Forecasting Bitcoin Risk Measures: A Robust Approach

Forecasting Bitcoin risk measures: A robust approach
Carlos Trucı́os∗
São Paulo School of Economics - FGV, Brazil
December 12, 2018
Abstract
Over the last few years, Bitcoin and other cryptocurrencies have attracted the in-
terest of many investors, practitioners and researchers. However, little attention has
been paid to the predictability of their risk measures. In this paper we compare the
predictability of the one-step-ahead volatility and Value-at-Risk of Bitcoin using sev-
eral volatility models. We also include procedures that take into account the presence
of outliers and estimate the volatility and Value-at-Risk in a robust fashion. Our re-
sults show that robust procedures outperform the non-robust ones when forecasting
the volatility and estimating the Value-at-Risk. These results suggest that the pres-
ence of outliers play an important role in the modelling and forecasting of Bitcoin risk
measures.
Keywords: Cryptocurrency, GARCH, Model confidence set, Outliers, Realised volatility,

Value-at-Risk.
JEL Classification: C51, C52, C53
2010 Mathematics Subject Classification: 62M10, 62F40, 62G35, 62P20, 91B84
∗
The author acknowledges financial support from São Paulo Research Foundation (FAPESP) grants
2016/18599-4 and 2018/03012-3 as well as the support of the Centre of Quantitative Studies in Eco-
nomics and Finance (CEQEF) and Centre for Applied Research on Econometrics, Finance and Statistics
(CAREFS). The author also thanks helpful comments and suggestions of João H. G. Mazzeu and Luiz K.
Hotta
1
Electronic copy available at: https://ssrn.com/abstract=3189446
1 Introduction
Since its creation in 2008, Bitcoin has attracted the interest of many investors, practitioners
and researchers. This interest has grown quickly over the last years, probably due to its
decentralised nature and by its large profits; see Nakamoto (2008).
Previous studies, such as Sapuric & Kokkinaki (2014), Baek & Elbeck (2015) and Briere
et al. (2015), have observed that Bitcoin is highly volatile. Thus, forecasting the volatility
and estimating the Value-at-Risk(VaR) as better as possible is crucial for good decisions of
investors and practitioners.
Bitcoin daily volatility has been previously studied by Dyhrberg (2016), Balcilar et al.
(2017), Chu et al. (2017), Katsiampa (2017), Liu et al. (2017), Pichl & Kaizoji (2017),
Naimy & Hayek (2018), Catania et al. (2018) and Conrad et al. (2018) among others. How-
ever, most of these studies have been focused on the in-sample analysis and the comparisons
have been made based on information criteria. In this context, Katsiampa (2017) estimates
the Bitcoin volatility using several GARCH-type models assuming Gaussian errors and con-
cludes that the best model to estimate the volatility is the AR(1)-CGARCH(1,1). analyse
Bitcoin and six other cryptocurrencies using GARCH-type models with different error dis-
tributions and conclude that the best models to estimate the Bitcoin volatility are the
IGARCH and GJRGARCH models with Gaussian distribution. Liu et al. (2017) compare
the GARCH model assuming the normal reciprocal inverse Gaussian (NRIG) distribution
against the Gaussian and Student-t error distributions and conclude that the GARCH model
with Student-t errors better estimates the volatility. Charles & Darné (2018) replicate the
study of Katsiampa (2017) and additionally take into account the presence of extreme ob-
servations. They find that, when using the jump-filtered returns as in Laurent et al. (2016),
the AR(1)-GARCH(1,1) model reports smaller information criteria than the models consid-
ered in Katsiampa (2017). On the other hand, Conrad et al. (2018) compare the classical
GARCH(1,1) with the GARCH-MIDAS model of Engle et al. (2013) and conclude that the
latter model outperforms the former.
In the out-of-sample context, Naimy & Hayek (2018) compare the one-step-ahead volatil-
ity forecasts estimated by GARCH and EGARCH models with Gaussian, Student-t and gen-
eralised error distributions. The authors compare the predicted volatility with the realised
2
volatility using the root mean square error (RMSE), mean absolute error (MAE) and mean
absolute percentage error (MAPE) and they conclude that EGARCH models present the
best performance. Catania et al. (2018) compare the Gaussian GARCH model with the gen-
eralised autoregressive score (GAS) models of Creal et al. (2013) and Harvey (2013). Their
comparison is based on the Quasi–Like (QLIKE) loss function and the predicted volatility is
compared with the squared observed returns. They conclude that the GARCH model is out-
performed by the GAS models. Peng et al. (2018) compare the volatility forecast obtained
by GARCH, EGARCH and GJR models assuming symmetric and asymmetric Gaussian
and Student-t errors against the Support Vector Regression GARCH model of Bezerra &
Albuquerque (2017) and they find that the latter yields more accurate forecasts. Although
out-of-sample comparisons are available in the literature, most of them are restrictive since
they do not consider models which presented better performance in previous studies, and,
additionally, they leave out several error distributions and models of the GARCH family.
In this sense, a comprehensive out-of-sample comparison is needed.
Nevertheless, most of the papers available in the literature do not consider the presence
of outliers, which, as mentioned by, for instance, Carnero et al. (2012), Boudt et al. (2013)
and Trucı́os & Hotta (2016), can affect drastically the volatility forecast and VaR estimation.
As far as we know, only Catania & Grassi (2017), Charles & Darné (2018) and Catania et al.
(2018) take into account the presence of outliers to estimate the Bitcoin volatility. Only
Catania et al. (2018) assess the Bitcoin volatility in an out-of-sample context, but they
only compare a Gaussian GARCH model with GAS models. In a VaR context, Chu et al.
(2017), Chan et al. (2017), Osterrieder & Lorenz (2017), Stavroyiannis (2018) and Gkillas
& Katsiampa (2018) estimate the VaR of Bitcoin. However, none of these works consider
the presence of outliers in a context of conditional heteroscedastic models.
The contribution of this paper is threefold. First, we carry out an extensive out-of-sample
comparison of the Bitcoin volatility forecast using several classical GARCH-type models
with different error distributions, filling, then, a gap in the literature and also summarising
the GARCH-type results found in the previous works. Second, we address the presence
of outliers and show that a better volatility forecast is obtained using robust procedures,
suggesting then, that outliers play a crucial role when modelling and forecasting the volatility
3
of Bitcoin. Finally, we compare the performance of the usual VaR estimation with the robust
bootstrap procedure proposed by Trucı́os et al. (2017).
The rest of the paper is organised as follows: Section 2 describes the models, realised
measures and loss functions used in the out-of-sample comparison. Section 3 describes the
data and reports the main findings. Finally, Section 4 presents the conclusions and future
researches.
2 Methodology
In this section, we briefly describe the volatility models used to forecast the Bitcoin volatility
as well as the realised measures used as volatility proxies. At the end of the section, we also
describe the loss functions used to evaluate the out-of-sample performance and the robust
bootstrap procedure of Trucı́os et al. (2017) to estimate the VaR.
2.1 Volatility models
Several approaches to model and forecast the volatility have been proposed in the literature;
see, for instance, Francq & Zakoian (2011), Bauwens et al. (2012) and Harvey (2013). In
this section, we briefly introduce the procedures used in this paper to forecast the daily
Bitcoin volatility. Subsection 2.1.1 introduces the classical GARCH-type models while the
subsequent subsections introduce alternative models.
2.1.1 GARCH models
The GARCH model has been widely used for researchers and practitioners to model and
forecast the second-order moments of the returns in economic and financial time series. Since
its introduction by Engle (1982) and Bollerslev (1986), several extensions have been proposed
in the literature. These extensions differ to each other in how the volatility equation is
defined.
Let rt be the observed return at time t and t the error term which follows a white noise
4
process. The GARCH(1,1) model is defined by
rt = σt t ,
σt2 = ω + αrt−1
2 2
+ βσt−1 ,
with σt2 being the conditional variance (or squared volatility) at time t and ω, α and β
parameters satisfying some stationary conditions. Sufficient conditions for stationarity are
given by ω > 0, α, β ≥ 0 and α + β < 1.
As mentioned above, different extensions of the GARCH model involve different spec-
ifications of the volatility equation (σt2 ). Table 1 describes the volatility equations of the
GARCH-type models used in this work. For good reviews of univariate GARCH-type mod-
els; see, for instance, Engle (1995), Teräsvirta (2009), Bollerslev (2010) and Rodrı́guez &
Ruiz (2012).
In all non-robust cases, several error distributions are assumed, namely, Normal, Skew
Normal, Student-t, Skew Student-t, GED, Skew GED, Normal Inverse Gaussian, General-
ized Hyperbolic and the Johnson’s reparametrized SU innovation distribution; see Ghalanos
(2018) for details about the parameterisation of the distributions and GARCH-type models
used in this paper.
Table 1: GARCH-type models

Model Volatility equation Proposed by
2
GARCH σt+1 = ω + αrt2 + βσt2 Bollerslev (1986)
IGARCH σt+1 = ω + αrt2 + (1 − α)σt2
2
Engle & Bollerslev (1986)
2
EGARCH log(σt+1 ) = ω + α2t + γ(|t | − E(|t |)) + βlog(σt2 ) Nelson (1991)
GJR σt+1 = ω + αrt2 + γI(rt < 0)rt2 + βσt2
2
Glosten et al. (1993)
δ
APARCH σt+1 = ω + α(|rt | − γrt )δ + βσtδ Ding et al. (1993)
CGARCH σt+1 = qt+1 + α(rt2 − qt ) + β(σt2 − qt )
2
qt+1 = ω + ρqt + φ(rt2 − σt2 ) Lee & Engle (1999)

TGARCH σt+1 = ω + ασt (|t | − η1 t ) + βσt Zakoian (1994)
AVGARCH σt+1 = ω + ασt (|t − η2 | − η1 (t − η2 )) + βσt Schwert (1990)
δ
NGARCH σt+1 = ω + ασtδ (|t |)δ + βσtδ Higgins & Bera (1992)
NAGARCH σt+1 = ω + ασt2 (|t − η2 |)2 + βσt2
2
Engle & Ng (1993)
δ
FGARCH σt+1 = ω + ασtδ (|
t −2 η2 | − η1 (t − η2 ))δ + βσtδ Hentschel (1995)
2 rt
Robust σt+1 = ω + γc αρ 2 + βσt2 , Boudt et al. (2013)
( σ t
1, if x > c,
GARCH with ρ(x) = with the modification of
x, if x ≤ c,
and γc being a constant to ensure consistency Trucı́os et al. (2017)
5
2.1.2 GARCH-MIDAS model
Engle et al. (2013) extended the work of Ghysels et al. (2006) and proposed a procedure
that decomposes the conditional variance into two components distinguishing short-run
from long-run movements. In that model, called GARCH-MIDAS, the short-run component
evolves as a GARCH(1,1) process while the long-run component is determined by past values
of the realised volatility.
Let ri,t be the observed return for day i of an arbitrary period t, i,t a white noise error
term and τt and gi,t be the unobserved long-run and short-run componentss respectively.
The GARCH(1,1)-MIDAS is defined as
p
ri,t = τt × gi,t t ,
2
ri−1,t
gi,t = (1 − α − β) + α + βgi−1,t ,
τt
K
X
τt = m + θ ϕk (ω)RVt−k ,
k=1
Nt
X
2
RVt = ri,t ,
i=1
where ϕk (ω) is a weight function, K stands for the number of lags used in the long-run
component and Nt is the number of observations within the arbitrary period t. In this
paper, we use K = 10 and the arbitrary period t equal to a month.
As an alternative to RVt , the model proposed by Engle et al. (2013) also allows the use
of macroeconomic variables; see, for instance, Walther & Klein (2018) and Conrad et al.
(2018). In this paper, we only use information available in the Bitcoin returns.
2.1.3 Realised GARCH model
The realised GARCH model proposed by Hansen et al. (2012) combines a model for realised
measures with a GARCH structure for the volatility where the squared return is replaced by
a realised measure. The approach is based on the argument that returns only offer a weak
information about the volatility and, consequently, they may lead to a poor performance
when volatility changes quickly. On the other hand, Andersen & Bollerslev (1998), Hansen
et al. (2012), Christoffersen et al. (2014) argue that realised measures are more informative
6
about the volatility than the squared returns.
Let rt be the observed daily return, σt2 its conditional variance, t a white noise error
term and xt a realised measure of variance. The realised GARCH(1,1) model is defined as
rt = σt t ,
2
log(σt2 ) = ω + α log(xt−1 ) + β log(σt−1 ),
log(xt ) = δ + γ log(σt2 ) + τ (t ) + ut
where xt is a realised measure at time t, τ (t ) is a function such that E(τ (t )) = 0 and ut
is a white noise process. In this paper, we use τ (t ) = η1 t + η2 (2t − 1). Note that with
this model we are able to forecast the volatility as well as the realised measure. Therefore,
we denote by RealGARCHx when the realised measure “x” is used as an auxiliary variable
to forecast the volatility and by RealGARCH when the volatility is used to forecast the
realised measure. Another model that uses a similar strategy is the high frequency based
volatility (HEAVY) model of Shephard & Sheppard (2010).
2.1.4 GAS models
Creal et al. (2013) and Harvey (2013) proposed a new class of models called Generalised
Autoregressive Score (GAS), also known as Dynamic Conditional Score (DCS) or Score-
Driven (SD) models. The general expression of the GAS model is given by

∂ log p(rt |ft )
ft+1 = ω + βft + αSt , (1)
∂ft
−λ
with St being a scaling function for the score.
" St is usuallyused # as It where λ = 0, 1/2
2
∂ log p(rt |ft )
or 1 and It is the Fisher information, Et−1 . The main difference with
∂ft
the classical GARCH model is in the evolution of the volatility equation (ft = σt2 ) which
depends on the past values of the score of the conditional distribution instead of the squared
returns. Since the score depends on the complete density instead of second-order moments
only, the GAS models allow us to better exploit the dynamic structure of the data.
Although well-known models such as GARCH and EGARCH can be obtained as special
cases of the GAS model by using p(·) as the Gaussian density and appropriate St and ft .
7
The main advantage of the GAS model comes from using p(·) as a non-Gaussian density.
An interesting and well-studied choice for p(·) is the Student-t density. This choice is
particularly attractive by its flexibility to capture the voltality dynamics and also by its
robustness to heavy tails and outliers; see, for instance, Harvey & Chakravarty (2008),
Harvey & Sucarrat (2014), Blazsek & Villatoro (2015) and Blasques et al. (2017).
In this paper, we use the betatEGARCH and betaSkewtEGARCH models proposed by
Harvey & Chakravarty (2008) and Harvey & Sucarrat (2014) as well as the t-GAS model;
see, Gao & Zhou (2016) for a brief explanation about the differences among these models.
2.2 Realised measures
Evaluating the predictive ability of different approaches to forecast the volatility is a chal-
lenging task since the volatility is a latent variable and consequently is not directly ob-
servable. A common practice in the literature is to use a proxy for the true conditional
variance. In this paper, we use the realised variance - RV (Andersen et al., 2003) and some
other realised measures robust to jumps and microstructure noise as a proxy of true squared
volatility. Specifically, we use the Bipower variation - BV (Barndorff-Nielsen & Shephard,
2004), MinRV (Andersen et al., 2012) and MedRV (Andersen et al., 2012). We prefer to
use realised measures instead of square observed daily returns as in Catania et al. (2018)
because realised measures have shown to be a better proxy of the true volatility (Alizadeh
et al., 2002; McAleer & Medeiros, 2008; Patton, 2011) and are most widely used nowadays.
Let rt (i) be the ith high-frequency return of day t, the realised measures used in this
paper are given in Table 2. The main difference between RV and the other realised measures
is that the latter are jump-robust and aim to estimate the integrated variance (IV) while
the former estimates the quadratic variation (QV). The relation between QV and IV is
that QVt = IVt + JVt , where JVt stands for the jump component. For more details about
realised measures as well as for asymptotic properties see, for instance, Andersen et al.
(2007), McAleer & Medeiros (2008), Andersen et al. (2012) and Liu et al. (2015).
8
Table 2: Realised measures.
Realised measure Formula
XN
RVt rt2 (i)
i=1
NX−1
π N
BVt |rt (i)||rt (i + 1)|
2 N −1
i=1
NX −1
π N
MinRVt min (|rt (i)|, |rt (i + 1)|)2
π−2 N −1
i=1
NX −1
π N
MedRVt √ med (|rt (i − 1)|, |rt (i)|, |rt (i + 1)|)2
6−4 3+π N − 2
i=1
2.3 Robust loss function
The evaluation of the volatility forecast is usually conducted by using volatility proxies
such as those described in the previous section. However, these proxies are imperfect since
they are estimates of the integrated variance. To avoid that these volatility proxies lead to
misleading results when comparing the volatility forecasts, Patton (2011) propose the use
of loss function robust to the microstructure. The general class of the robust loss function
is defined by Patton (2011) and it is given by

2 2 σ̂ 2
−


 h σ̂ + σ̂ log( ), for b = -1,



 2 h
σ̂ σ̂ 2
L(σ̂ 2 , h, b) = − log( ) − 1, for b = -2,

 h h
(σ̂ 2b+4 − h2b+4 ) hb+1 (σ̂ 2 − h)


−


 , otherwise,
(b + 1)(b + 2) b+1
where σ̂ 2 is the squared forecasts volatility and h is the squared volatility proxy. In this
paper we consider the robust loss function of Patton (2011) with three different values of
b: b = −2 (QLIKE), b = −1 (MSE) and b = 0 (hereafter denoted by RLF). Observe that
when h = σ̂ 2 all loss functions are equal to zero.
2.4 VaR estimation
Assuming that returns are zero mean, the one-step-ahead VaR is usually estimated as
VaRα = Qα σ̂T+1|T where Qα is the α quantile of the assumed error distribution (scaled
to have unit variance) and σ̂T +1|T is the one-step-ahead volatility forecast.
To estimate the VaR in a robust way, we use the robust bootstrap procedure recently
9
proposed by Trucı́os et al. (2017). The procedure is based on a residual-based bootstrap
scheme combining with the robust GARCH estimator defined previously and robust filters
for the volatility. The procedure can be summarized in the following steps:
• Step 1: Estimate the parameters ω, α and β in a robust way and obtain the stan-
rt
dardized residuals ˆt = σ̂t
. Denote by F̂ the empirical distribution of these centred
standardised residuals.
• Step 2: Using ∗t (bootstrap extractions from F̂ ), generate bootstrap series through
the following recursion.
rt∗ = σt∗ ∗t ,

rt∗2
(2)
∗2
σt+1 = ω̂ + α̂σt∗2 cγ rc + β̂σt∗2 ,
σt∗2
where σ1∗2 = σ̂12 and the filter rc (·) is similar to the filter defined in Table 1 but
large values are replaced by a new squared bootstrap extractions from F̂ . With the
bootstrap series obtained in (2), estimate the parameters ω̂ ∗ , α̂∗ and β̂ ∗ using the same
estimator in Step 1.
• Step 3: Obtain h-steps-ahead forecasts as
r̂T∗ +h|T = ∗T +h σ̂T∗ +h|T ,

(3)
!
rT∗2+h−1|T
σ̂T∗2+h|T = ω̂ ∗ + α̂∗ σ̂T∗2+h−1|T cγ rc + β̂ ∗ σ̂T∗2+h−1|T ,
σ̂T∗2+h−1|T
for h = 1, ..., H, and where r̂T∗ |T = rT , ∗T +h are bootstrap extractions from F̂ and

2
rt−1
∗2 ∗2 ∗ ∗ ∗2
σ̂T |T is obtained through the recursion σ̂t|T = ω̂ + α̂ σ̂t−1|T cγ rc σ̂∗2 + β̂ ∗ σ̂t−1|T
∗2
for
t−1|T
∗2
t = 2, ..., T , with σ̂1|T = σ̂12 .
∗(1) ∗(B)
• Step 4: Repeat steps 2 and 3 B times to obtain B bootstrap replicates (r̂T +h|T , ..., r̂T +h|T ),
the VaRα is estimated as the h-step-ahead α empirical quantile of the bootstrap repli-
cates.
10
3 Data and results
We use daily Bitcoin closing prices (in US dollar) traded on Bitstamp from September 13,
2011, to December 31, 2017 (2280 observations). Since Bitcoin is traded 24/7, the daily
closing prices are considered as the last price traded at each day and were constructed using
the tick-by-tick data obtained from bitcoincharts 1 .
Returns are calculated as rt = log(Pt /Pt−1 ) with Pt being the closing price at day
t. Table 3 reports descriptive statistics and Figure 1 shows the daily returns as well as
the autocorrelations of returns and squared returns. The confidence bands for the returns
were constructed using the generalised non-parametric Bartlett’s formula for non-linear
process (Francq & Zakoı̈an, 2009) while the confidence bands for the squared returns were
constructed using the Bartlett’s formula (Bartlett, 1946).
We can observe that, Bitcoin is highly volatile with an annualised standard deviation of
√
0.8350 (0.0526 × 252) and returns also present asymmetry and large Kurtosis. The large
Kurtosis is probably explained by the presence of extreme returns, as observed in Figure 1.
Because the returns series does not exhibit serial correlation, no ARMA filter is applied to
the data, so that the series is only centred to have zero mean.
The largest return (in absolute value) corresponds to April 10, 2013, when Bitcoin prices
drop down around 60% of its value. This event was attributed to a DDoS (distributed
denial of service) attack2 . The largest value in 2012 corresponds to August 19. According
to Forbes3 , after the Bitcoin prices doubled between July 1 and August 18, 2012, its value
dropped down considerably. More than 25% of the loss between June 2014 and November
2016 happened on January 13, 2015, a week after the temporary suspension of the services
of Bitstamp and the recent news that Russia started banning websites related to Bitcoin4 .
Table 3: Descriptive statistics of Bitcoin daily returns

Mean Std. Dev. Min Q1 Med. Q3 Max Skewness Kurtosis
0.0034 0.0526 -0.6639 -0.0111 0.0024 0.0200 0.4455 -1.4011 28.4802
To evaluate the out-of-sample performance of the models we use a rolling windows scheme
1
https://api.bitcoincharts.com/v1/csv/bitstampUSD.csv
2
https://bitcoinmagazine.com/articles/the-bitcoin-crash-an-examination-1365911041/
3
https://www.forbes.com/sites/timothylee/2013/04/11/an-illustrated-history-of-bitcoin-
crashes/#640b1c940397
4
https://www.cnbc.com/2015/01/14/bitcoin-falls-below-200-making-some-investors-worry-about-
downward-spiral.html
11
0.50
0.25
0.00
0.1
returns
−0.25
−0.50
ACF(returns)
2011−09 2013−02 2014−06 2015−11 2017−03

0.0
0.1
0.2
ACF(squared returns)
ACF(returns)
0.0
0.1
−0.1
−0.1 0.0
0 0 10 20 10 30 40 20 50 0 30 10 20 40 30 40 50 50
lag lag lag
Figure 1: Daily returns (top panel), sample autocorrelation function of returns with their
corresponding generalised non-parametric Bartlett 95% confidence bands (left bottom panel)
and sample autocorrelation of squared returns with their corresponding Bartlett 95% con-
fidence bands (right bottom panel).
with a windows size equal to 1000 days. In each window, the one-step-ahead conditional
variance is estimated. As the daily volatility is not observed, we use as a proxy the realised
measures described in Section 2.2 (using five-minutes high-frequency data). Table 4 reports
the descriptive statistics of those realised measures in the out-of-sample period and Figure
2 shows its evolution over time5 . As expected, the RV is larger than the other realised
measures. The MinRV, MedRV and BV are all robust to jumps and microstructure noise
which are desirable properties in volatility proxies as pointed out by Barndorff-Nielsen &
Shephard (2004) and Andersen et al. (2012) which make these measures preferred to RV.
The extreme value observed when the RV is computed can explain the large values of the
Skewness and Kurtosis.
Unlike Charles & Darné (2018) which takes into account the presence of outliers using
5
To better visualise the results, the values were cut-off to 0.0023. However, there is a large RV (0.0151)
on April 17, 2016
12
Table 4: Descriptive statistics (multiplied by 100) of the realised measures in the out-of-
sample period.
Mean Std. Dev. Min Q1 Med. Q3 Max Skewness Kurtosis
RV 0.0333 0.0479 0.0011 0.0150 0.0255 0.0434 1.5133 23.3025 714.2308
MinRV 0.0171 0.0151 0.0002 0.0070 0.0128 0.0226 0.1485 2.3828 12.6226
MedRV 0.0203 0.0163 0.0001 0.0088 0.0156 0.0271 0.1155 1.7340 7.1631
BV 0.0154 0.0124 0.0006 0.0068 0.0122 0.0205 0.1184 2.1705 11.4692
BV MedRV
0.0020
0.0015
0.0010
0.0005
Realised measure
0.0000
MinRV RV
0.0020
0.0015
0.0010
0.0005
0.0000
2014−06 2015−11 2017−03 2014−06 2015−11 2017−03
Figure 2: Realised measures in the out-of-sample period. BV (left top panel), MedRV (right
top panel), MinRV (left bottom panel) and RV (right bottom panel).
jump-filtered returns as in Laurent et al. (2016), we estimate the conditional variance in

a robust way using the robust estimator of Boudt et al. (2013) with the modification in-
troduced by Trucı́os et al. (2017). Alternative robust procedures such as the proposed by
Muler & Yohai (2008) and Carnero et al. (2012) could also be used, although, the results in
Trucı́os et al. (2015) shown that the procedure of Boudt et al. (2013) with the modification
introduced by Trucı́os et al. (2017) has the best performance in a forecasting context.
Table 5 reports the MSE, QLIKE and RLF between σ̂T2 +1 obtained by the GARCH-type
models and the realised measures RV, MinRV, MedRV, BV used as a proxy of the true
value σT2 +1 . The shadowed cells are the set of models with best out-of-sample performance
obtained using the MCS approach (Hansen et al., 2011) at 75% significant level.6 In bold,
the best model (first position in the MCS rank) in each case.
The MSE selects a large set of models reflecting a low power to distinguish between
different models. Indeed, when looking at the MSE, just 15 of the 100 models (16 when
6
We used the R package MCS of Bernardi (2017).
13
considering the MinRV as a volatility proxy) are left out. Other works have also found that
the MSE has a low power to distinguish between different models; see, for instance, Patton
& Sheppard (2009) and Liu et al. (2015). The QLIKE and RLF loss functions generally
selects a small number of models as best models7 .
In general, we can observe that the MCS methodology selects the same models by loss
function regardless of the realised measure used. An exception is observed when the QLIKE
measure is used, in which case the CGARCH model belongs to the group of models with
the best performance only when the RV is used.
Considering all loss functions and realised measures, only five models are in the set
of best models in all cases, namely, TARCH and AVGARCH, both assuming GED (with
standard and skew versions) and the robust GARCH model. The CGARCH model which
was the best model according to Katsiampa (2017), now appear in the set only when using
the MSE and RLF loss functions and or when using the QLIKE loss function with the RV as
volatility proxy. The IGARCH model which was the best model in the in-sample analyses
of Chu et al. (2017) only appears in the set when considering the MSE loss functions.
In all cases, the GARCH model estimated in a robust way reports the best results
regardless of the loss function and realised measure used. This result is extremely important
since most of the studies available in the literature about Bitcoin do not take into account
the presence of outliers. As we can see in Table 5, the forecasts can be substantially improved
using a robust approach, so that a simple model as the robust GARCH model considered in
this paper outperforms more sophisticated models. These results are in concordance with
Charles & Darné (2018) which also find, in an in-sample context, that, sophisticated models
are outperformed by a GARCH model when the presence of outliers is considered in the
analysis.
The results in Table 5 compare most of the (if not all) classical GARCH-type models used
in the literature to model and forecast the Bitcoin volatility. Furthermore, we go one step
further and forecast the volatility using the alternative approaches described in Subsections
2.1.2 to 2.1.4 whose results are reported in Table 6. For the sake of comparison, we also
7
It is important to note that even when the average MSE considering the TARCH model with SGED
innovations distribution is extremely large. The MCS still selects this model in the set of best models, this
can be explained by the fact that this high average value is given by just an extreme case and the remaining
cases reported small values of the MSE
14
Table 5: 2
Average MSE, QLIKE and RLF of σ̂T +1 (obtained using GARCH-type models) and the realised measures. The
shadowed cells stand for the set of models selected by the MCS.
Innov MSE (× 105 ) QLIKE RLF (× 103 )
Distrib. RV MinRV MedRV BV RV MinRV MedRV BV RV MinRV MedRV BV
norm 0.6462 0.6924 0.6799 0.6932 4.6611 15.7753 12.8778 14.0071 0.2137 0.3374 0.3023 0.3356
snorm 0.6420 0.6879 0.6755 0.6886 4.5618 15.5269 12.6780 13.7686 0.2110 0.3334 0.2987 0.3316
std 0.6301 0.6750 0.6626 0.6756 4.3189 14.8890 12.1316 13.1683 0.2089 0.3293 0.2950 0.3271
GARCH
sstd 0.6301 0.6751 0.6627 0.6756 4.3255 14.9089 12.1485 13.1858 0.2091 0.3297 0.2952 0.3274
ged 0.6223 0.6663 0.6541 0.6668 4.2173 14.6294 11.9493 12.9009 0.2039 0.3222 0.2885 0.3199
sged 0.6233 0.6675 0.6553 0.6680 4.2448 14.7110 12.0190 12.9733 0.2046 0.3233 0.2895 0.3210
nig 0.6298 0.6745 0.6622 0.6750 4.2573 14.7438 12.0325 13.0099 0.2067 0.3262 0.2921 0.3239
ghyp 0.6301 0.6748 0.6624 0.6753 4.2558 14.7365 12.0229 13.0057 0.2068 0.3263 0.2922 0.3240
jsu 0.6282 0.6729 0.6605 0.6734 4.2763 14.7824 12.0522 13.0590 0.2074 0.3272 0.2930 0.3250
norm 0.6848 0.7331 0.7202 0.7339 4.7743 16.1611 13.2284 14.3087 0.2237 0.3511 0.3150 0.3491
snorm 0.6777 0.7255 0.7126 0.7262 4.6886 15.9360 13.0387 14.1000 0.2210 0.3472 0.3114 0.3451
std 0.6347 0.6799 0.6675 0.6805 4.3386 14.9463 12.1776 13.2208 0.2101 0.3311 0.2965 0.3289
IGARCH
sstd 0.6347 0.6799 0.6675 0.6805 4.3451 14.9660 12.1944 13.2380 0.2103 0.3314 0.2968 0.3292
ged 0.6324 0.6769 0.6646 0.6774 4.2577 14.7373 12.0339 12.9998 0.2069 0.3262 0.2922 0.3238
sged 0.6324 0.6770 0.6647 0.6775 4.2854 14.8159 12.0990 13.0714 0.2075 0.3272 0.2931 0.3249
nig 0.6348 0.6797 0.6673 0.6802 4.2773 14.8017 12.0791 13.0627 0.2080 0.3280 0.2938 0.3257
ghyp 0.6349 0.6798 0.6674 0.6803 4.2736 14.7908 12.0680 13.0539 0.2080 0.3280 0.2938 0.3257
jsu 0.6329 0.6778 0.6654 0.6783 4.2947 14.8362 12.0955 13.1079 0.2087 0.3289 0.2946 0.3266
norm 0.3997 0.4384 0.4268 0.4397 4.2819 14.6638 11.9263 13.0381 0.1808 0.2942 0.2614 0.2934
snorm 0.3816 0.4187 0.4073 0.4199 4.1724 14.2734 11.5432 12.7374 0.1756 0.2865 0.2542 0.2857
EGARCH
std 1.3587 1.4412 1.4209 1.4434 8.3776 26.2082 21.2621 23.6404 0.4427 0.6451 0.5872 0.6438
sstd 4.1221 4.2203 4.1960 4.2242 12.4856 33.6221 27.5656 31.3650 0.7206 0.9637 0.8926 0.9662
ged 0.4687 0.5082 0.4964 0.5091 4.2841 14.7140 11.8625 13.1082 0.1953 0.3115 0.2778 0.3100
sged 0.4772 0.5175 0.5055 0.5184 4.3620 14.9607 12.0615 13.3298 0.1989 0.3168 0.2826 0.3153
nig 0.5874 0.6336 0.6205 0.6347 4.8527 16.4305 13.2608 14.6555 0.2302 0.3598 0.3223 0.3582
ghyp 0.6182 0.6666 0.6531 0.6676 4.9544 16.7707 13.5804 14.9264 0.2372 0.3700 0.3317 0.3682
jsu 0.7094 0.7620 0.7476 0.7632 5.4314 18.0941 14.5942 16.1862 0.2653 0.4078 0.3667 0.4061
norm 0.6660 0.7127 0.7001 0.7135 4.7435 15.9572 12.9768 14.2207 0.2187 0.3438 0.3083 0.3420
snorm 0.6389 0.6840 0.6717 0.6848 4.6482 15.6192 12.6296 13.9678 0.2145 0.3375 0.3024 0.3358
GJR-GARCH
std 0.6722 0.7172 0.7048 0.7177 4.5216 15.3809 12.4347 13.7138 0.2205 0.3438 0.3085 0.3417
sstd 0.6751 0.7203 0.7078 0.7209 4.5321 15.4303 12.4847 13.7505 0.2211 0.3447 0.3094 0.3426
ged 0.6565 0.7007 0.6884 0.7012 4.3948 15.0588 12.2121 13.3783 0.2138 0.3346 0.3001 0.3324
sged 0.6636 0.7083 0.6960 0.7088 4.4448 15.2279 12.3615 13.5232 0.2161 0.3380 0.3033 0.3358
nig 0.6720 0.7171 0.7047 0.7176 4.4515 15.2494 12.3697 13.5464 0.2179 0.3403 0.3054 0.3381
ghyp 0.6676 0.7125 0.7001 0.7130 4.4402 15.2048 12.3250 13.5114 0.2173 0.3395 0.3046 0.3373
jsu 0.6733 0.7183 0.7059 0.7188 4.4802 15.3081 12.3999 13.6200 0.2193 0.3422 0.3071 0.3400
norm 0.5586 0.6027 0.5902 0.6038 4.7028 15.8339 12.8170 14.1272 0.2115 0.3350 0.2995 0.3337
snorm 0.5675 0.6110 0.5986 0.6120 4.6737 15.6700 12.6122 14.0424 0.2126 0.3354 0.3000 0.3341
std 1.7452 1.8356 1.8142 1.8373 8.3640 26.8273 21.8816 23.9507 0.4686 0.6800 0.6200 0.6771
APARCH
sstd 1.8165 1.9096 1.8877 1.9113 8.5378 27.4649 22.4993 24.4475 0.4809 0.6969 0.6357 0.6940
ged 0.6816 0.7289 0.7158 0.7296 4.6978 16.0421 12.9647 14.2607 0.2317 0.3605 0.3235 0.3583
sged 0.6922 0.7402 0.7270 0.7410 4.7671 16.2775 13.1611 14.4651 0.2352 0.3656 0.3282 0.3635
nig 0.8510 0.9064 0.8918 0.9072 5.2732 17.8979 14.5316 15.8679 0.2698 0.4134 0.3723 0.4108
ghyp 0.9272 0.9856 0.9705 0.9863 5.3893 18.3389 14.9618 16.2017 0.2803 0.4280 0.3859 0.4252
jsu 1.0107 1.0730 1.0571 1.0739 5.8575 19.6262 15.9345 17.4342 0.3075 0.4645 0.4197 0.4619
norm 0.5051 0.5437 0.5324 0.5442 3.7672 13.2390 10.7400 11.7096 0.1779 0.2865 0.2553 0.2845
snorm 0.5101 0.5494 0.5381 0.5499 3.7854 13.2734 10.7744 11.7376 0.1784 0.2872 0.2560 0.2852
CGARCH
std 0.5592 0.5996 0.5881 0.5999 3.7305 13.2083 10.7250 11.6396 0.1846 0.2949 0.2633 0.2924
sstd 0.5644 0.6056 0.5939 0.6058 3.8005 13.3810 10.8571 11.8077 0.1870 0.2986 0.2666 0.2960
ged 0.5320 0.5713 0.5600 0.5715 3.7362 13.1697 10.7139 11.5997 0.1803 0.2888 0.2577 0.2863
sged 0.5280 0.5672 0.5559 0.5674 3.7172 13.1183 10.6648 11.5505 0.1799 0.2883 0.2572 0.2858
nig 0.5441 0.5842 0.5728 0.5844 3.7599 13.2428 10.7656 11.6695 0.1832 0.2929 0.2614 0.2903
ghyp 0.5472 0.5876 0.5760 0.5878 3.7487 13.2312 10.7534 11.6463 0.1837 0.2937 0.2621 0.2911
jsu 0.5493 0.5903 0.5786 0.5905 3.7040 13.1620 10.6864 11.5745 0.1837 0.2943 0.2626 0.2917
norm 0.4556 0.4977 0.4855 0.4989 4.4336 15.1155 12.2453 13.4620 0.1931 0.3116 0.2773 0.3106
snorm 0.4272 0.4673 0.4554 0.4685 4.3205 14.7282 11.8736 13.1577 0.1866 0.3021 0.2685 0.3012
std 1.5828 1.6712 1.6500 1.6731 8.3242 26.6331 21.6692 23.8363 0.4582 0.6671 0.6075 0.6648
TARCH
sstd 1.6504 1.7416 1.7199 1.7436 8.5075 27.2939 22.2991 24.3616 0.4708 0.6845 0.6236 0.6821
ged 0.3065 0.3125 0.3068 0.3128 3.0786 8.9822 6.6485 8.5162 0.1422 0.2063 0.1860 0.2055
sged 6.2929 6.3005 6.2951 6.3018 4.2964 10.5292 8.2206 10.2862 0.2816 0.3505 0.3309 0.3530
nig 0.7427 0.7960 0.7817 0.7969 5.1558 17.5626 14.2467 15.5755 0.2572 0.3974 0.3571 0.3952
ghyp 0.8061 0.8622 0.8474 0.8631 5.2881 18.0411 14.7000 15.9507 0.2678 0.4122 0.3709 0.4098
jsu 0.8974 0.9577 0.9420 0.9587 5.7644 19.3435 15.6862 17.1989 0.2962 0.4503 0.4061 0.4480
norm 0.5495 0.5940 0.5813 0.5951 4.5948 15.2893 12.2950 13.7239 0.2074 0.3293 0.2941 0.3281
snorm 0.5027 0.5443 0.5321 0.5454 4.4396 15.0180 12.1520 13.3952 0.1988 0.3169 0.2827 0.3158
AVGARCH
std 1.5283 1.6143 1.5937 1.6160 8.3379 26.7001 21.7781 23.8574 0.4505 0.6564 0.5978 0.6538
sstd 1.5985 1.6881 1.6667 1.6900 8.6625 27.6488 22.6233 24.6775 0.4685 0.6805 0.6201 0.6781
ged 0.2903 0.2957 0.2902 0.2960 3.0270 8.8674 6.5561 8.4003 0.1382 0.2010 0.1811 0.2002
sged 0.3010 0.3067 0.3012 0.3071 3.1075 9.0838 6.7185 8.6091 0.1423 0.2063 0.1860 0.2055
nig 0.7162 0.7690 0.7548 0.7699 5.1853 17.7134 14.4072 15.6786 0.2553 0.3951 0.3549 0.3929
ghyp 0.8057 0.8616 0.8469 0.8624 5.3933 18.3932 15.0112 16.2479 0.2710 0.4163 0.3747 0.4138
jsu 0.8512 0.9107 0.8953 0.9118 5.7763 19.1370 15.3250 17.1624 0.2912 0.4439 0.4001 0.4416
norm 0.5416 0.5858 0.5733 0.5869 4.6604 15.7700 12.7876 14.0398 0.2081 0.3310 0.2956 0.3296
snorm 0.5319 0.5753 0.5629 0.5763 4.5226 15.3895 12.4799 13.6896 0.2030 0.3237 0.2890 0.3223
NGARCH
std 1.6741 1.7651 1.7436 1.7668 7.9521 25.7754 21.2326 22.8155 0.4470 0.6538 0.5953 0.6508
sstd 1.7375 1.8309 1.8089 1.8327 8.1307 26.3710 21.7885 23.2955 0.4593 0.6702 0.6105 0.6671
ged 0.6240 0.6704 0.6575 0.6712 4.4331 15.3760 12.5222 13.5547 0.2159 0.3404 0.3046 0.3381
sged 0.6309 0.6780 0.6650 0.6788 4.4978 15.5595 12.6575 13.7321 0.2191 0.3450 0.3089 0.3428
nig 0.7848 0.8391 0.8248 0.8399 4.9346 16.9856 13.9009 14.9371 0.2505 0.3887 0.3493 0.3861
ghyp 0.8490 0.9057 0.8910 0.9065 4.9718 17.2124 14.1918 15.0570 0.2560 0.3966 0.3567 0.3938
jsu 0.9421 1.0036 0.9878 1.0045 5.4913 18.6339 15.2523 16.4191 0.2873 0.4389 0.3958 0.4362
norm 0.6528 0.6993 0.6868 0.7001 4.6372 15.7110 12.8230 13.9435 0.2139 0.3375 0.3025 0.3357
snorm 0.6234 0.6681 0.6559 0.6688 4.5686 15.4497 12.5384 13.7637 0.2100 0.3314 0.2968 0.3295
NAGARCH
std 0.6804 0.7263 0.7137 0.7269 4.7956 16.2179 13.0605 14.4966 0.2309 0.3581 0.3217 0.3559
sstd 0.6805 0.7266 0.7139 0.7271 4.8023 16.2470 13.0900 14.5172 0.2311 0.3585 0.3220 0.3563
ged 0.6477 0.6920 0.6797 0.6924 4.4874 15.3887 12.4775 13.6599 0.2158 0.3378 0.3030 0.3354
sged 0.6462 0.6908 0.6785 0.6913 4.5119 15.4574 12.5342 13.7219 0.2162 0.3387 0.3037 0.3364
nig 0.6691 0.7147 0.7021 0.7151 4.6438 15.8441 12.8098 14.0973 0.2243 0.3495 0.3136 0.3471
ghyp 0.6664 0.7120 0.6994 0.7124 4.6296 15.7878 12.7553 14.0513 0.2238 0.3488 0.3130 0.3464
jsu 0.6746 0.7203 0.7077 0.7207 4.7149 16.0246 12.9284 14.2904 0.2276 0.3538 0.3176 0.3515
norm 0.6313 0.6769 0.6642 0.6779 4.6570 15.7984 12.8002 14.0545 0.2202 0.3456 0.3095 0.3440
snorm 0.6314 0.6753 0.6628 0.6762 4.5357 15.3349 12.3748 13.7062 0.2181 0.3414 0.3058 0.3398
Full GARCH
std 1.5654 1.6517 1.6310 1.6535 8.4590 26.9654 21.9542 24.1603 0.4575 0.6647 0.6057 0.6622
sstd 1.6544 1.7434 1.7221 1.7452 8.5986 27.4219 22.3873 24.4923 0.4713 0.6826 0.6223 0.6799
ged 0.6032 0.6485 0.6357 0.6492 4.5789 15.7680 12.7706 13.9733 0.2213 0.3468 0.3105 0.3447
sged 0.6112 0.6573 0.6443 0.6581 4.6636 16.0417 12.9991 14.2162 0.2249 0.3521 0.3153 0.3500
nig 0.7458 0.7986 0.7844 0.7994 5.2114 17.7875 14.4494 15.7488 0.2592 0.3994 0.3591 0.3970
ghyp 0.8244 0.8805 0.8657 0.8813 5.4332 18.5296 15.0815 16.3849 0.2744 0.4205 0.3786 0.4179
jsu 0.8738 0.9329 0.9174 0.9338 5.7881 19.4739 15.8039 17.2912 0.2934 0.4461 0.4023 0.4436
Robust 0.1027 0.1099 0.1041 0.1102 1.8795 7.2874 5.6621 6.4628 0.0644 0.1192 0.1020 0.1182
15
Table 6: 2
Average MSE, QLIKE and RLF of σ̂T +1 (obtained using alternative models) and the realised measures. The
shadowed cells stand for the set of models selected by the MCS.
Models MSE (× 105 ) QLIKE RLF (× 103 )
RV MinRV MedRV BV RV MinRV MedRV BV RV MinRV MedRV BV
Robust GARCH 0.1027 0.1099 0.1041 0.1102 1.8795 7.2874 5.6621 6.4628 0.0644 0.1192 0.1020 0.1182
GARCH-MIDAS 0.5492 0.5887 0.5773 0.5894 4.1104 14.0998 11.4015 12.5635 0.1902 0.3026 0.2704 0.3009
RealGARCHRV 0.1477 0.1725 0.1631 0.1768 3.1193 10.0758 7.7690 9.6237 0.1080 0.1896 0.1643 0.1958
RealGARCHMinRV 0.2456 0.2806 0.2691 0.2868 5.0407 14.6723 11.6306 14.0496 0.1758 0.2802 0.2483 0.2895
RealGARCHMedRV 0.2314 0.2655 0.2541 0.2714 4.7684 13.9661 10.9824 13.4170 0.1658 0.2676 0.2365 0.2765
RealGARCHBV 0.2189 0.2514 0.2404 0.2571 4.6060 13.5970 10.7120 13.0180 0.1578 0.2564 0.2262 0.2648
betatEGARCH 0.0703 0.0695 0.0655 0.0690 1.0767 4.6618 3.7484 3.9375 0.0375 0.0734 0.0621 0.0716
betatEGARCH levarage 0.0700 0.0694 0.0654 0.0689 1.0742 4.6699 3.7641 3.9354 0.0374 0.0734 0.0620 0.0715
betaSkewtEGARCH 0.0722 0.0713 0.0674 0.0708 1.1505 4.8617 3.8758 4.1499 0.0399 0.0765 0.0649 0.0747
betaSkewtEGARCH leverage 0.0700 0.0687 0.0649 0.0683 1.1369 4.8128 3.8380 4.1045 0.0391 0.0752 0.0637 0.0734
t-GAS 0.0337 0.0051 0.0066 0.0038 3.6695 2.9776 3.1786 2.9232 0.0315 0.0155 0.0187 0.0139
include in Table 6 the results obtained with the robust GARCH procedure already presented
in Table 5. As expected, the GARCH-MIDAS and the Realised GARCH (regardless the
realised measures used as an auxiliary variable) outperform the Gaussian GARCH model
(Table 5). However, they are outperformed by the Robust GARCH model.
In order to include additional robust procedures to model and forecast the volatility, Ta-
ble 6 reports results using the betatEGARCH, betaSkewtEGARCH and the t-GAS models,
which are all particular cases of the general class of GAS models. For the sake of compari-
son we also include the leverage effect in the betatEGARCH and betaSkewtEGARCH; see
Gao & Zhou (2016) for a brief explanation. The results reveal that the GAS-type models
used in this paper are, in general, better to capture the dynamic of Bitcoin volatility than
the robust GARCH procedure. The best result is achieved mostly using the t-GAS model
and only in one case the best performance is achieved by the betatEGARCH model with
leverage. These results are not surprising since the literature about GAS models argue that
scores are more informative than the squared returns to capture the volatility dynamics and
additionally down weight extreme observations.
Given that high-frequency data is available, in Table 7 we also compare the predictability
of the realised measures. To this aim, we use the ARFIMA model as suggested in Ander-
sen et al. (2003) and Koopman et al. (2005), the Heterogeneous Auto-Regressive (HAR)
model of Corsi (2009), the HEAVY model of Shephard & Sheppard (2010) and the re-
alised GARCH model described in Subsection 2.1.3. ARFIMAL and HARL stand for the
ARFIMA and HAR models applied to the logarithm of the realised measure. The results
using the logarithm are always better than using directly the realised measures. Note also
that, except for the ARFIMA and HAR models8 , the procedures that forecast the realised
8
We exclude of the comparison the HAR model when using MinRV because in two days they were
forecasted as negative values
16
Table 7: Average MSE, QLIKE and RLF of the one-step-ahead forecast realised measures and its true value. The shadowed
cells stand for the set of models selected by the MCS.
Models MSE (× 105 ) QLIKE RLF (× 103 )
RV MinRV MedRV BP RV MinRV MedRV BP RV MinRV MedRV BP
ARFIMA 0.0513 0.0647 0.0274 0.0233 0.8511 2.6159 1.7177 1.4447 0.0252 0.0540 0.0284 0.0274
ARFIMAL 0.0266 0.0017 0.0018 0.0011 0.1680 0.3593 0.3287 0.2394 0.0065 0.0035 0.0033 0.0025
HAR 0.0357 - 0.0066 0.0050 0.7770 - 1.2820 1.0710 0.0177 - 0.0125 0.0111
HARL 0.0214 0.0018 0.0019 0.0012 0.1688 0.4226 0.4565 0.2689 0.0053 0.0036 0.0035 0.0027
HEAVY 0.0360 0.0019 0.0020 0.0013 0.2584 0.4928 0.4734 0.3225 0.0098 0.0042 0.0041 0.0032
RealGARCH 0.0205 0.0015 0.0016 0.0010 0.1206 0.2960 0.2897 0.1911 0.0043 0.0028 0.0028 0.0021
measures report a better performance than the t-GAS model and the best performance is
achieved using the realised GARCH model. These results are not surprising since these
models are properly designed to forecast directly the realised measure and also fully exploit
the information available in the intra-day data while the models in Tables 5 and 6 forecast
the daily volatility and use (mostly) daily information.
The 1% VaR is also computed and backtesting procedures for the VaR are carried out.
The models compared are the AVGARCH (best performance among the non-robust proce-
dure), Robust GARCH (best performance among the classical GARCH-type models), t-GAS
(best performance to forecast the daily volatility) and Realised GARCH (best performance
to forecast the realised measures). In all cases, the usual VaR estimator was used, except for
the Robust GARCH model where the robust bootstrap procedure of Trucı́os et al. (2017)
was applied. Results are shown in Table 8 and report the proportion of fails (returns smaller
than the 1% VaR), p-values of the unconditional coverage - UC (Kupiec, 1995), conditional
coverage - CC (Christoffersen, 1998), dynamic quantile -DQ (Engle & Manganelli, 2004)
tests as well as the average quantile loss function of González-Rivera et al. (2004).
Results reveal that the non-robust procedures underestimate the VaR and the backtest-
ing tests reject the null hypothesis that the proportion of fails is equal to 0.01. Only the
robust bootstrap procedure of Trucı́os et al. (2017) showed a good performance in all cases
(tests fail to reject the null hypothesis and the smallest average quantile loss function is
observed). The t-GAS reports an improvement with respect to the performance obtained
by the AVGARCH and Realised GARCH. However, the tests UC, CC and DQ reject the
null hypothesis. Figure 3 reports the returns of the out-of-sample period and the 1% VaR
estimated in a robust way.
17
Table 8: Proportion of returns smaller than the 1% VaR, p-values of the UC, CC, DQ test
and average quantile loss function (AQLF) (×103 ).
Models Prop. Fails UC CC DQ AQLF
AVGARCHGED 0.1023 0.0000 0.0000 0.0000 2.9776
AVGARCHSGED 0.1023 0.0000 0.0000 0.0000 2.9710
Robust bootstrap 0.0063 0.1475 0.3331 0.6602 1.4283
t-GAS 0.0266 0.0000 0.0000 0.0000 1.7676
RealGARCH 0.0867 0.0000 0.0000 0.0000 3.4591
0.2
0.0
−0.2
−0.4
2014−06 2015−11 2017−03
Figure 3: Return in the out-of-sample period (black solid line) and the estimated 1% VaR
(red dashed line) obtained using the robust bootstrap procedure of Trucı́os et al. (2017).
4 Conclusions and future works
In this paper, we have carried out a comprehensive out-of-sample comparison in the context
of the daily volatility forecast of Bitcoin using several volatility models, including robust
procedures and procedures based on intra-day data. Their VaR performance was also ad-
dressed.
The best model to forecast the volatility among the non-robust classical GARCH-type
models was the AVGARCH assuming errors coming from a GED (symmetric and skew
versions). However, this model was outperformed by the robust procedures.
We emphasize the fact that the presence of outliers should not be neglected and sub-
stantial improvements can be obtained by using robust procedures. The robust GARCH
procedure and the GAS model outperformed all the non-robust procedures to forecast the
volatility, being the best performance achieved when using the t-GAS model.
Procedures to forecast directly the realised measures were also applied. Procedures
applied to the logarithm of the realised measure showed a better performance than the
procedures applied directly to the realised measures. The best performance to forecast the
realised measures was obtained using the realised GARCH model.
In the non-robust procedures, the VaR estimation reported a large proportion of fails
18
and the backtesting tests reject the null hypothesis that the proportion of fails is equal to
1%, giving a misleading picture of what can be expected in the future. On the other hand,
the VaR estimated using the robust bootstrap procedure of Trucı́os et al. (2017) shown a
good performance and none of the test reject the null hypothesis.
In the absence of high-frequency data, we suggest using the t-GAS model to forecast the
volatility while when using high-frequency data the realised GARCH model seems to be a
good alternative. In a context of VaR estimation, we suggest using the robust bootstrap
procedure of Trucı́os et al. (2017). Extensions of this procedure to be applied in other
volatility models such as t-GAS or realised GARCH models is an interesting research topic
and could provide a good alternative to better estimate the VaR.
Additional comparisons considering different robust procedures as well as considering
switch regime models applied to cryptocurrencies (Ardia et al., 2018; Caporale & Zekokh,
2018) is another interesting research topics. In the same spirit, a more extensive comparison
using models to forecast realised measures is important to be conducted.
Some papers such as Dyhrberg (2016), Balcilar et al. (2017) and Cermak (2017) have also
considered explanatory variables to better estimate the volatility. Thus, one step further in
this research is to analyse the performance of the out-of-sample volatility forecast considering
explanatory variables and robust procedures.
This paper is in concordance with the results of Carnero et al. (2012), Trucı́os & Hotta
(2016) and Trucı́os et al. (2017) which show the dramatic effect of additive outliers in the
estimation and prediction of the volatility and VaR. For a good review about outliers in
GARCH models, we refer to Hotta & Trucı́os (2018).
In a multivariate framework Boudt et al. (2013), Grané et al. (2014) and Trucı́os et al.
(2018b) shown the dramatic effect of outliers in the estimation and prediction of the volatil-
ities and co-volatilities. In this sense, it is important that future studies considering Bitcoin
with other cryptocurrencies take into account the presence of outliers. In this sense, the
procedures proposed by, for instance, Croux et al. (2010), Boudt & Croux (2010), Boudt
et al. (2013), Iqbal (2013), Trucı́os et al. (2018a) and Trucı́os et al. (2018a) could be useful.
Other recent published paper are in concordance with our findings. For instance, Troster
et al. (2018) also concludes that GAS models have a better performance than the GARCH
19
models in a Bitcoin risk measure context. Finally, Chaim & Laurini (2018) find that the
presence of jumps should be taken into account when modelling the Bitcoin volatility.
The author acknowledges financial support from São Paulo Research Foundation (FAPESP)
grants 2016/18599-4 and 2018/03012-3 as well as the support of the Centre of Quantitative
Studies in Economics and Finance (CEQEF) and Centre for Applied Research on Econo-
metrics, Finance and Statistics (CAREFS). The author thanks helpful comments and sug-
gestions of João H. G. Mazzeu and Luiz K. Hotta, the editors and two anonymous referees.
References
Alizadeh, S., Brandt, M. W., & Diebold, F. X. (2002). Range-based estimation of stochastic
volatility models. The Journal of Finance, 57 (3), 1047–1091.
Andersen, T. G., & Bollerslev, T. (1998). Answering the skeptics: Yes, standard volatility
models do provide accurate forecasts. International Economic Review , (pp. 885–905).
Andersen, T. G., Bollerslev, T., & Diebold, F. X. (2007). Roughing it up: Including jump
components in the measurement, modeling, and forecasting of return volatility. The
Review of Economics and Statistics, 89 (4), 701–720.
Andersen, T. G., Bollerslev, T., Diebold, F. X., & Labys, P. (2003). Modeling and forecasting
realized volatility. Econometrica, 71 (2), 579–625.
Andersen, T. G., Dobrev, D., & Schaumburg, E. (2012). Jump-robust volatility estimation
using nearest neighbor truncation. Journal of Econometrics, 169 (1), 75–93.
Ardia, D., Bluteau, K., & Rüede, M. (2018). Regime changes in Bitcoin GARCH volatility
dynamics. Finance Research Letters.
Baek, C., & Elbeck, M. (2015). Bitcoins as an investment or speculative vehicle? a first
look. Applied Economics Letters, 22 (1), 30–34.
Balcilar, M., Bouri, E., Gupta, R., & Roubaud, D. (2017). Can volume predict Bitcoin
returns and volatility? a quantiles-based approach. Economic Modelling, 64 , 74–81.
Barndorff-Nielsen, O. E., & Shephard, N. (2004). Power and bipower variation with stochas-
tic volatility and jumps. Journal of Financial Econometrics, 2 (1), 1–37.
Bartlett, M. S. (1946). On the theoretical specification and sampling properties of auto-

correlated time-series. Supplement to the Journal of the Royal Statistical Society, 8 (1),
27–41.
URL http://www.jstor.org/stable/2983611
Bauwens, L., Hafner, C. M., & Laurent, S. (2012). Handbook of Volatility Models and Their
Applications, vol. 3. New Jersey: John Wiley & Sons.
20
Bernardi, L. C. . M. (2017). MCS: Model Confidence Set Procedure. R package version
0.1.3.
URL https://CRAN.R-project.org/package=MCS
Bezerra, P. C. S., & Albuquerque, P. H. M. (2017). Volatility forecasting via SVR–GARCH

with mixture of Gaussian kernels. Computational Management Science, 14 (2), 179–196.
Blasques, F., Koopman, S. J., & Lucas, A. (2017). Maximum likelihood estimation for
score-driven models. Timbergen Institute Discussion Paper 14-029/III .
Blazsek, S., & Villatoro, M. (2015). Is beta-t-EGARCH (1, 1) superior to GARCH (1, 1)?
Applied Economics, 47 (17), 1764–1774.
Bollerslev, T. (1986). Generalized autoregressive conditional heteroskedasticity. Journal of

Econometrics, 31 (3), 307–327.
Bollerslev, T. (2010). Glossary to ARCH (GARCH), Volatility and Time Series Economet-
rics: Essays in Honor of Robert F. Engle, Chapter 8.
Boudt, K., & Croux, C. (2010). Robust M-estimation of multivariate GARCH models.
Computational Statistics & Data Analysis, 54 (11), 2459–2469.
Boudt, K., Danielsson, J., & Laurent, S. (2013). Robust forecasting of dynamic conditional
correlation GARCH models. International Journal of Forecasting, 29 (2), 244–257.
Briere, M., Oosterlinck, K., & Szafarz, A. (2015). Virtual currency, tangible return: Portfolio
diversification with bitcoin. Journal of Asset Management, 16 (6), 365–373.
Caporale, G. M., & Zekokh, T. (2018). Modelling volatility of cryptocurrencies using

markov-switching GARCH models. CESifo Group Munich, No. 7167 .
Carnero, M. A., Peña, D., & Ruiz, E. (2012). Estimating GARCH volatility in the presence
of outliers. Economics Letters, 114 (1), 86–90.
Catania, L., & Grassi, S. (2017). Modelling crypto-currencies financial time-series. CEIS
Working Paper. Available at SSRN: https://ssrn.com/abstract=3084109 , 417 .
Catania, L., Grassi, S., & Ravazzolo, F. (2018). Predicting the volatility of cryptocurrency
time–series. CAMP Working Paper Series, 3 .
Cermak, V. (2017). Can bitcoin become a viable alternative to fiat currencies? an em-
pirical analysis of bitcoin’s volatility based on a GARCH model. Available at SSRN:
https://ssrn.com/abstract=2961405 .
Chaim, P., & Laurini, M. P. (2018). Volatility and return jumps in bitcoin. Economics
Letters, 173 , 158–163.
Chan, S., Chu, J., Nadarajah, S., & Osterrieder, J. (2017). A statistical analysis of cryp-
tocurrencies. Journal of Risk and Financial Management, 10 (2), 12.
Charles, A., & Darné, O. (2018). Volatility estimation for bitcoin: Replication and robust-
ness. International Economics.
Christoffersen, P., Feunou, B., Jacobs, K., & Meddahi, N. (2014). The economic value of
realized volatility: Using high-frequency returns for option valuation. Journal of Financial
and Quantitative Analysis, 49 (3), 663–697.
21
Christoffersen, P. F. (1998). Evaluating interval forecasts. International Economic Review ,
29 (4), 841–862.
Chu, J., Chan, S., Nadarajah, S., & Osterrieder, J. (2017). GARCH modelling of cryptocur-
rencies. Journal of Risk and Financial Management, 10 (4), 17.
Conrad, C., Custovic, A., & Ghysels, E. (2018). Long-and short-term cryptocurrency volatil-
ity components: A GARCH-MIDAS analysis. Journal of Risk and Financial Management,
11 (2), 23.
Corsi, F. (2009). A simple approximate long-memory model of realized volatility. Journal

of Financial Econometrics, 7 (2), 174–196.
Creal, D., Koopman, S. J., & Lucas, A. (2013). Generalized autoregressive score models
with applications. Journal of Applied Econometrics, 28 (5), 777–795.
Croux, C., Gelper, S., & Mahieu, K. (2010). Robust exponential smoothing of multivariate
time series. Computational Statistics & Data Analysis, 54 (12), 2999–3006.
Ding, Z., Granger, C. W., & Engle, R. F. (1993). A long memory property of stock market
returns and a new model. Journal of Empirical Finance, 1 (1), 83–106.
Dyhrberg, A. H. (2016). Bitcoin, gold and the dollar–a GARCH volatility analysis. Finance
Research Letters, 16 , 85–92.
Engle, R. F. (1982). Autoregressive conditional heteroscedasticity with estimates of the

variance of United Kingdom inflation. Econometrica, 50 (4), 987–1007.
Engle, R. F. (1995). ARCH: Selected Readings: Advanced Texts in Econometrics. Oxford:

Oxford University.
Engle, R. F., & Bollerslev, T. (1986). Modelling the persistence of conditional variances.
Econometric Reviews, 5 (1), 1–50.
Engle, R. F., Ghysels, E., & Sohn, B. (2013). Stock market volatility and macroeconomic
fundamentals. Review of Economics and Statistics, 95 (3), 776–797.
Engle, R. F., & Manganelli, S. (2004). CAViaR: Conditional autoregressive value at risk by
regression quantiles. Journal of Business & Economic Statistics, 22 (4), 367–381.
Engle, R. F., & Ng, V. K. (1993). Measuring and testing the impact of news on volatility.
The Journal of Finance, 48 (5), 1749–1778.
Francq, C., & Zakoı̈an, J.-M. (2009). Bartlett’s formula for a general class of nonlinear
processes. Journal of Time Series Analysis, 30 (4), 449–465.
Francq, C., & Zakoian, J.-M. (2011). GARCH Models: Structure, Statistical Inference and
Financial Applications. New Jersey: John Wiley & Sons.
Gao, C.-T., & Zhou, X.-H. (2016). Forecasting VaR and ES using dynamic conditional score
models and skew Student distribution. Economic Modelling, 53 , 216–223.
Ghalanos, A. (2018). Introduction to the rugarch package.(version 1.3-1). Tech. rep., avail-
able at http://cran. r-project. org/web/packages/rugarch.
22
Ghysels, E., Santa-Clara, P., & Valkanov, R. (2006). Predicting volatility: getting the most
out of return data sampled at different frequencies. Journal of Econometrics, 131 (1-2),
59–95.
Gkillas, K., & Katsiampa, P. (2018). An application of extreme value theory to cryptocur-
rencies. Economics Letters, 164 , 109–111.
Glosten, L. R., Jagannathan, R., & Runkle, D. E. (1993). On the relation between the
expected value and the volatility of the nominal excess return on stocks. The Journal of
Finance, 48 (5), 1779–1801.
González-Rivera, G., Lee, T.-H., & Mishra, S. (2004). Forecasting volatility: A reality
check based on option pricing, utility function, value-at-risk, and predictive likelihood.
International Journal of Forecasting, 20 (4), 629–645.
Grané, A., Veiga, H., & Martı́n-Barragán, B. (2014). Additive level outliers in multivariate
GARCH models. In V. Melas, S. Mignani, P. Monari, & L. Salmaso (Eds.) Topics in
Statistical Simulation, vol. 114, (pp. 247–255). New York: Springer.
Hansen, P. R., Huang, Z., & Shek, H. H. (2012). Realized GARCH: a joint model for returns
and realized measures of volatility. Journal of Applied Econometrics, 27 (6), 877–906.
Hansen, P. R., Lunde, A., & Nason, J. M. (2011). The model confidence set. Econometrica,
79 (2), 453–497.
Harvey, A., & Chakravarty, T. (2008). Beta-t-(E)GARCH. Cambridge Working Papers in

Economics CWPE0840, University of Cambridge.
Harvey, A., & Sucarrat, G. (2014). EGARCH models with fat tails, skewness and leverage.
Computational Statistics & Data Analysis, 76 , 320–338.
Harvey, A. C. (2013). Dynamic mModels for Volatility and Heavy Tails: With Applications
to Financial and Economic Time Series, vol. 52. New York: Cambridge University Press.
Hentschel, L. (1995). All in the family nesting symmetric and asymmetric GARCH models.
Journal of Financial Economics, 39 (1), 71–104.
Higgins, M. L., & Bera, A. K. (1992). A class of nonlinear arch models. International
Economic Review , (pp. 137–158).
Hotta, L. K., & Trucı́os, C. (2018). Inference in (M)GARCH Models in the Presence of
Additive Outliers: Specification, Estimation and Prediction. In C. Lavor, & F. A. M. G.
Neto (Eds.) Advances in Mathematics and Applications, (pp. 179–202). Springer.
Iqbal, F. (2013). Robust estimation for the orthogonal GARCH model. The Manchester
School , 81 (6), 904–924.
Katsiampa, P. (2017). Volatility estimation for bitcoin: A comparison of GARCH models.

Economics Letters, 158 , 3–6.
Koopman, S. J., Jungbacker, B., & Hol, E. (2005). Forecasting daily variability of the S&P
100 stock index using historical, realised and implied volatility measurements. Journal of
Empirical Finance, 12 (3), 445–475.
Kupiec, P. H. (1995). Techniques for verifying the accuracy of risk measurement models.
The Journal of Derivatives, 3 (2), 73–84.
23
Laurent, S., Lecourt, C., & Palm, F. C. (2016). Testing for jumps in conditionally Gaussian
ARMA–GARCH models, a robust approach. Computational Statistics & Data Analysis,
100 , 383–400.
Lee, G., & Engle, R. (1999). A permanent and transitory component model of stock return
volatility. In Cointegration Causality and Forecasting A Festschrift in Honor of Clive WJ
Granger , (pp. 475–497). New York: Oxford University Press.
Liu, L. Y., Patton, A. J., & Sheppard, K. (2015). Does anything beat 5-minute RV? a
comparison of realized measures across multiple asset classes. Journal of Econometrics,
187 (1), 293–311.
Liu, R., Shao, Z., Wei, G., & Wang, W. (2017). GARCH model with fat-tailed distributions
and bitcoin exchange rate returns. Journal of Accounting, Business and Finance Research,
1 (1), 71–75.
McAleer, M., & Medeiros, M. C. (2008). Realized volatility: A review. Econometric Reviews,
27 (1-3), 10–45.
Muler, N., & Yohai, V. J. (2008). Robust estimates for GARCH models. Journal of Statis-
tical Planning and Inference, 138 (10), 2918–2940.
Naimy, V. Y., & Hayek, M. R. (2018). Modelling and predicting the bitcoin volatility
using GARCH models. International Journal of Mathematical Modelling and Numerical
Optimisation, 8 (3), 197–215.
Nakamoto, S. (2008). Bitcoin: A peer-to-peer electronic cash system. Manuscript available

at https://bitcoin.org/bitcoin.pdf .
Nelson, D. B. (1991). Conditional heteroskedasticity in asset returns: A new approach.

Econometrica, 59 (2), 347–370.
Osterrieder, J., & Lorenz, J. (2017). A statistical risk assessment of bitcoin and its extreme
tail behavior. Annals of Financial Economics, 12 (01), 1–19.
Patton, A. J. (2011). Volatility forecast comparison using imperfect volatility proxies. Jour-
nal of Econometrics, 160 (1), 246–256.
Patton, A. J., & Sheppard, K. (2009). Optimal combinations of realised volatility estimators.
International Journal of Forecasting, 25 (2), 218–238.
Peng, Y., Albuquerque, P. H. M., de Sá, J. M. C., Padula, A. J. A., & Montenegro, M. R.
(2018). The best of two worlds: Forecasting high frequency volatility for cryptocurrencies
and traditional currencies with support vector regression. Expert Systems with Applica-
tions, 97 , 177–192.
Pichl, L., & Kaizoji, T. (2017). Volatility analysis of bitcoin price time series. Quantitative
Finance and Economics, 1 (4), 474–485.
Rodrı́guez, M. J., & Ruiz, E. (2012). Revisiting several popular GARCH models with
leverage effect: Differences and similarities. Journal of Financial Econometrics, 10 (4),
637–668.
Sapuric, S., & Kokkinaki, A. (2014). Bitcoin is volatile! isn’t that right? In International
Conference on Business Information Systems, (pp. 255–265). Springer.
24
Schwert, G. W. (1990). Stock volatility and the crash of ’87. The Review of Financial
Studies, 3 (1), 77–102.
Shephard, N., & Sheppard, K. (2010). Realising the future: forecasting with high-frequency-
based volatility (HEAVY) models. Journal of Applied Econometrics, 25 (2), 197–231.
Stavroyiannis, S. (2018). Value-at-risk and related measures for the bitcoin. The Journal
of Risk Finance, 19 (2), 127–136.
Teräsvirta, T. (2009). An introduction to univariate GARCH models. In T. Mikosch, J.-P.

Kreiß, R. A. Davis, & T. G. Andersen (Eds.) Handbook of Financial Time Series, (pp.
17–42). Berlin, Heidelberg: Springer Berlin Heidelberg.
Troster, V., Tiwari, A. K., Shahbaz, M., & Macedo, D. N. (2018). Bitcoin returns and risk:
A general GARCH and GAS analysis. Finance Research Letters.
Trucı́os, C., & Hotta, L. K. (2016). Bootstrap prediction in univariate volatility models
with leverage effect. Mathematics and Computers in Simulation, 120 , 91–103.
Trucı́os, C., Hotta, L. K., & Pereira, P. L. V. (2018a). On the robustness of the prin-
cipal volatility components. CEQEF Working Paper Series 47 available at SSRN:
https://ssrn.com/abstract=3143870 .
Trucı́os, C., Hotta, L. K., & Ruiz, E. (2015). Robust bootstrap forecast densities for GARCH
models: returns, volatilities and value-at-risk. UC3M Working Papers Statistics and
Econometrics, 15 (22).
Trucı́os, C., Hotta, L. K., & Ruiz, E. (2017). Robust bootstrap forecast densities for GARCH
returns and volatilities. Journal of Statistical Computation and Simulation, 87 (16), 3152–
3174.
Trucı́os, C., Hotta, L. K., & Ruiz, E. (2018b). Robust bootstrap densities for dynamic
conditional correlations: implications for portfolio selection and value-at-risk. Journal of
Statistical Computation and Simulation, 88 (10), 1976–2000.
Walther, T., & Klein, T. (2018). Exogenous drivers of cryptocurrency volatility-a mixed data
sampling approach to forecasting. University of St.Gallen, School of Finance Research
Paper No. 2018/19. Available at SSRN: https://ssrn.com/abstract=3192474 .
Zakoian, J.-M. (1994). Threshold heteroskedastic models. Journal of Economic Dynamics

and Control , 18 (5), 931–955.
25

Forecasting Bitcoin Risk Measures: A Robust Approach

Hochgeladen von

Dokumentinformationen

Copyright

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Forecasting Bitcoin Risk Measures: A Robust Approach

Hochgeladen von

Copyright:

Forecasting Bitcoin risk measures: A robust approach

São Paulo School of Economics - FGV, Brazil

December 12, 2018

predictability of the one-step-ahead volatility and Value-at-Risk of Bitcoin using sev-

Keywords: Cryptocurrency, GARCH, Model confidence set, Outliers, Realised volatility,

2.1 Volatility models

2.1.1 GARCH models

Table 1: GARCH-type models

qt+1 = ω + ρqt + φ(rt2 − σt2 ) Lee & Engle (1999)

2.1.3 Realised GARCH model

log(xt ) = δ + γ log(σt2 ) + τ (t ) + ut

2.1.4 GAS models

2.2 Realised measures

2.3 Robust loss function

2.4 VaR estimation

rt∗ = σt∗ ∗t ,

• Step 3: Obtain h-steps-ahead forecasts as

r̂T∗ +h|T = ∗T +h σ̂T∗ +h|T ,

Table 3: Descriptive statistics of Bitcoin daily returns

2011−09 2013−02 2014−06 2015−11 2017−03

jump-filtered returns as in Laurent et al. (2016), we estimate the conditional variance in

4 Conclusions and future works

Bartlett, M. S. (1946). On the theoretical specification and sampling properties of auto-

Bezerra, P. C. S., & Albuquerque, P. H. M. (2017). Volatility forecasting via SVR–GARCH

Bollerslev, T. (1986). Generalized autoregressive conditional heteroskedasticity. Journal of

Caporale, G. M., & Zekokh, T. (2018). Modelling volatility of cryptocurrencies using

Corsi, F. (2009). A simple approximate long-memory model of realized volatility. Journal

Engle, R. F. (1982). Autoregressive conditional heteroscedasticity with estimates of the

Engle, R. F. (1995). ARCH: Selected Readings: Advanced Texts in Econometrics. Oxford:

Harvey, A., & Chakravarty, T. (2008). Beta-t-(E)GARCH. Cambridge Working Papers in

Katsiampa, P. (2017). Volatility estimation for bitcoin: A comparison of GARCH models.

Nakamoto, S. (2008). Bitcoin: A peer-to-peer electronic cash system. Manuscript available

Nelson, D. B. (1991). Conditional heteroskedasticity in asset returns: A new approach.

Teräsvirta, T. (2009). An introduction to univariate GARCH models. In T. Mikosch, J.-P.

Zakoian, J.-M. (1994). Threshold heteroskedastic models. Journal of Economic Dynamics

Das könnte Ihnen auch gefallen

log(xt ) = δ + γ log(σt2 ) + τ (t ) + ut

rt∗ = σt∗ ∗t ,

r̂T∗ +h|T = ∗T +h σ̂T∗ +h|T ,