Sie sind auf Seite 1von 8

Statistics and Probability Letters 80 (2010) 26–33

Contents lists available at ScienceDirect

Statistics and Probability Letters


journal homepage: www.elsevier.com/locate/stapro

Explosive volatilities for threshold-GARCH processes generated by


asymmetric innovations
S.Y. Hwang ∗ , J.S. Baek, J.A. Park, M.S. Choi
Sookmyung Women’s University, Seoul, Republic of Korea

article info abstract


Article history: The threshold-asymmetric GARCH (TGARCH, for short) models have been useful for analyz-
Received 26 May 2009 ing asymmetric volatilities arising mainly from financial time series. Most of the research
Received in revised form 12 September on TGARCH has been directed to the stationary case. In this article, motivated by unsta-
2009
ble features in recent time series in Korea amid worldwide financial crisis, we introduce
Accepted 14 September 2009
Available online 18 October 2009
‘‘explosive volatilities’’ in TGARCH processes. The term of explosive volatility in TGARCH
context is defined and is justified. Moreover, asymmetric innovations such as normal mix-
MSC:
tures are considered in modeling explosive TGARCH and hence we are concerned with a
primary 62M10 class of explosive TGARCH models generated by asymmetric innovations. Assuming nor-
mal mixture innovations, maximum likelihood (ML) estimation method is discussed and
procedures for computing ML-estimates are described. To illustrate, exchange rate data of
Korea–Won to US dollars are analyzed and it is observed that the data exhibit a certain ex-
plosive volatility and in turn, our model performs better than various competing models.
© 2009 Elsevier B.V. All rights reserved.

1. Motivation and introduction

Since the seminal paper of Engle (1982), the ARCH (GARCH) processes have served as useful models for analyzing
volatility (conditional variance, denoted by ht throughout) inherent in time series data. It is noted that these ARCH (GARCH)
processes belong to symmetric models for which volatility is formulated as a linear function of squared past values. In
the past two decades, there has been growing interest in threshold-asymmetric models in the GARCH context in a natural
response to the empirical evidences of asymmetric volatilities arising mainly from financial time series. News-impact-curves
discussed by Engel and Ng (1993) can be employed as a simple but efficient tool to detect asymmetric volatilities. Gourieroux
(1997, p. 90) argued that heteroscedasticity effect depends on whether the recent past values are positive or negative,
necessitating threshold-asymmetric GARCH (TGARCH, hereafter) models. The time series {Xt } is said to follow the first order
TGARCH, i.e., TGARCH(1, 1) model when
p
Xt = ht · e t (1.1)
ht = α0 + α1 (Xt+−1 )2 + α2 (Xt−−1 )2 + β ht −1
where α0 > 0, α1 , α2 ≥ 0, β ≥ 0 and the notation

x+ = max(x, 0) and x− = min(−x, 0) (1.2)


will be used throughout. Here, the innovation {et } is a sequence of iid random variables with mean zero and unit variance.
Refer to, among others, Higgins and Bera (1992), Li and Li (1996), Hwang and Basawa (2004) and Pan et al. (2008) for
discussions on probabilistic structures, inferential problems and data applications for TGARCH models.

∗ Corresponding author.
E-mail address: shwang@sookmyung.ac.kr (S.Y. Hwang).

0167-7152/$ – see front matter © 2009 Elsevier B.V. All rights reserved.
doi:10.1016/j.spl.2009.09.008
S.Y. Hwang et al. / Statistics and Probability Letters 80 (2010) 26–33 27

Most of the research on TGARCH models, however, has been directed to stable (in volatility) case where the volatility
process {ht } is a covariance-stationary process and therefore l-step ahead volatility converges almost surely to a constant
which turns out to be stationary moment E (ht ) as the lead time l goes to infinity. Assuming for the time being that the
innovation et is symmetrically distributed, for instance, as standard normal distribution and standardized t-distributions,
consider the key parameter φ given by
φ = (α1 + α2 )/2 + β. (1.3)
The stable case corresponds to φ < 1. Recently, in analogy with extensions of GARCH models to IGARCH models by Nelson
(1990) and Park et al. (2009) proposed and investigated a ‘‘integrated’’ TGARCH process (I-TGARCH, say) for which φ = 1.
They showed that I-TGARCH demonstrates a persistent property in volatility in the sense that current volatility ht continues
to remain for all l-step ahead volatilities contrary to the stable case (φ < 1).
This article is mainly concerned with the ‘‘explosive in volatility’’ case of φ > 1 in TGARCH models where the effect
of current volatility ht on future l-step ahead volatilities is explosively growing with the lead time l. This case will be
interesting in its own right in theoretical aspects, complementing the literature in TGARCH context. In fact, ‘‘explosive’’
case was motivated by current financial time series in Korea amid worldwide financial crisis. We have observed a feature
of threshold-asymmetric ‘‘explosive’’ volatility from the recent exchange rate data of Korea-won to US dollar (USD). See
Section 4 for detailed analysis. Further, we are willing to extend TGARCH class by allowing asymmetric innovations including
mixture of normal distributions. See, for instance, Bai et al. (2003) and Lee and Lee (2009) for discussions on asymmetric
innovations in standard GARCH models. Consequently, this article is concerned with a class of explosive TGARCH processes
generated by asymmetric innovations. We proceed as follows. The model is specified and explosive properties (in volatility)
of the model are obtained in Section 2. The asymmetric case when innovation et follows a mixture of normals is discussed
and related algorithm for obtaining parameter estimates are presented in Section 3. To illustrate, an empirical analysis for the
exchange rate of Korea–Won to USD is conducted in Section 4. Section 5 addresses concluding remarks regarding possible
extensions to higher order models.

2. The model specification and explosive volatility

Consider the TGARCH(1, 1) time series {Xt } defined in (1.1) with the conditional variance (i.e., volatility ht ) of Xt given
the past, which is formulated by
p
Xt = ht · et with ht = α0 + α1 (Xt+−1 )2 + α2 (Xt−−1 )2 + β ht −1 (2.1)
where α0 > 0, α1 , α2 ≥ 0, β ≥ 0. Note that α1 = α2 (= α, say) in (2.1) reduces to the standard GARCH(1, 1) process. It will
be assumed throughout that the innovation process {et } is a sequence of iid random variables with mean zero and variance
unity, and
(C1) The distribution of et is absolutely continuous with respect to Lebesgue measure.
Note that P (et = 0) = 0 under (C1). Define a measure of asymmetry τe given by

τe = E (e−
t )
2
(2.2)
and note that E (et ) = 1 − τe . In particular when et is symmetrically distributed, τe = 1/2 and thus a value of τe away from
+ 2

1/2 indicates asymmetry of et . In this paper, et is not necessarily symmetrically distributed and asymmetric case (τe 6= 1/2)
will be focused on. The key parameter φ is then defined by
θ = (1 − τe )α1 + τe α2 + β (2.3)
which is non-negative. When τe = 1/2, θ reduces to φ = (α1 + α2 )/2 + β in (1.3).
It can be verified that {Xt } following TGARCH(1, 1) permits a unique strictly stationary solution with finite second moment
if and only if θ < 1. Refer to Theorem 1 of Hwang and Basawa (2004). See also Theorem 6 in Pan et al. (2008). For the
integrated case (θ = 1), Park et al. (2009) identified persistent volatilities for TGARCH(1, 1) models only assuming symmetric
et , i.e., τe = 1/2. Their arguments on persistent volatility may be extended to cover possibly asymmetric et case of θ = 1.
This article is concerned with the remaining case.

Definition (Explosive TGARCH(1, 1)). The TGARCH(1, 1) process is said to be explosive in volatility when the key parameter
θ > 1.
To justify the term ‘‘explosive’’, note first that TGARCH(1,1) model in (2.1) can be rewritten in terms of ARMA(1,1) time
series with random coefficient. To see this, let Ft denote the σ -field generated by Xt , Xt −1 , . . .. Note that ht = E (Xt2 |Ft −1 )
and consider the following sequence of prediction errors {ηt } for the squared observation Xt2 .

ηt = Xt2 − ht (2.4)
which constitutes a sequence of zero mean martingale differences with respect to {Ft }. Substituting (2.4) for (2.1), we obtain

Xt2 = α0 + wt −1 Xt2−1 + ηt − βηt −1 (2.5)


28 S.Y. Hwang et al. / Statistics and Probability Letters 80 (2010) 26–33

where the random coefficient wt −1 (as a function of Xt −1 ) is given by

wt −1 = α1 + (α2 − α1 )I[Xt −1 <0] + β (2.6)

where I[·] denotes indicator function.


Now, fix t and define l-step ahead volatility ht (l) as

ht (l) = Var(Xt +l |Ft ) = E (Xt2+l |Ft ), l ≥ 1. (2.7)

Note that ht (1) = ht +1 . In view of the ARMA (with random coefficient) representation (2.5) of the TGARCH(1, 1), {ηt } plays
a role of innovations for the series {Xt2 } of squared observations. Thus, the partial derivative

λl = ∂ ht (l)/∂ηt (2.8)

measures a certain contribution of innovation ηt (at time t) to the l-step ahead volatility. Baillie et al. (1996) referred to λl
as cumulative impulse response (in volatility) and noted that λl goes to zero as l tends to infinity for a class of stable GARCH
processes. It will be shown however that explosive TGARCH(1, 1) entails explosive cumulative impulse response λl as l goes
to infinity, justifying the term ‘‘explosive in volatility’’.

Theorem 2.1. Fix t and for the explosive TGARCH (1, 1), we obtain

(i) The current volatility is given by ht (1) = α0 + wt Xt2 − βηt with wt in (2.6).
(ii) ht (l) = (1 + θ + θ 2 + · · · + θ l−2 )α0 + θ l−1 ht (1), l ≥ 2.
(iii) λl = α1 θ l−1 I[Xt >0] + α2 θ l−1 I[Xt <0] , l ≥ 2.

Remarks. It follows from (ii) that the effect of current volatility ht (1) = ht +1 on the future volatility is indeed explosive. In
addition, λl goes exponentially to infinity (almost surely) as l tends to infinity.

Proof. (i) Note that wt and ηt are Ft -measurable and it readily follows from (2.5).

ht (1) = E (Xt2+1 |Ft ) = α0 + wt Xt2 − βηt .

For (ii), consider

ht (2) = E (Xt2+2 |Ft ) = E [E (Xt2+2 |Ft +1 )|Ft ]


= E [E (α0 + wt +1 Xt2+1 + ηt +2 − βηt +1 |Ft +1 )|Ft ]
= α0 + E (wt +1 Xt2+1 |Ft ).

Since

E (wt +1 Xt2+1 |Ft ) = E [α1 Xt2+1 + (α2 − α1 )(Xt−+1 )2 + β Xt2+1 |Ft ]


= α1 ht (1) + (α2 − α1 ) · τe · ht +1 + β ht (1) = θ ht (1),

we conclude

ht (2) = α0 + θ ht (1).

Similarly, one can obtain recursively ht (l) = α0 + θ ht (l − 1), for l ≥ 3 which in turn gives (ii).
(iii) For l = 1, since ht (1) = α0 + wt Xt2 − βηt , via partial derivative with respect to ηt , we have λ1 = wt − β where
∂ Xt2 /∂ηt = 1 is used. For l ≥ 2, it follows from (ii) that

∂ ht (l)/∂ηt = θ l−1 (∂ ht (1)/∂ηt ).

Consequently, λl = θ l−1 λ1 = θ l−1 (wt − β), l ≥ 2 which readily yields (iii). 

As an example of asymmetric innovations, we consider in the next section a mixture of two normals for et . Refer to
Bai et al. (2003) and Lee and Lee (2009) for applications of normal mixtures in the context of standard stationary-GARCH
processes.
S.Y. Hwang et al. / Statistics and Probability Letters 80 (2010) 26–33 29

3. Innovations following normal mixtures

In order to reflect heavy-tailed and asymmetric distribution, compared to the standard normal, let {et } be a sequence of
iid random variables with two-component normal mixture density f (·) which is formulated as
2
X
f (·) = pk fk (·; µk , σk2 ) (3.1)
k=1

where p1 + p2 = 1 and fk denotes normal density with mean µk and variance σk2 (k = 1, 2), i.e.

1 h (· − µ )2 i
k
fk (·; µk , σk2 ) = q exp − (3.2)
2π σk2 2σk2

satisfying

p1 µ1 + p2 µ2 = 0 and p1 σ12 + p2 σ22 = 1. (3.3)


Two normal components in (3.1) may be easily extended to multi-component case and thus we will retain (3.1) for simplicity.
Note that τe in (2.2) turns out to be
Z 0
τe = [p1 x2 f1 (x) + p2 x2 f2 (x)]dx = p1 E (Z1− )2 + p2 E (Z2− )2 (3.4)
−∞

where Zk is used for denoting N (µk , σk2 ) random variable (k = 1, 2).


Assume that n-observations X1 , . . . , Xn are available. Collecting parameters appearing in ht , denote ζ1 = (α0 , α1 , α2 , β)
and write ht (ζ1 ) for ht in order to emphasize the dependence on ζ1 . The other parameters coming from normal mixture
density (3.1) are denoted by ζ2 = (p1 , p2 , µ1 , µ2 , σ12 , σ22 ). The log-likelihood based on the data is given by

 (X − µ √h (ζ ))2 
( )
n 2
X X t k t 11
L(ζ1 , ζ2 ) = log pk q exp − . (3.5)
t =1 k=1 2π σk2 ht (ζ1 ) 2σk2 ht (ζ1 )

The maximum likelihood estimator (MLE) ζˆn = (ζˆ1n , ζˆ2n ) of ζ = (ζ1 , ζ2 ) is obtained by maximizing L(ζ1 , ζ2 ). It
is noted that normal mixture structure involves additional parameter ζ2 and this adds complexity to L(ζ1 , ζ2 ), thereby
requiring EM algorithm (cf. Demster et al. (1977)). Refer to, for instance, Pawitan (2001, Ch. 12) for illustrative EM algorithm
related to normal mixture models. In addition, to obtain a refined value of ζˆ1n of ζ1 , iterative Newton–Raphson algorithm
is employed. The procedures for obtaining MLE ζˆn are presented below for interested readers to pursue. In the context of
standard stationary-GARCH processes, Lee and Lee (2009) discussed asymptotic normality of quasi-MLE (QMLE) obtained
by maximizing normal mixture likelihood and suggested algorithm for computing QMLE. Algorithm given below is a
modification and an adaptation of that in Lee and Lee (2009) to our case of the explosive TGARCH(1, 1) processes.
[S1] Start with an appropriate initial value ζ˜1n and construct residuals
Xt
e˜t = q , t = 1, 2, . . . , n.
ht (ζ˜1n )

[S2] Applying the EM algorithm to the residuals e˜1 , . . . , e˜n , obtain ζ˜2n for mixture parameter ζ2 and set the initial values
(0) (0)
ζˆ1 = ζ˜1n , ζˆ2 = ζ˜2n .
(i+1) (i+1)
In [S3] and [S4] below, we refine ζˆ1 , ζˆ2 iteratively through i = 0, 1, 2, . . . .
(i+1) (i+1) (i+1) (i+1)
[S3] Compute ζˆ2 = (pˆ1 , pˆ2 , µ̂2 (i+1) , σˆ1 2(i+1) , σˆ2 2(i+1) ) via
, µ̂1
(i) (i)
pˆk fk (et ; ζˆ2 )
q
(i) (i)
at , k = with et = Xt / ht (ζˆ1 )
2 (i)
(i)
p̂j fj (et ; ζˆ2 )
P
j =1

(i+1) 1
Pn (i)
and in turn pˆk = n t =1 at ,k , and with the restriction (3.3)
n  X −µ̂(i+1) (i)
q
n  ht (ζ̂1 )
2
(i) (i)

P Xt
P t k
q
(i)
at , k q
(i)
at , k
ht (ζ̂1 ) ht (ζ̂1 )
µ̂k(i+1) = σ̂k2(i+1) =
t =1 t =1
n
and n
.
P (i) P (i)
at , k at , k
t =1 t =1
30 S.Y. Hwang et al. / Statistics and Probability Letters 80 (2010) 26–33

0.8
0.6
Density
0.4
0.2
0.0

-4 -2 0 2 4
bandwidth=0.1

Fig. 1. Histogram and kernel density of estimated innovations from (4.2) of K1 .

(i+1)
[S4] To obtain ζ1 = (α0(i+1) , α1(i+1) , α2(i+1) , β (i+1) ), we use the following procedures of Newton–Raphson algorithm
 ∂ 2 L(ζ (i) , ζ (i) ) −1  ∂ L(ζ (i) , ζ (i) ) 
ζ1(i+1) = ζ1(i) + − 1 2 1 2
,
∂ζ1 ∂ζ10 ∂ζ1
where

∂ L(ζ1 , ζ2 ) X n X 2  1  ∂ h (ζ ) n  1  X − µ √h (ζ )  o
t 1 t k t 1
= at , k 1− Xt
∂ζ1 t =1 k=1
h t (ζ 1 ) ∂ζ1 σ 2
k
h t (ζ 1 )
and

∂ 2 l(ζ1 , ζ2 ) X n X 2  1 nh 1  ∂ ht (ζ1 ) 2  ∂ 2 ht (ζ1 ) ih Xt (Xt − µk ht (ζ1 )) i
= a t , k − + 1 −
∂ζ1 ∂ζ10 t =1 k =1
ht (ζ1 ) ht (ζ1 ) ∂ζ1 ∂ζ1 ∂ζ10 σk2 ht (ζ1 )
 1  ∂ h (ζ ) 2  X  √
t 1 t µk ht (ζ1 ) o
+ 2 X t − .
ht (ζ1 ) ∂ζ1 σk2 2

[S5] Iterative steps [S3] and [S4] stop when the difference between ζ (i+1) and ζ (i) is deemed acceptably small and the final
value is regarded as the MLE ζˆn of ζ .

4. Illustrative data analysis : exchange rate data of Korea–Won to US dollar

Recent financial crisis has added further volatilities to financial time series in Korea and in particular, foreign exchange
rate data exhibit unstable patterns in volatility. This section considers the daily exchange rate data of Korea–Won to US
dollar (USD) in a span from January 3, 2008 to November 28, 2008, consisting of 226 observations. Let Wt denote the
daily closing Korea–Won which is equivalent to one USD. The log-return series, i.e., log differenced series is denoted by
Yt = log Wt − log Wt −1 . The first order autocorrelation of the log-return series {Yt } is calculated as 0.33589 which is
significantly different from zero. To whitening {Yt }, i.e., to delete autocorrelation from {Yt }, define

Xt = (Yt − Ȳ ) − 0.33589(Yt −1 − Ȳ ) (4.1)


Xt2
where Ȳ is the average of {Yt }. It is found via Portmanteau tests that the squared series { } is serially correlated whereas
{Xt } is uncorrelated series, and the kurtosis of {Xt } is identified as 12.06 which is considerably greater than 3 (which is the
reference value of normal distribution), indicating a leptokurtic pattern of {Xt }.
First consider the TGARCH(1,1) model (denoted by K1 ) for {Xt } with standard normal innovations {et }. The estimated
model is given by

K1 : ht = 0.036234(10−4 ) + 0.640138(x+
t −1 ) + 0.121498(xt −1 ) + 0.644650ht −1
2 − 2
(4.2)
which readily yields the key parameter value θ = 1.0255, implying explosive volatility. Note that τe = 1/2 due to symmetric
innovation. It is also noted that α1 is estimated as 0.640138 which is approximately five times larger than 0.121498 for α2
and this validates threshold modeling. The estimated innovation series {et } is obtained after fitting K1 and [Fig. 1] shows
histogram and corresponding kernel density (with bandwidth = 0.1) of estimated innovations from (4.2), suggesting instead
a normal mixture model for et .
S.Y. Hwang et al. / Statistics and Probability Letters 80 (2010) 26–33 31

Table 1
Summary statistics.
θ τe Log-likelihood AIC BIC

KM 1.0471 0.4582 766.9136 −1519.827 −1495.883


K1 1.0255 0.5000 758.3900 −1506.780 −1489.677
K2 1.0000 0.5000 757.7510 −1507.502 −1493.820

We are thus led to consider a TGARCH(1,1) model with two-component normal mixture innovation et defined in (3.1)
and (3.2). This is referred to as KM . The steps described in the previous section are implemented, using estimated values in
(4.2) of K1 for the starting value ζ˜1n in [S1]. The estimated equation has the form

KM : ht = 0.04(10−4 ) + 0.6401384(x+
t −1 ) + 0.1214984(xt −1 ) + 0.6446504ht −1
2 − 2
(4.3)

and ζ2 = (p1 , p2 , µ1 , µ2 , σ , σ ) is identified as


2
1
2
2

(p1 , p2 , µ1 , µ2 , σ12 , σ22 ) = (0.6548, 0.3452, − 0.1496, 0.2837, 0.3977, 2.1426) (4.4)
which in turn gives θ = 1.0471 and τe = 0.4582. The evidence of asymmetry of innovation is reflected in the value of
τe = 0.4582, and the key parameter θ = 1.0471 again supports explosive volatility in the data. Regarding ζ1 , we admit
that it may not be possible to distinguish between (4.2) and (4.3) except the last digits. It is worth noting however that the
difference in the key parameter θ and asymmetry measure τe appears to be considerable between the two models KM and K1 .
It is revealed that KM dominates K1 according to the log-likelihood, AIC and BIC criterion. See Table 1.
Another model competing with KM is considered. The case of θ = 1 is characterized by the persistent volatility. This was
investigated and named as integrated-TGARCH (I-TGARCH) by Park et al. (2009). The I-TGARCH(1, 1) model with standard
normal innovations (denoted by K2 ) is applied to obtain

K2 : ht = 0.037600(10−4 ) + 0.583255(x+
t −1 ) + 0.127792(xt −1 ) + 0.644476ht −1
2 − 2
(4.5)
for which θ is given by one.
Table 1 summarizes θ and τe for each model and reports the corresponding log-likelihood, AIC and BIC statistics. As is
seen from Table 1, explosive volatility is shown for KM and K1 . It is also found that KM has a better fit than the other models
in terms of the maximum log-likelihood and the minimum AIC and BIC.
The most useful feature of the model KM is to capture a certain type of explosive volatilities as is seen from Fig. 2. The
forecast lead time is chosen as l = 1, 5, 22 and 66 corresponding respectively to daily, weekly, monthly and quarterly
forecasts. It is noted from Fig. 2 that ht (l) tends to become large as l increases, indicating explosive volatilities and the
‘‘explosiveness’’ does not seem to be severe, in view of θ = 1.0471. In each four panel, the recent part is much more volatile,
mainly due to the bankruptcy of Lehman Brothers at September 2009. If one were to mistakenly employ stable (θ < 1)
TGARCH(1, 1), ht (l) would be flat, converging to a constant, in particular when the lead time l is large.

5. Concluding remarks

We have defined explosive TGARCH(1, 1) models with possibly asymmetric innovations via a measure of asymmetry τe
and the key parameter θ in (2.3). Extensions to higher order models can be made as follows. The TGARCH(p, q) process is of
the form
p
X q
X
ht = α0 + [αi1 (Xt+−i )2 + αi2 (Xt−−i )2 ] + βj ht −j . (5.1)
i =1 j =1

The key parameter θ is then given by


p
X p
X q
X
θ = (1 − τe ) αi1 + τe αi2 + βj (5.2)
i =1 i=1 j =1

and TGARCH(p, q) process is defined to be explosive when θ > 1. A broader class of Box–Cox transformed TGARCH processes
is defined by
p q
δ
hδt = α0 + [αi1 (Xt+−i )2δ + αi2 (Xt−−i )2 ] + βj hδt −j
X X
(5.3)
i=1 j =1

where δ > 0 denotes power transformation. Refer to Pan et al. (2008) and Hwang and Kim (2004). Eq. (5.3) requires a
modification of τe = E (e−
t ) in (2.2) and thus define
2

p p q

X X X
τe (δ) = E (e−
t ) and θ (δ) = (1 − τe (δ)) αi1 + τe (δ) αi2 + βj (5.4)
i=1 i=1 j =1
32 S.Y. Hwang et al. / Statistics and Probability Letters 80 (2010) 26–33

Fig. 2. Estimated volatilities ht (l) from the model KM for l = 1, 5, 22 and 66.

where θ (δ) with δ = 1 reduces to θ in (5.2). The class of Box–Cox transformed TGARCH processes specified by (5.3) is
referred to as explosive, persistent and stable according to θ (δ) > 1, θ (δ) = 1 and θ (δ) < 1. When θ (δ) > 1, explosive
properties for (5.3) can be derived analogously to Theorem 2.1. Details are omitted.

Acknowledgements

We thank the referee for the careful reading of the paper. This work was supported by a grant from the National Research
Foundation of Korea (NRF-2009-0084772).

References

Bai, X., Russell, J.R., Tiao, G.C., 2003. Kurtosis of GARCH and stochasitic volatility models with non-normal innovations. Journal of Econometrics 114, 349–360.
Baillie, R.T., Bollerslev, T., Mikkelsen, H.O., 1996. Fractionally integrated generalized autoregressive conditional heteroskedasticity. Journal of Econometrics
74, 3–30.
Demster, A.P., Laird, N.M., Rubin, D.B., 1977. Maximum-likilihood from imcomplete data via the EM algorithm. Journal of the Royal Statistical Series 39,
1–38.
Engle, R.F., 1982. Autoregressive conditional heteroscedasticity with estimates of the variance of UK inflation. Econometrika 50, 987–1008.
Engel, R.F., Ng, V.K., 1993. Measuring and testing the impact of news on volatility. Journal of Finance 48, 1749–1778.
S.Y. Hwang et al. / Statistics and Probability Letters 80 (2010) 26–33 33

Gourieroux, C., 1997. ARCH Models and Financial Applications. Springer-Verlag, New York.
Hwang, S.Y., Basawa, I.V., 2004. Stationarity and moment structure for Box-Cox transformed threshold GARCH (1, 1) processes. Statistics & Probability
Letters 68, 209–220.
Hwang, S.Y., Kim, T.Y., 2004. Power transformation and threshold modeling for ARCH innovations with applications to tests for ARCH sturcture. Stochastic
Processes and their Applications 110, 295–314.
Higgins, M.L., Bera, A.K., 1992. A class of nonlinear ARCH models. Internatational Economic Review 33, 137–158.
Lee, T.W., Lee, S., 2009. Normal mixture quasi-maximum likilihood estimator for GARCH models. Scandinavian Journal of Statistics 36, 157–170.
Li, C.W., Li, W.K., 1996. On a double-threshold autoregressive heteroscedastic time series model. Journal of Applied Econometrics 11, 253–274.
Nelson, D.B., 1990. Stationarity and persistence in the GARCH(1, 1) model. Econometric Theory 6, 318–334.
Pan, J.Z., Wang, H., Tong, H., 2008. Estimation and tests for power-transformed and threshold GARCH models. Journal of Econometrics 142, 352–378.
Park, J.A., Baek, J.S., Hwang, S.Y., 2009. Persistent threshold-GARCH processes: Model and application. Statistics & Probability Letters 79, 907–914.
Pawitan, Y., 2001. In all likelihood: Statistical modelling and Inference using likelihood. Oxford.

Das könnte Ihnen auch gefallen