●●
●● ●● ●
● ●
●●
●●
●●
●
● ●
●● ●●
●●
●●
●●
● ●
●
●●●●● ●●●●●●● ● ●
●● ●
● ●●●● ●● ●●●● ●
●● ●●
●●●
●
●●
●●
●
●●
●●
●●
●
●●
●●
●● ●
●●●●
●●
●●
●●
● ●● ●
●●
●
●●
●
●
●●
●
●
●●
●●●
●●
● ●
●●●
●●●
●●● ●●●
●●●
●●
● ●●●
●●
●●
●●
●
●●●
● ●● ● ●
●●
●●
●●●
●●●
●
●●●
●●
●
●●●
●●●
●●●
●●
●●
●
●
● ●
●●
●
●●●●●
● ●
●●●
●●
●●●
●
●●●
● ●●
●
●
●
●●
●●
●
●●●
●●●●
● ●●●●
● ●●●● ●●●
●
●
●●
●
●●●
●
●
●● ●● ●
●
●● ●● ●●●●● ●●●●●● ●●●
● ●●
●
●●
●●
●
●●
●
●●● ●●●●●
● ●●●●
●●●●● ●●●
●
●
●
●
●
●
●●
●●
● ●
●●
●●
●●
●
●●
●
●
●●●
●
●●
●
●
● ●
●●
●●
●●●
●●
●●
●●
●●
● ●
●
●●
● ● ● ●● ●
●●●●●
●●●●
●● ●●●●●
●●●●
●● ●●●
●●●●
●●●
●●
●●
● ●
●●●
●●●
●
● ●●
●●
●●
●
●●
●
●●
●
●●●● ●
● ● ●●● ●●●●
●● ●● ●●●
●
●
●●●
●
●●
● ●● ●
● ●●●● ●● ●●●●●●●
● ● ●
● ●
●● ●
● ●● ●● ●
●● ●●●
●
●
● ●●●●●●●● ●●
●●● ●
● ●●●
● ● ● ●●●
●● ●● ●●● ●● ● ●●●
●● ● ● ●●
● ●
●
● ●
● ●
● ●●●
●●
●
●●
●●
● ●
● ●
●
● ●
●● ● ●●●●●●●●
●
●●
● ●
●●
●●●
● ●●
●● ●●●● ●● ●
●
●
●
●
●● ●
●●
●●●
● ● ● ●
● ● ● ● ● ● ●● ●● ●●● ● ● ● ●● ● ● ●
● ● ● ● ● ● ● ● ● ●● ● ●
●● ●● ●●●● ●●● ●●● ●● ●●
● ●●●● ● ●●●● ● ●●● ●
● ● ●
●●● ● ● ●● ● ● ●
●●●● ●●
● ●●●●● ●
● ● ●●●● ●
●● ●
● ●●● ● ● ● ●●
● ● ●
●● ●●●●●● ●●●
● ●
●
●●●● ● ●● ●●
● ●●
●●●●●● ● ● ●
● ● ● ● ● ●●●● ●
● ● ● ●●● ●● ● ●● ●● ● ● ●●● ● ● ● ●●
● ● ● ●● ●
●● ●●● ● ● ● ●
● ● ● ● ●● ●
● ● ● ●● ●● ● ● ● ●●●●●
● ● ●
● ● ● ● ● ● ● ●
● ●● ● ● ● ● ●● ●
● ● ● ●●● ● ● ●●●
● ● ●
● ● ●
● ●●
●
●
● ●

−0.05

2003−12−03 2004−12−17 2006−01−02 2007−01−18 2008−02−05 2009−02−23 2010−03−10 2011−03−23 2012−04−04

Cumulative log predictive Bayes factors in favor of heteroskedasticity
SV
SVt
1000 GARCH

800

600

400

200

2003−12−03 2004−12−17 2006−01−02 2007−01−18 2008−02−05 2009−02−23 2010−03−10 2011−03−23 2012−04−04

Figure 3: Log predictive one-day-ahead Bayes factors in favor of SV, SV-t, and GARCH errors
over the homoskedastic model. The final log predictive Bayes factors aggregate to 1107.44
(SV), 1115.36 (SV-t), and 1033.53 (GARCH), respectively, thus providing strong evidence for
the SV-t model.
 Gregor Kastner 5

3. Necessary modifications in the sampling scheme

The Student’s t distribution appearing in Equation 1 can be conveniently expressed as a scale
mixture of normal distributions,
yt |ht , τt ∼ N (0, τt exp ht ) ,
τt |ν ∼ G −1 (ν/2, ν/2) ,
where N µ, ση2 denotes the normal distribution with mean µ and variance ση2 , and G −1 (a, b)

denotes the inverse gamma distribution with shape and scale parameters a and b, respectively.
√
Treating τ = (τ1 , . . . , τn )> as latent data and letting ỹt = yt / τt for t ∈ {1, . . . , n}, we have
ỹt |ht ∼ N (0, exp ht ) ,
and the AWOL sampler described in Kastner and Frühwirth-Schnatter (2014) can directly be
applied to the transformed data. To obtain draws from the newly introduced variables τ and
ν, two additional steps are required.

3.1. Sampling the auxiliary variables

It is easy to see that the marginal posterior is given through
2
 
−1 ν + 1 ν + yt exp(−ht )
τt |yt , ht , ν ∼ G , ,
2 2
independently for each t ∈ {1, . . . , n}. Obtaining draws from this distribution is straightfor-
ward.

3.2. Sampling the degrees of freedom parameter

The full conditional posterior of the degrees of freedom parameter, ν|·, only depends on τ .
Its density is given through the product of n univariate truncated inverse gamma densities
which may be written as
n
!−ν/2 ( n
)
 ν nν/2  ν −n Y νX 1
p(ν|·) = p(ν|τ ) ∝ Γ τt exp − (4)
2 2 2 τt
t=1 t=1

for ν ∈ (a, b) and zero elsewhere.

For obtaining draws from this distribution, we use an independence Metropolis-Hastings up-
date. We follow Chib and Greenberg (1994), who introduced the idea of specifying an inde-
pendence proposal through numerical maximization of the log-density. For the problem at
hand, we consequently aim for optimizing
n  
nν νX 1
log p(ν|τ ) = log(ν/2) − n log Γ(ν/2) − log τt + + C, (5)
2 2 τt
t=1

with first and second derivatives given through

n  
∂ log p(ν|τ ) n (0)
 1X 1
= 1 + log(ν/2) − ψ (ν/2) − log τt + , (6)
∂ν 2 2 τt
t=1
∂ 2 log p(ν|τ ) n n
= − ψ (1) (ν/2), (7)
∂ν 2 2ν 4
6 Heavy-Tailed Innovations in the R Package stochvol

where ψ (m) denotes the polygamma function of order m. Using the above, it is easy to
numerically find

ν̂ = arg max log p(ν|τ ),

ν
 2 
∂ log p(ν|τ ) 
Bν̂ = −1 ,
∂ν 2 
ν=ν̂

and a proposal candidate νprop may be drawn from a normal distribution with mean ν̂ and
variance Bν̂ (the Laplace approximation). Letting φ(x|ν̂, Bν̂ ) denote the corresponding den-
sity function, the acceptance probability is equal to min{1, R} with

p(νprop |τ ) φ(νold |ν̂, Bν̂ )

R= × .
p(νold |τ ) φ(νprop |ν̂, Bν̂ )

4. Conclusion
We have shown how a simply data augmentation trick can be utilized to generalize the core
sampler in stochvol in order to cater for potentially heavier-tailed innovation distributions.
However, several caveats are called for:

• Even though the uniform prior for ν has been used widely, more robust alternatives are
probably preferred, cf. Frühwirth-Schnatter and Pyne (2010) and the references therein.

• Sampling the degrees of freedom parameter ν conditionally on τ can be very inefficient

if ν becomes large (and thus the plain vanilla SV model suffices). We recommend to
resort to the original SV sampler in this case.

• Leaving aside the additional computational burden, it is trivial to incorporate this ex-
tension into samplers employing stochvol as part of a larger MCMC scheme (e.g. Huber
2014; Kastner, Frühwirth-Schnatter, and Lopes 2014; Dovern, Feldkircher, and Huber
2015). Nevertheless, at the current stage of development, this should be conducted with
caution by carefully investigating the convergence of the posterior draws.

Acknowledgments
The author would like to thank the attendants of the Spring 2015 Brown Bag Seminar of the
Institute for Statistics and Mathematics, WU Vienna University of Economics and Business,
in particular Sylvia Frühwirth-Schnatter, Mark Jensen, and Kurt Hornik, for stimulating
comments and suggestions.

References

Chib S, Greenberg E (1994). “Bayes Inference in Regression Models with ARMA(p, q) Errors.”
Journal of Econometrics, 64, 183–206. doi:10.1016/0304-4076(94)90063-9.
 Gregor Kastner 7

Chib S, Nardari F, Shephard N (2002). “Markov Chain Monte Carlo Methods for Stochastic
Volatility Models.” Journal of Econometrics, 108, 281–316. doi:10.1016/S0304-4076(01)
00137-3.

Delatola EI, Griffin JE (2011). “Bayesian Nonparametric Modelling of the Return Distribution
with Stochastic Volatility.” 6, 901–926. doi:10.1214/11-BA632.

Dovern J, Feldkircher M, Huber F (2015). “Does Joint Modeling of the World Economy
Pay Off? Evaluating GVAR Forecasts from a Multivariate Perspective.” Discussion Paper
Series 590, University of Heidelberg, Department of Economics. URL http://www.ub.
uni-heidelberg.de/archiv/18586.

Frühwirth-Schnatter S, Pyne S (2010). “Bayesian inference for finite mixtures of univariate

and multivariate skew-normal and skew-t distributions.” Biostatistics, 11(2), 317–336.
doi:10.1093/biostatistics/kxp062.

Harvey AC, Ruiz E, Shephard N (1994). “Multivariate Stochastic Variance Models.” The
Review of Economic Studies, 61(2), 247–264. doi:10.2307/2297980.

Huber F (2014). “Density Forecasting using Bayesian Global Vector Autoregressions with
Common Stochastic Volatility.” Department of Economics Working Paper Series 179, WU
Vienna University of Economics and Business. URL http://epub.wu.ac.at/id/eprint/
4280.

Jensen MJ, Maheu JM (2010). “Bayesian semiparametric stochastic volatility modeling.”

Journal of Econometrics, 157(2), 306–316. doi:10.1016/j.jeconom.2010.01.014.

Kastner G (2016a). “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.

Kastner G (2016b). stochvol: Efficient Bayesian Inference for Stochastic Volatility (SV)
Models. R package version 1.2.3, URL http://CRAN.R-project.org/package=stochvol.

Kastner G, Frühwirth-Schnatter S (2014). “Ancillarity-Sufficiency Interweaving Strategy

(ASIS) for Boosting MCMC Estimation of Stochastic Volatility Models.” Computational
Statistics & Data Analysis, 76, 408–423. doi:10.1016/j.csda.2013.01.002.

Kastner G, Frühwirth-Schnatter S, Lopes HF (2014). “Analysis of Exchange Rates via Mul-

tivariate Bayesian Factor Stochastic Volatility Models.” In E Lanzarone, F Ieva (eds.), The
Contribution of Young Researchers to Bayesian Statistics – Proceedings of BAYSM2013,
volume 63 of Springer Proceedings in Mathematics & Statistics, pp. 181–185. Springer-
Verlag. doi:10.1007/978-3-319-02084-6_35.

Nakajima J, Omori Y (2012). “Stochastic Volatility Model with Leverage and Asymmetrically
Heavy-Tailed Error Using GH Skew Student’s t-Distribution.” 56(11), 3690–3704. doi:
10.1016/j.csda.2010.07.012.

R Core Team (2016). R: A Language and Environment for Statistical Computing. R Founda-
tion for Statistical Computing, Vienna, Austria. URL http://www.R-project.org/.
8 Heavy-Tailed Innovations in the R Package stochvol

Silva RS, Lopes HF, Migon HS (2006). “The Extended Generalized Inverse Gaussian Distri-
bution for Log-Linear and Stochastic Volatility Models.” Brazilian Journal of Probability
and Statistics, 20(1), 67–91.

Affiliation:
Gregor Kastner
Institute for Statistics and Mathematics
Department of Finance, Accounting and Statistics
WU Vienna University of Economics and Business
Welthandelsplatz 1, Building D4, Level 4
1020 Vienna, Austria
Telephone: +43/1/31336-5593
Fax: +43/1/31336-90-5593
E-mail: [email protected]
URL: http://statmath.wu.ac.at/~kastner/