Kocaeli Üniversitesi Sosyal Bilimler Enstitüsü Dergisi (14) 2007 / 2 : 78-109

The Use of ARCH and GARCH Models for Estimating and

Forecasting Volatility

Bahadtin Rüzgar ∗
İsmet Kale ∗∗

Abstract: This paper presents the performance of 11 ARCH-type models each with
four different distributions combined with ARMA specifications in conditional mean
in estimating and forecasting the volatility of IMKB 100 stock indices, using daily
data over a 9 years period. The results suggest that fractionally integrated asym-
metric models outperform the non-FI versions and, using skewed-t and student-t
distributions provide better fit to the data for almost every model in estimating
volatility. In forecasting volatility a clear improvement is not observed by altering a
specific model component or distribution.
Keywords: GARCH; EGARCH; GJR; APARCH; IGARCH; FIGARCH; FIAPARCH;
FIEGARCH; HYGARCH; ARMA; GED; Skewed-t; Ox; G@RCH

1. Introduction

Until the 80’s most of the analytical research was on finding the relation between
factors and outcomes. For this purpose it was a mostly used assumption for simplic-
ity that errors were random constants. The models that best represent relations were
those that produced minimum errors. As a result the errors were minimized in mod-
els, and remained out of the subject of quantitative prediction. However in some
cases it is exactly the quantity of those errors, the prediction of which is important.
The world of finance is one of the first who supported the research as they realized
that this error can be interpreted as what we may call the risk.
Generalized AutoRegressiv Conditional Heteroskedasticity (GARCH) is a model
of errors. It is mostly used in other models to represent volatility. The models that

∗
Yrd. Doç. Dr. Bahadtin Rüzgar, Marmara Üniversitesi Bankacılık ve Sigortacılık Yüksekokulu Aktüerya
Bölümü’n’de öğretim üyesidir.
∗∗
İsmet Kale, Marmara Üniversitesi Bankacılık ve Sigortacılık Yüksekokulu Bankacılık Bölümü’nde Yüksek
Lisans öğrencisidir.
 The Use Of ARCH And GARCH Models…. 79

make use of GARCH vary from predicting the spread of toxic gases in the
atmosphere to simulating neural activity. But finance is still the leading area and
dominates the research on GARCH.
ARCH class models were first introduced by Nobel price awarded Engle (1982)
with the ARCH model. Since then, numerous extensions have been put forward, all
of them modelling the conditional variance as a function of past (squared) returns
and associated characteristics.
In recent years, the tremendous growth of trading activity and trading losses of
financial institutions has led financial regulators and supervisory committee of banks
to favor quantitative techniques which appraise the possible loss that these
institutions can incur. Value-at-Risk (VaR) has become one of the most sought-after
techniques. The computation of the VaR for a collection of returns requires the
computation of the empirical quantile at level α of the distribution of the returns of
the portfolio. Because quantiles are direct functions of the variance in parametric
models, ARCH class models immediately translate into conditional VaR models.
These conditional VaR models are important for characterizing short term risk for
intradaily or daily trading positions.
In this paper we investigate the estimating and forecasting capabilities of
GARCH models when applied to daily IMKB 100 index data. We furthermore aim
to understand whether IMKB data exhibits the common caracteristics of financial
time series observed in developed countries. We thereby wish to contribute to the
risk management research in Turkey, the outcomes of which will be of crucial value
after the implementation of Basel II regulations in 2007.
The rest of the paper is organized in the following way. In Section 2, we describe
ARMA and GARCH processes as the building blocks of analysed variance models.
These models are applied to daily stock index data in Section 3 where we assess
their performances and conclude.

2. Formation of variance models

GARCH models are designed to capture certain characteristics that are commonly as-
sociated with financial time series: fat tails, volatility clustering and leverage effects.
Probability distributions for asset returns often exhibit fatter tails than the
standard normal, or Gaussian, distribution. Time series that exhibit a fat tail
distribution are often referred to as leptokurtic. In addition, financial time series
usually exhibit a characteristic known as volatility clustering, in which large changes
tend to follow large changes, and small changes tend to follow small changes. In
either case, the changes from one period to the next are typically of unpredictable
80 Bahadtin Rüzgar, İsmet Kale

sign. Large disturbances, positive or negative, become part of the information set
used to construct the variance forecast of the next period's disturbance. In this
manner, large shocks of either sign are allowed to persist, and can influence the
volatility forecasts for several periods. Volatility clustering, or persistence, suggests
a time-series model in which successive disturbances, although uncorrelated, are
nonetheless serially dependent.
Finally, certain classes of asymmetric GARCH models are also capable of captur-
ing the so-called leverage effect, in which asset returns are often observed to be
negatively correlated with changes in volatility.
A standard approach of time series analysis is to take a time series that exhibits com-
plicated behavior and try to convert it to a simpler form. Optimally, such simplification
would yield time series that were so simple that they could reasonably be modeled
as independent and identically distributed (IID). In practice, and especially in financial
applications, this is rarely possible. Stationarity is a condition similar to IID, but not
as strong. Two different forms of stationarity are defined:
i) A process is said to be strictly stationary if the unconditional joint distribution
of any segment (yt, yt+1, ..., yt+r) is identical to the unconditional joint distribution of
any other segment (yt+s, yt+s+1, ..., yt+s+r) of the same length.
ii) A process is said to be covariance stationary if the unconditional joint distribution
of any segment (yt, yt+1, ..., yt+r) has means, standard deviations and correlations that are
identical to the corresponding means, standard deviations and correlations of the uncon-
ditional joint distribution of any other segment (yt+s, yt+s+1, ..., yt+s+r) of equal length.
Correlations include autocorrelations and cross correlations.
Strict stationarity is appealing because it affords a form of homogeneity across terms
without requiring that they be independent. Covariance stationarity is the condition that
is more frequently assumed in GARCH models. It does require that all first and second
moments exist whereas strict stationarity does not. In this one respect, covariance sta-
tionarity is a stronger condition (Holton, 1996).
In this paper we are going to construct linear models combining mean and variance
equations holding either covariance or strict stationary. We will use ARMA for mean
and GARCH for variance spesification.

2.1 ARMA (R, S) And ARFIMA (R, D, S) Processes In The Conditional Mean

Box and Jenkins introduced a flexible family of time series models capable of ex-
pressing a variety of short-range serial relationships in terms of linear regression,
where the predictors are previous observations and previous residual errors.
 The Use Of ARCH And GARCH Models…. 81

One component of the Box-Jenkins framework is the autoregressive (AR) equa-

tion, where predicting variables are previous observations. An AR(r) model, where
the predictors are the previous r terms, is defined as
r
yt = ∑ φi yt −i + ε t (1)
i =1

In equation 1, the current value yt is partly based on the value at time t - i (i≤r),
and partly based on a random variable ε, typically Gaussian noise. The influence of
prior values is usually assumed to decay over time, such that φ 1> φ 2> …> φ r .
A second component in the Box-Jenkins framework is the moving average (MA)
process. In an MA process, the observation yt is dependent not on the previous val-
ues of yt, but rather on the values of the noise random variable ε. A moving average
model of order s, MA(s), is defined by
s
yt = ∑ θ j ε t − j + ε j (2)
j =1

where yt depends on the previous s errors εt-j (j≤s) and the current error εt.
Ooms and Doornik (1999) present the basic ARMA(r,s) model as

yt = φ1 yt −1 +... +φr yt −r + εt +θ1εt −1 +... +θsεt−s (3)

and the general ARMA (r,s) is in the form
r s
yt = φ0 + ∑
i =1
φi yt−i + ε t + ∑φ
j =1
j ε j−i
(4)

where r is the order of the AR(r) part, φi its parameters, s the order of the MA(s)

part, θ j its parameters and εt normally and identically distributed noise or innovation
process.
The family of ARMA models as defined by equation 4 is flexible and able to con-
cisely describe the serial dependencies of seemingly complex time series in terms of
the number of parameters (i.e., the order or history) of the AR and MA components,
and the values of these parameters.
In fields such as physics and economics, phenomena that fluctuate over time of-
ten display long-range serial correlations. In order to correctly identify and parsimo-
niously describe processes that give rise to persistent serial correlations, traditional
ARMA time series models can be extended to allow for fractional integration to cap-
ture long-range correlations. The resulting ARFIMA models, popular in economet-
rics and hydrology, allow for simultaneous maximum likelihood estimation of the
parameters of both short-range and long-range processes.
82 Bahadtin Rüzgar, İsmet Kale

Following the description of Laurent and Peters (2002), by using lag polynomi-
nals and introducing a mean μ, the equation 4 becomes
φ ( L)( yt − μ ) = θ ( L)ε t (5)
where L is the lag operator, μ is the unconditional mean of yt,
r s
φ (L) = 1 − ∑i =1
φ i L i and θ ( L ) = 1 + ∑
i =1
θ i L i are the autoregressiv and

the moving average operators in the lag operator. They are polinominals of order r
and s respectively. With a fractional integration parameter d, the ARFIMA(r, d, s)
model is written as
φ ( L )(1 − L ) d ( yt − μ ) = θ ( L )ε t (6)
The fractional differencing operator (1-L)d is a notation for the following infinite
polynominal: ∞
Γ(i − d ) ∞
(1 − L) d = ∑ Γ(i + 1)Γ(−d )
i =0
∑
Li ≡ π i (d ) Li
i =0
(7)

where π i ( z ) ≡ Γ(i − z ) / Γ(i + 1)Γ(− z ) and Γ(.) is the Standard gamma func-
tion. To ensure statitionary and invertibility of the process yt, d lies between -0.5 and
0.5. Given data series yt one can use conditional or exact likelyhood method to spec-
ify the order and parameters. The Ljung-Box statistics of residuals can check the fit.
Bhardwaj and Swanson (2004) found that ARFIMA models perform better for
greater forecast horizons and that they under certain conditions provide significantly
better out-of-sample predictions than AR, MA, ARMA, GARCH, simple regime
switching, and related models.
Throughout the paper ARMA spesification will only be used to model the mean
of returns. ARMA (0,0) implies a constant mean, ARMA(1,0) is simply AR(1). It is
also possible to make the conditional mean a function of the conditional variance. In
that case the conditional variance derived from the GARCH model will be a variable
in the mean equation. This then will be the so called ARCH-in-mean model, which
we denote in this paper with (-m) in naming our models.

2.2 GARCH Processes to Model The Conditional Variance

If the value of the stock market index at time t is marked Pt , the return of the index
at time t is given by yt = ln (Pt / Pt-1) where ln denotes natural logarithm.
For the log return series yt, we assume its mean is ARMA modelled, then let
εt = yt - μt be the mean corrected log return. Stock market index returns can be mod-
elled with the help of the following equation:
yt = μ + εt, (8)
 The Use Of ARCH And GARCH Models…. 83

where µ is the mean value of the return, which is expected to be zero; t is a ran-
dom component of the model, not autocorrelated in time, with a zero mean value.
Sequence εt may be considered a stochastic process, expressed as:
εt = zt σt (9)
where zt is a sequence of independently and identically distributed random vari-
ables, with a distribution E(zt) = 0 and Var(zt) = 1. By definition εt is serially uncor-
related with a mean equal to zero, but its conditional variance equals σt2 and there-
fore may change over time, contrary to what is assumed in the standard regression
model. The conditional variance is the measure of our uncertainty about a variable
given a model and information set.
Following Markowitz' definition of volatility as standard deviation of the ex-
pected return, σt is the volatility of log returns at time t, the changes of which will be
modelled by means of the following ARCH-type models.

2.2.1 The ARCH Model

Volatility clustering, or persistence, suggests a time-series model in which succes-

sive disturbances, although uncorrelated, are nonetheless serially dependent. Rob
Engle had the great insight to introduce and study the class of autoregressive condi-
tionally heteroscedastic (ARCH) time series models for modeling the time-varying
volatility clustering phenomenon (Engle, 1982). He used a weighted average of
squared past residuals over a long period with higher weights on the recent past and
small but non-zero weights on the distant past.
The ARCH (q) model can be expressed as
εt=ztσt (10)
zt ~ i.i.d D(0,1)
σt2 = σ2 (εt-1 , εt-2, …, t, xt, b) = σ2 (σ t-1z t-1 , σ t-2z t-2, …, t, xt, b)
where εt denotes the prediction error at time t, xt is a vector of lagged exogenous
variables, b is a vector of parameters, D(.) is distribution.
The conditional variance of εt given the information at time t-1 is σt2. For the
parameterization of this variance many possibilities are suggested in the literature. In
its original form ARCH can be written as

σ t2 = α + α ε t2−1 + ... + α qε t2−q

0 1
(11)
using ∑ operator equation 11 becomes
q
σ t2 = α + ∑ α iε t2−i
0 (12)
i =1

or by replacing εt = ztσt , to more clearly notice the autoregression, we get

84 Bahadtin Rüzgar, İsmet Kale

q
σ t2 = α + ∑ α iσ t2−i zt2−i
0 (13)
i =1

The ARCH model can describe volatility clustering through the following mecha-
nism: if εt-1 was large in absolute value, σt2 and thus εt is expected to be large in abso-
lute value as well. Even if the conditional variance of an ARCH model is time vary-
q
ing, the unconditional variance of εt is constant provided that α0 > 0 and ∑α
i =1
i < 1.

Conditional variance σt2 has to be positive for all t. Sufficient conditions are when
α0 > 0 and αi ≥ 0. Evidence has shown that a high ARCH order has to be selected to
catch the dynamics of the conditional variance. This involves the estimation of a
large number of parameters. The generalized ARCH (GARCH) model of Bollerslev
(1986) is based on an infinite ARCH spesification and it allows reducing the number
of estimated parameters by imposing nonlinear restrictions on them.

2.2.2 The GARCH Model

The GARCH model additionally assumes that forecasts of variance changing in time
also depend on the lagged conditional variances of capital assets. An unexpected in-
crease or fall in the returns of an asset at time t will generate an increase in the vari-
ability expected in the period to come.
Introduced by Engle (1982) and Bollerslev (1986) the mostly used GARCH (p,q)
models make σt2 a linear function of lagged conditional variances and squared past
residual
σ t2 = α 0 + α 1ε t2−1 + ... + α qε t2−q + β 1σ t2−1 + ... + β pσ t2− p (14)
using ∑ operator
q p
σ t = α + ∑ α iε t −i + ∑ β iσ t −i
2
0
2 2 (15)
i =1 j =1
where p is the degree of GARCH; q is the degree of the ARCH process, α0 >0, αi ≥0,
q p
βj≥0 . The covariance stationary condition is ∑α + ∑ β
i =1
i
j =1
j < 1. Since the equation

expresses the dependence of the variability of returns in the current period on data
(i.e. the values of the variables εt-i2 and σt-j2) from previous periods, we denote this
variability as conditional.
One can observe that an important feature of the GARCH (p,q) model is that it can
be regarded as an ARMA (r,s), where r is the larger of p and q. This result allows
econometricians to apply the analysis of ARMA process to the GARCH model.
Using the lag operator, the GARCH (p,q) model can be rewritten as: (ω=α0)
 The Use Of ARCH And GARCH Models…. 85

σt2 = ω + α(L) εt2 + β(L) σt2 (16)

where L denotes the lag operator and α(L) and β(L) denote the AR and MA poly-
nominals respectively, with α(L) = α1(L) + α2(L)2 + … + αq(L)q and
β(L) = β1(L) + β2(L)2 + … + βp(L)p.
If all the roots of the polynominal |1- β(L)| =0 lie outside of the unit circle, we get:
σt2 = ω | 1- β(L) |-1 + α(L) | 1- β(L) |-1 εt2, (17)
which may be regarded as an ARCH ( ∞ ) process, since the conditional variance
linearly depends on all previous squared residuals. The unconditional variance is
given by:
ω
σ 2 ≡ E ( ε t2 ) = q p
(18)
1− ∑ αi − ∑ β j
i =1 j =1
The basic and most widespread model is GARCH (1, 1), which can be reduced to:
σ t2 = ω + αε t2−1 + βσ t2−1 (19)
As the variance is expected to be positive, we expect that the regression coefficients
ω, α, β are always positive (α and β can also take the value 0), while the stationarity
of the variance is preserved, if the the sum of α and β is smaller than 1. Conditional
variability of the returns defined in equation 19 is determined by three effects:
1. The constant part, which is given by the coefficient ω;
2. Part of variance expressed by the relationship αεt–12 and designated as ARCH
component;
3. Part given by the predicted variability from the previous period and expressed
by the relationship βσt-12.
The sum of regression coefficients (α+β) expresses the influence of the variabil-
ity of variables from the previous period on the current value of the variability. This
value is usually close to 1, which is a sign of increased effects of shocks on the vari-
ability of returns on financial assets.
While the basic GARCH model allows a certain amount of leptokurtic behaviour
this is often insufficient to explain real world data. We therefore use 3 distributions
other than normal in our analysis, namely Student-t, Skewed-t (Lambert and
Laurent, 2000-2001) and Generalized Error Distributions which help to allow for the
fat tails in the distribution.
The choice of the quadratic form for the conditional variance has the important
consequence that the impact of the past values of the innovation on the current vola-
tility is only a function of their magnitude not of its sign. The principal disadvantage
of the GARCH model is therefore its unsuitability or modelling the frequently ob-
served asymmetry that occurs when a different volatility is recorded systematically
in the case of good and bad news.
86 Bahadtin Rüzgar, İsmet Kale

Falls and increases in the returns can be interpreted as good and bad news. If a
fall in returns is accompanied by an increase in volatility greater than the volatility
induced by an increase in returns, we may speak of a ‘leverage effect’. Following
classes of asymmetric GARCH models are capable of capturing this effect.

2.2.3 The GJR and TARCH model

The GJR model is an asymmetric model. It is proposed by Glosten, Jagannathan and

Runkle (1993). The generalized version may be written as
q q p
σ = α0 + ∑α ε
t
2 2
1 t −1 + ∑γ iS ε 2
t −i t −i + ∑ β jσ t2− j
(20)
i =1 i =1 j =1

where St is a dummy variable with St-i =1, if εt-i < 0 and St-i =0, if εt-i ≥ 0.
In this model, it is assumed that the impact of εt2 on the conditional variance σt2 is
different when εt is positive or negative. The TARCH model of Zakoian (1994) is
very similar to the GJR, where he preferred to model the standard deviation instead
of the conditional variance. Its basic variant is GJR (1,1), which is expressed by:
σ t2 = ω + αε t2−1 + γσ t2−1 + βε t2−1St −1 (21)
The model can be interpreted that unexpected (unforeseen) changes in the returns of
the index yt expressed in terms of εt, have different effects on the conditional vari-
ance of stock market index returns. An unforeseen increase is presented as good
news and contributes to the variance in the model through multiplicator α. An un-
foreseen fall, which is a bad news, generates an increase in volatility through multi-
plicators α and β. The asymmetric nature of the returns is then given by the nonzero
value of the coefficient β, while a positive value of β indicates a ‘leverage effect’.
q p
The covariance stationary condition is ∑ α (1 + γ
t =1
i i
2
) + ∑ β j < 1.
j =1

2.2.4 The EARCH Model

The exponential GARCH (EGARCH) model is introduced by Nelson (1991). In this

model, the conditional variance may be expressed as follows:
q p
ln σ t = α 0 + ∑ α i s( zt −1 ) + ∑ β j ln(σ t − j )
2 2 (22)
i =1 j =1
where zt = εt / σt is the normalized residuals series. The function s(.) can be written as
s(zt) = δ1 zt + δ2 { |zt| - E(|zt|) } (23)
 The Use Of ARCH And GARCH Models…. 87

Therefore δ1 zt adds the effect of the sign of εt whereas δ2 {|zt| - E (|zt|) adds its
magnitude effect. E (|zt|) depends on the choice of the distribution of return series.
For the normal distribution E (|zt|) =2 .
π
Its basic variant is EGARCH (1, 1) with normal distribution is expressed by:

ε t −1 ⎡ ε t −1 2⎤
ln σ t2 = ω + α (δ 1 + δ2 ⎢ − ⎥ ) + β ln σ t −1
2
(24)
σ t −1 ⎣ σ t −1 π ⎦
The asymmetric nature of the returns is then given by the nonzero value of the coef-
ficient δ1, while a positive value of δ1 indicates a ‘leverage effect ’.
The use of ln transformation ensures that σt2 is always positive and consequently
there are no restrictions on the sign of the parameters. Moreover external unexpected
shocks will have a stronger influence on the predicted volatility than TARCH or
GJR.

2.2.5 The APARCH Model

In general, the inclusion of a power term acts so as to emphasise the periods of rela-
tive tranquillity and volatility by magnifying the outliers in that series. Squared
terms are therefore so often used in models. If a data series is normally distributed
than we are able to completely characterise its distribution by its first two moments
(McKenzie and Mitchell, 2001). If we accept that the data may have a non-normal
error distribution, other power transformations may be more appropriate.
Recognising the possibility that a squared power term may not necessarily be op-
timal, Ding, Granger and Engle (1993) introduced a new class of ARCH model
called the Power ARCH (PARCH) model. Rather than imposing a structure on the
data, the Power ARCH class of models estimates the optimal power term.
Ding, Granger and Engle (1993) also specified a generalised asymmetric version
of the Power ARCH model (APARCH). The APARCH (p,q) model can be ex-
pressed as: q p
δ
t 0 ∑
σ =α + α (ε −γ ε ) + β σδ
i =1
i t −i
δ
i t −i ∑
j =1
j t− j
(25)

where α0 > 0, δ ≥ 0, βj ≥ 0, αi ≥ 0 and -1 < γi < 1.

This model couples the flexibility of a varying exponent δ with the asymmetry
coefficient γi to take the “leverage effect” into account. Moreover, the APARCH in-
cludes ARCH, GARCH and GJR as special cases:
ARCH when δ = 2, γi = 0 (i = 1,. . . ,p) and βj = 0 (j = 1,. . . ,p),
GARCH when δ = 2 and γi = 0 (i = 1,. . . ,p) and
88 Bahadtin Rüzgar, İsmet Kale

GJR when δ = 2
It also includes four other ARCH extentions which are not tested in this paper
TARCH when δ = 1,
NARCH when γi = 0 (i = 1,. . . ,p) and βj = 0 (j = 1,. . . ,p),
The Log-ARCH, when δ → 0 and
Taylor / Schwert GARCH when δ = 1, and γi = 0 (i = 1, . . . , p).
A stationary solution for APARCH model exists. See Ding, Granger and Engle
(1993) for details.

2.2.6 The IGARCH Model

In explaining the GARCH (p, q) model it was mentioned that GARCH may be re-
garded as an ARCH (∞) process, since the conditional variance linearly depends on
all previous squared residuals. Moreover it was stated that a GARCH process is co-
q p
variance stationary if and only if ∑ α i + ∑ β j < 1. But strict stationarity does not
i =1 j =1

require such a stringent restriction that the unconditional variance does not depend
q p
on t, in fact we often find in estimation that ∑αi + ∑ β j
i =1 j =1
is close to 1.

Lets denote h as the timelag between the present shock and future conditional
variance. Then a shock to the conditional variance σt2 has a decaying impact on σt+h2.
When h increases this impact becomes neglectable indicating a short memory.
q p
However if ∑α + ∑ β
i =1
i
j =1
j ≥ 1, the effect on σt+h2 does not die out even for a very

high h. This property is called persistence in the literature. In many high frequency
time series applications, the conditional variance estimated using GARCH (p,q)
process exhibits a strong persistence.
It was also mentioned that the GARCH (p, q) process can be seen as an ARMA
process. It is known that such an ARMA process has a unit root when
q p

∑ α i + ∑ β j = 1. When
i =1 j =1
the sum of all AR coefficients and MA coefficients is

equal to one, the ARMA process is integrated (ARIMA). Due to their similarity to
ARMA models GARCH models are symetric and have short memory.
 The Use Of ARCH And GARCH Models…. 89

q p
A GARCH model that satisfies ∑αi + ∑ β j =
i =1 j =1
1 (or equally rewritten

as α ( L) + β ( L) = 1 ) is known as an integrated GARCH (IGARCH) process

(Engle and Bollerslev, 1986), meaning that current information remains of impor-
tance when forecasting the volatility for all horizons. IGARCH (p, q) models are a
kind of ARIMA models for volatilities. Recall that GARCH model is:
define υt ≡ ε 2t − σ 2t
σ t2 = ω + α ( L)ε 2 + β ( L)σ 2(26)
ε 2t = ω + (α ( L ) + β ( L ))ε 2 + β ( L)υt + υt (27)
This is an ARMA {max (p, q), p} model for the squared innovations. If
α ( L) + β ( L) = 1 then we have an Integrated GARCH model (IGARCH).
The IGARCH can be expressed as:
φ ( L )(1 − L )ε t2 = ω + [1 − β ( L ) ]ν t or (28)

φ ( L ) (1 − L ) ε t2 = ω + [1 − β ( L ) ] ( ε t2 − σ t
2
), (29)

where θ ( L ) = [1 − α ( L ) − β ( L ) ] (1 − L ) − 1 is a polynominal of order {max (p,q)-1 }

We can also express the conditional variance as a function of the squared residuals,
and then an IGARCH (p, q) becomes:

σ t2 =
ω
[1 − β ( L ) ] {
+ 1 − φ ( L )(1 − L ) [1 − β ( L ) ]
−1
}ε t
2 (30)

Although a process yt that follows an IGARCH process is not covariance station-

ary, and its unconditional variance is infinite, IGARCH process is still important
since the unconditional density of yt is the same for all t, and thus the process can
come from a strictly stationary process. However we may suspect that IGARCH is
more a product of omitted structural breaks than the result of true IGARCH behav-
ior. An integrated process will be hereafter denoted as I(1), a non-integrared process
I(0).
The assumption of short memory such as in GARCH models is usually not ful-
filled. Ding Granger and Engle (1993) during their research for the APARCH model
have found that the absolute returns and their power transformations have a highly
significant long-term memory property as the returns are highly correlated. For ex-
ample, significant positive autocorrelations were found at over 2,700 lags in 17,054
daily observations of the S&P 500. That makes 2700 lags / 252 trading days = 10.7
years. On the other hand the implications of IGARCH models are too strong which
leads to the consideration of fractionally integrated models.
90 Bahadtin Rüzgar, İsmet Kale

2.2.7 The FIGARCH Model

As shown in Ding, Granger, and Engle (1993) among others, the effects of a shock
can take a considerable time to decay. Therefore, the distinction between I(0) and
I(1) processes seems to be far too restrictive. In an I(0) process the propagation of
shocks occurs at an exponential rate of decay so that it only captures the short-
memory, while for an I(1) process the persistence of shocks is infinite.
In the conditional mean, the ARFIMA specification has been proposed to fill the
gap between short and complete persistence, so that the short-run behavior of the
time-series is captured by the ARMA parameters, while the fractional differencing
parameter allows for modelling the long-run dependence.
The first long memory GARCH model was the fractionally integrated GARCH
(FIGARCH) introduced by Ballie, Bollerslev and Mikkelsen (1996). The FIGARCH
(p, d, q) model is a generalization of the IGARCH model by replacing the operator
(1-L) of the IGARCH equation by (1-L)d , where d is the memory parameter.
φ ( L) ⎡⎣(1 − α ) + α (1 − L) d ⎤⎦ ε t2 = ω + [1 − β ( L)]ν t (31)

where θ ( L ) = [1 − α ( L ) − β ( L ) ] (1 − L ) − 1 is a polynominal of order {max (p,q)-

1 }(same as IGARCH). We can also express the conditional variance as a function
of the squared residuals, then a FIGARCH (p,d,q) becomes:

σ t2 =
ω
[1 − β ( L ) ] {
+ 1 − φ ( L )(1 − L ) d [1 − β ( L ) ]
−1
}ε t
2 (32)

FIGARCH models exhibit long memory. They include GARCH models (for d=1)
and IGARCH models (for d=1). In contrast to ARFIMA models, where the memory
parameter d is -0.5 < d <+0.5, FIGARCH d is 0 < d < 1.
FIGARCH-processes are non-stationary like IGARCH-processes. This shows that
the concept of unit roots can hardly be generalized from linear to nonlinear proc-
esses. Furthermore, the interpretation of the memory parameter d is difficult in the
FIGARCH set up.

2.2.8 The HYGARCH Model

Davidson (2001) extended the class of FIGARCH models to HYGARCH(p,α,d,q)

models which stands for hyperbolic GARCH. HYGARCH-models replace the op-
erator (1-L)d in FIGARCH equation by [(1- α)+α(1-L)d]. The parametrization of
HYGARCH-models is given by
 The Use Of ARCH And GARCH Models…. 91

σ t2 =
ω
[1 − β ( L ) ] { { }}
+ 1 − [1 − β ( L ) ] φ ( L ) 1 + α ⎡⎣ (1 − L ) d ⎤⎦ ε t2
−1
(33)

The parameters α and d are assumed to be non-negative. HYGARCH-models nest

GARCH models (for α = 0), FIGARCH-processes (for α = 1) and IGARCH-models
(for α = d = 1).

2.2.9 FI process of Chung v.s. Ballie, Bollerslev and Mikkelsen (BBM)

Chung (1999) underscores some little drawbacks in the BBM model: there is a struc-
tural problem in the BBM specification since the direct implementation of the
ARFIMA framework originally designed for the conditional mean equation is not
perfect for the use in conditional variance equation, leading to difficult interpreta-
tions of the estimated parameters.
Indeed the fractional differencing operator applies to the constant term in the
mean equation (ARFIMA) while it does not in the variance equation (FIGARCH).
Chung (1999) proposes a slightly different process:
{ }
σ 2 = σ 2 + 1 − [1 − β ( L ) ] φ ( L ) (1 − L ) d ( ε 2 − σ 2 )
t
−1
t
(34)

or
σ t
2
=σ 2
+ λ ( L ) ( ε t2 − σ 2 ) (35)
λ (L) is an infinite summation which, in practice, has to be truncated. BBM propose
to truncate λ (L) at 1000 lags and initialize the unobserved εt2 at their unconditional
moment. Contrary to BBM, Chung (1999) proposes to truncate λ (L) at the size of
the information set (t-1) and to initialize the unobserved (εt2- σ2) at 0. In our analysis
we hold the proposal of BBM and truncate at 1000 lags.

2.2.10 The FIEGARCH and FIAPARCH Model

The idea of fractional integration has been extended to other GARCH types of mod-
els, including the Fractionally Integrated EGARCH (FIEGARCH) of Bollerslev and
Mikkelsen (1996) and the Fractionally Integrated APARCH (FIAPARCH) of Tse
(1998).
Similarly to the GARCH (p, q) process, the EGARCH (p, q) can be extended to
account for long memory by factorizing the autoregressive polynomial
[1 − β ( L ) ] = φ ( L ) (1 − L ) d where all the roots of φ (z) = 0 lie outside the unit cir-
cle. The FIEGARCH (p, d, q) is specified as follows:
ln(σ t2 ) = ω + φ ( L)−1 (1 − L)− d [1 + α ( L)] s( zt −1 ) (36)
92 Bahadtin Rüzgar, İsmet Kale

And the FIAPARCH (p, d, q) model can be written as:

{ }
σ tδ = ω + 1 − [1 − β ( L ) ] φ ( L )(1 − L ) d ( ε t − γε t ) δ
−1
(37)

3. Empirical Applications

In order to study the estimation and forecasting performances of different GARCH

processes, 11 models are applied to Istanbul Stock Exchange 100 Index log returns.
We used daily data from 3.1.1996 to 15.12.2004. From a total of 2200 trading days
2000 are used for estimation and 200 are left to test the forecasts.
The models applied are GARCH, EGARCH, GJR, APARCH, IGARCH,
FIGARCH of BBM, FIGARCH of Chung, FIEGARCH, FIAPARCH of BBM,
FIAPARCH of Chung and HYGARCH. The (p,q) = (1,1) variant of all models are
systematically tested with four different distributions, namely, Gaussian Normal, Stu-
dent-t, Generalized Error Distribution (GED) and Skewed-t distribution. That makes
44 basic models. Moreover for matters of observation 50 more models of higher order
combined with ARMA are unsystematically experimented to examine the effects on
estimation and forecasting performances with an emphasis on Maximum Likelihood.
We have written an Ox code named IMKB_Estimate&Forecast.ox on the basis
of the examples and objects provided in the Ox 3.40 and its G@RCH 3.00 package
and used it for our analysis. G@RCH 3.0 of Laurent and Peters (2002) is a package
dedicated to GARCH models and many of its extensions. It is written in the Ox pro-
gramming language (see Doornik, 1999). G@RCH 3.0 can be downloaded free of
charge for academic purposes at http://www.egss.ulg.ac.be/garch/.
The program proved to be very flexible and fast. Currently only sharewares pro-
vided for academic research are capable of analyzing such recent variety of models.
Moreover most standard software do not allow for such a flexible combination of
processes like we applied in our analysis. With open-source code are able to add or
modify specifications, processes or graphics in the future.
Estimation results are evaluated on the basis of ML, Akaike, Schwarz, Shibata
and Hannan-Quinn values, whereas forecasting results are ranked according to Min-
cer Zarnowitz regression R2, Root Mean Square Error, Mean Square Error and Mean
Absolute Error value criteria.

3.1 Estimation Results

It is apparent that t distributions shall be preferred if one aims to obtain a better rep-
resentation of the existing data. Among the first 15 best basic estimating models ac-
cording to all five criteria all was either student-t or skewed-t. GED and Normal dis-
 The Use Of ARCH And GARCH Models…. 93

tributions follow. Therefore in comparing the estimation powers of models, we re-

strict our comments to student and skewed-t distributions. Test statistics of some
models in estimating performances and forecasting performances according to dif-
ferent criteria are given Table 3.1.5 and Table 3.2.5.

Table 3.1.1: First 20 And Last Ten Models With Constant Mean İn
Estimating Performances According To Different Criteria
Akaike Schwartz Shibata Hannan-Quinn
Log-L LogL k LogL k log[log(n)] LogL log (k ) LogL ⎛ n + 2k ⎞
−2 + 2 −2 +2 −2 +2 −2 + log⎜ ⎟
n n n n n n n ⎝ n ⎠
FIAparchCh11Skt FIAparchBBM11St-t FIgarchCh11St-t FIAparchBBM11St-t FIgarchCh11St-t
5

5
FIAparchBBM11Skt FIAparchCh11Skt FIgarchBBM11St-t FIAparchCh11Skt FIgarchBBM11St-t
FIAparchBBM11St-t FIAparchBBM11Skt HYGarch11St-t FIAparchBBM11Skt FIAparchBBM11St-t
FIAparchCh11St-t FIAparchCh11St-t FIgarchCh11Skt FIAparchCh11St-t FIAparchCh11St-t
HYGarch11Skt FIgarchCh11St-t FIGarchBBM11Skt FIgarchCh11St-t HYGarch11St-t
HYGarch11St-t FIgarchBBM11St-t FIAparchBBM11St-t FIgarchBBM11St-t FIgarchCh11Skt
10

10
FIgarchCh11Skt HYGarch11St-t Igarch11St-t HYGarch11St-t FIGarchBBM11Skt
FIGarchBBM11Skt FIgarchCh11Skt FIAparchCh11St-t FIgarchCh11Skt FIAparchCh11Skt
FIgarchCh11St-t FIGarchBBM11Skt Garch11St-t FIGarchBBM11Skt FIAparchBBM11Skt
FIgarchBBM11St-t HYGarch11Skt Gjr11St-t HYGarch11Skt HYGarch11Skt
Aparch11Skt Gjr11St-t HYGarch11Skt Gjr11St-t Gjr11St-t
15

15
Gjr11Skt Gjr11Skt FIAparchCh11Skt Gjr11Skt Gjr11Skt
Aparch11St-t Aparch11St-t FIAparchBBM11Skt Aparch11St-t Garch11St-t
Gjr11St-t Aparch11Skt Igarch11Skt Aparch11Skt Aparch11St-t
FIAparchCh11GED Garch11St-t Garch11Skt Garch11St-t Igarch11St-t
FIAparchBBM11GED Garch11Skt Gjr11Skt Garch11Skt Garch11Skt
20

20

Garch11Skt FIAparchCh11GED FIGarchCh11GED FIAparchCh11GED Aparch11Skt

HYGarch11GED FIAparchBBM11GED Aparch11St-t FIAparchBBM11GED FIGarchCh11GED
Garch11St-t FIGarchCh11GED FIGarchBBM11GED FIGarchCh11GED FIGarchBBM11GED
FIGarchCh11GED FIGarchBBM11GED Aparch11Skt FIGarchBBM11GED IGarch11Skt
……………………………………………………………………………………………
FIgarchBBM11N HYGarch11N Garch11N HYGarch11N HYGarch11N
10

10

EGarch11Skt Gjr11N Igarch11N Gjr11N Garch11N

Aparch11N Garch11N Gjr11N Garch11N Gjr11N
Gjr11N Aparch11N Aparch11N Aparch11N Aparch11N
Garch11N EGarch11Skt EGarch11Skt EGarch11Skt Igarch11N
Igarch11N Igarch11N EGarch11GED Igarch11N EGarch11Skt
5

EGarch11GED EGarch11GED FIEgarch11N EGarch11GED EGarch11GED

FIEgarch11St-t FIEgarch11St-t FIEgarch11St-t FIEgarch11St-t FIEgarch11N
FIEgarch11N FIEgarch11N EGarch11N FIEgarch11N FIEgarch11St-t
EGarch11N EGarch11N FIAparchBBM11GED EGarch11N EGarch11N
Log-L = log likelihood value, n = number of observations, k = number of estimated parame-
ters
94 Bahadtin Rüzgar, İsmet Kale

For optimizing maximum likelihood, skewed-t performs better than student-t for
all models. On the other hand if we evaluate according to the other four criteria, by
which more complicated models are penalized for the inclusion of additional pa-
rameters, skewed-t looses its apparent advantage, because it requires an additional
skewness parameter. Especially Hannan-Quinn test seems to judge according to the
distribution rather than the model specification and prefer student-t.
We found that the choice of models is at least as important as the choice of
distributions, because best performing models combined with both distributions
found place in the front ranks, mostly successively. For ranking models with
constant mean in estimating performances according to different criteria, test
statistics, illustrated on Table 3.1.5, are used.

Table 3.1.2: Minimum Sum Of Rankings Of Different GARCH And

FIAPARCH Specifications
Log- L Akaike Shibata Hannan-Quinn Total
AR1Garch22Skt 2 1 1 2 6
Garch22Skt 3 2 2 1 8
ARMA22Garch-m33Skt 1 3 3 7 14
AR1Garch-m11Skt 4 4 4 6 18
AR1Garch11Skt 5 5 5 5 20
Garch11St-t 7 6 6 3 22
Garch11Skt 6 7 7 4 24
Garch11GED 8 8 8 8 32
AR1Garch11N 9 9 9 9 36
Garch11N 10 10 10 10 40
……………………………………………………………………………………
Log-L Akaike Schwarz Shibata Hannan-Quinn Total
AR1FIAparchCh11Skt 3 2 3 2 3 13
ARMA11FIAparchCh21Skt 1 1 7 1 6 16
AR1FIAparchCh21Skt 2 4 5 3 5 19
AR1FIAparchCh21St-t 4 3 4 4 4 19
FIAparchCh11Skt 5 5 2 5 2 19
FIAparchCh11St-t 6 6 1 6 1 20
FIAparchCh11GED 8 7 6 7 7 35
AR1FIAparchCh21GED 7 8 8 8 8 39
AR1FIAparchCh21N 9 9 10 9 10 47
FIAparchCh11N 10 10 9 10 9 48

Log Likelihood results are consistent with aggregate results. Normal

distribution estimates worse for all models. Higher orders alone improve results
 The Use Of ARCH And GARCH Models…. 95

more than ARMA specifications alone. Together they improve more but the
marginal benefit decreases.
Among the basic models fractionally integrated ones, especially FIAPARCH of
BBM and that of Chung combined with student-t and skewed-t distributions perform
outstanding based on all criteria. It is also to note that based on the estimation
power, the methods of BBM and Chung report only slight differences. We can con-
clude that among models with the same distribution and same mean specification a
FI (1) model is a better estimator than its FI (0) counterpart. This is a clear indicator
that IMKB 100 indices shows strong persistence and the effect of shocks influence
future returns for long periods. That FI (1) performs better then I (1) show that the
persistence is not completely permanent.

Table 3.1.3: Comparison Of APARCH And FIAPARCH Models

Model Structure
Model Mean Equation Variance Equation Distribution Log-L
ARMA11Aparch33Skt ARMA(1,1) Aparch (3,3) Skewed-t 4239.96
AR1Aparch22Skt ARMA(1,0) Aparch (2,2) Skewed-t 4238.40
Aparch22Skt ARMA(0,0) Aparch (2,2) Skewed-t 4237.11
Aparch33Skt ARMA(0,0) Aparch (3,3) Skewed-t 4236.68
AR1Aparch11Skt ARMA(1,0) Aparch (1,1) Skewed-t 4230.44
Aparch11Skt ARMA(0,0) Aparch (1,1) Skewed-t 4228.85
Aparch11St-t ARMA(0,0) Aparch (1,1) Student-t 4227.98
ARMA11Aparch-m11GED ARMA(1,1) Aparch (1,1) GED 4221.97
Aparch11GED ARMA(0,0) Aparch (1,1) GED 4218.36
Aparch11N ARMA(0,0) Aparch (1,1) Normal 4170.24
…………………………………………………………………………………………
AR1FIAparchBBM21Skt ARMA(1,0) FIAparchBBM (2,d,1) Skewed-t 4239.05
AR1FIAparchBBM11St-t ARMA(1,0) FIAparchBBM (1,d,1) Student-t 4238.10
FIAparchBBM11Skt ARMA(0,0) FIAparchBBM (1,d,1) Skewed-t 4237.78
FIAparchBMM11St-t ARMA(0,0) FIAparchBBM (1,d,1) Student-t 4236.83
AR1FIAparchBBM11GED ARMA(1,0) FIAparchBBM (1,d,1) GED 4227.41
FIAparchBMM11GED ARMA(0,0) FIAparchBBM (1,d,1) GED 4226.64
FIAparchBBM11N ARMA(0,0) FIAparchBBM (1,d,1) Normal 4181.37

A ranking among the specifications would be the existence of FI followed by the

existence of asymetry. The coefficients indicating asymmetry are mostly significant
and showing that there exist a leverage effect in IMKB 100. The simple GARCH
(1,1) with t distributions follow the more complex models with t distributions but is
clearly better than any model, even the FI(1) and asymmetric models, combined
96 Bahadtin Rüzgar, İsmet Kale

with GED or Normal distribution. No EGARCH model could reach strong

convergence during numerical optimization and results are misleading.
With distributions and mean specifications equal, the FI (1) version outperforms
the FI (0) version based on Log Likelihood.
Given the same distribution and same GARCH orders, using an autoregression
AR(1) or autoregression+moving average ARMA (1,1) or an ARCH-in-mean effect
in the mean equation improves the performance of all models if the ranking is based
on maximum likelihood. Again the other four criteria take the additional parameters
in account, however most of the time the improvement in maximum likelihood is
large enough to compensate for the estimation burden of additional parameters. The
marginal improvement decreases as the mean equation gets more complex.
The stand-alone effect of increasing the orders is significantly more than
standalone effect of manipulating the mean equation. Orders of (2, 2) or (3, 3)
perform well, but the chances of non-convergence and getting misleading results
also increases.

Table 3.1.4: Best 20 Estimating Models Based On Log Likelihood

Model Structure
Model Mean Equation Variance Equation Distribution Log-L R2
ARMA11FIAparchCh21Skt ARMA(1,1) FIAparchBBM(2,d,1)Skewed-t 4241.82 0.011570
AR1FIAparchCh21Skt ARMA (1,0) FIAparchCh (2,d,1) Skewed-t 4239.99 0.012283
ARMA11Aparch33Skt ARMA (1,1) Aparch (3,3) Skewed-t 4239.96 0.008243
ARMA11GJR33Skt ARMA (1,1) GJR (3,3) Skewed-t 4239.43 0.020429
AR1FIAparchCh11Skt ARMA (1,0) FIAparchCh (1,d,1) Skewed-t 4239.07 0.012136
AR1FIAparchBBM21Skt ARMA (1,0) FIAparchBBM(2,d,1)Skewed-t 4239.05 0.011850
AR1FIAparchCh21St-t ARMA (1,0) FIAparchCh (2,d,1) Student-t 4238.99 0.011743
AR1Gjr22Skt ARMA (1,0) GJR (2,2) Skewed-t 4238.46 0.008727
ARMA22Garch-m33Skt ARMA (2,2) ARCH-mGARCH (3,3) Skewed-t 4238.41 0.014413
AR1Aparch22Skt ARMA (1,0) Aparch (2,2) Skewed-t 4238.40 0.008754
AR1FIAparchBBM11St-t ARMA (1,0) FIAparchBBM (1,d,1) Student-t 4238.10 0.011746
FIAparchCh11Skt ARMA (0,0) FIAparchCh (1,d,1) Skewed-t 4237.81 0.012526
FIAparchBBM11Skt ARMA (0,0) FIAparchBBM(1,d,1)Skewed-t 4237.78 0.012499
Aparch22Skt ARMA (0,0) Aparch (2,2) Skewed-t 4237.11 0.008964
FIAparchBBM11St-t ARMA (0,0) FIAparchBBM(1,d,1)Student-t 4236.83 0.012130
Gjr33Skt ARMA (0,0) GJR (3,3) Skewed-t 4236.68 0.021404
Aparch33Skt ARMA (0,0) Aparch (3,3) Skewed-t 4236.68 0.021544
FIAparchCh11St-t ARMA (0,0) FIAparchCh (1,d,1) Student-t 4236.47 0.008248
AR1HYGarch22Skt ARMA (1,0) HYGarch (2,d,2) Skewed-t 4236.00 0.012553
AR1HYGarch11Skt ARMA (1,0) HYGarch (1,d,1) Skewed-t 4235.85 0.011973
 The Use Of ARCH And GARCH Models…. 97

FIAPARCH models of different specifications and distributions dominate.

Including mean specifications clearly improve Log-L results. Skewed-t distribution
seems to be the solution for fat tails.
The combination of more complex mean equation with higher orders perform
overall better with the cost of additional parameters. As a result we can conclude
that, based on the maximum likelihood, the more complex a model is the better it
fits to the data.
Combined rankings allow for the conclusion that maximum likelihood is a
consistent evaluation criterion. While Akaike, Shibata, Schwarz and Hannan-Quinn
may result in complete different rankings, the aggregate results are consistent with
that of the maximum likelihood.
If skewed-t or student-t distributions are used, it is always possible to increase the
likelihood by adding more tailored processes, implying a better fit to the data on the
basis of numbers. However using the graphs we can show that even the simplest
GARCH (1, 1) is a satisfactory model in estimation. The differences are not subtle
and using a GARCH (1, 1) model would not lead to a different decision than a
decision based on a more complex ARMA (1, 1) FIAPARCH (2, 1) model. GARCH
models in general succeed in reproducing volatility clustering, persistence, leverage
effect and fat tail behavior of real world data.

Graph 3.1.1: ARMA(1,1)FIARCHCh(2,1)Skewed-t.

98 Bahadtin Rüzgar, İsmet Kale

Best estimator in test, Log-L=4241,82. Note that the series and the residuals are
almost identical implying a good reproduction of characteristics, outliers are
perfectly cached.

Graph 3.1.2: GARCH(1,1)Normal.

One of the worse estimators in comparison, Log-L=4168.74. Reproduction of the

data was able to catch the important outliers but tends to stay closer around mean.
Residuals graph is “thicker”. Note the highest two points around data 1200
(November-December 2000). Conditional variance is higher where medium size
residuals of around ± 0.1 cluster than where single big residual of around -0.2. This
is the opposite in above graph. GARCH (1, 1) Normal has slower responses.

Table 3.1.5: Test Statistics Of Each Model

Information Criterium (minimize)
Model Log-L Akaike Schwarz Shibata Hannan-Quinn
FIAparchCh11Skt 4237.81 -4.230929 -4.20571 -4.230969 -4.221670
FIAparchBBM11Skt 4237.78 -4.230895 -4.20568 -4.230935 -4.221636
FIAparchBBM11St-t 4236.83 -4.230948 -4.20854 -4.230980 -4.222719
FIAparchCh11St-t 4236.47 -4.230588 -4.20818 -4.230620 -4.222358
HYGarch11Skt 4234.68 -4.228795 -4.20638 -4.228827 -4.220565
HYGarch11St-t 4234.04 -4.229157 -4.20955 -4.229182 -4.221956
FIgarchCh11Skt 4233.97 -4.229089 -4.20948 -4.229114 -4.221888
FIGarchBBM11Skt 4233.81 -4.228926 -4.20931 -4.228950 -4.221725
FIgarchCh11St-t 4233.39 -4.229502 -4.21269 -4.229520 -4.223330
FIgarchBBM11St-t 4233.25 -4.229369 -4.21256 -4.229387 -4.223197
Aparch11Skt 4228.85 -4.222966 -4.20055 -4.222998 -4.214736
 The Use Of ARCH And GARCH Models…. 99
Table 3.1.5 devamı
Gjr11Skt 4228.82 -4.223935 -4.20432 -4.223959 -4.216734
Aparch11St-t 4227.98 -4.223096 -4.20349 -4.223120 -4.215895
Gjr11St-t 4227.97 -4.224081 -4.20727 -4.224099 -4.217909
FIAparchCh11GED 4226.78 -4.220889 -4.19848 -4.220920 -4.212659
FIAparchBBM11GED 4226.64 -4.220747 4.19834 -4.220779 -4.212518
Garch11Skt 4225.11 -4.221221 -4.20441 -4.22124 -4.215049
HYGarch11GED 4224.61 -4.219716 -4.20011 -4.219741 -4.212515
Garch11St-t 4224.59 -4.221704 -4.20770 -4.22172 -4.216561
FIgarchCh11GED 4224.27 -4.220382 -4.20357 -4.220400 -4.214210
FIgarchBBM11GED 4224.05 -4.220160 -4.20335 -4.220178 -4.213987
Igarch11Skt 4222.01 -4.219117 -4.20511 -4.219129 -4.213973
Igarch11St-t 4221.30 -4.219411 -4.20821 -4.219419 -4.215296
Gjr11GED 4218.37 -4.214472 -4.19766 -4.214490 -4.208300
Aparch11Ged 4218.36 -4.213471 -4.19386 -4.213496 -4.206270
Garch11GED 4216.12 -4.213226 -4.19922 -4.21324 -4.208082
Igarch11GED 4212.15 -4.210254 -4.19905 -4.210262 -4.206139
EGarch11St-t 4205.20 -4.200303 -4.18069 -4.20033 -4.193102
FIEgarch11Skt 4203.91 -4.197007 -4.17179 -4.197047 -4.187749
FIEgarch11GED 4192.53 -4.186621 -4.16421 -4.186653 -4.178392
FIAparchCh11N 4182.80 -4.177891 -4.15828 -4.177916 -4.170690
FIAparchBBM11N 4181.37 -4.176462 -4.15685 -4.176486 -4.169261
FIgarchCh11N 4179.29 -4.176376 -4.16237 -4.176389 -4.171233
HYGarch11N 4179.07 -4.175160 -4.15835 -4.175178 -4.168988
FIgarchBBM11N 4178.99 -4.176076 -4.16207 -4.176088 -4.170932
EGarch11Skt 4172.21 -4.166289 -4.14388 -4.16632 -4.158059
Aparch11N 4170.24 -4.166320 -4.14951 -4.166338 -4.160148
Gjr11N 4170.24 -4.167319 -4.15331 -4.167331 -4.162175
Garch11N 4168.74 -4.166826 -4.15562 -4.16683 -4.162711
Igarch11N 4163.80 -4.162879 -4.15447 -4.162883 -4.159793
EGarch11GED 4154.92 -4.149993 -4.13038 -4.15002 -4.142792
FIEgarch11St-t 4122.17 -4.116231 -4.09382 -4.116263 -4.108002
FIEgarch11N 4120.69 -4.115746 -4.09614 -4.115770 -4.108545
EGarch11N 4096.69 -4.092739 -4.07593 -4.09276 -4.086567

3.2 Forecasting Results

As we expected, the best models for estimation are not necessarily the best ones for
forecasting. The same thing is also true for the distributions. The specification of the
model has a more clear and predictable effect on Mincer Zarnowitz regression R2.
As explained by Laurent and Peters (2002) the Mincer-Zarnowitz regression has
been largely used to evaluate forecasts in the conditional mean. For the conditional
variance, it is computed by regressing the forecasted variances on the actual
variances.
σt2 = α + βσˆt2 +υt (34)
100 Bahadtin Rüzgar, İsmet Kale

The other criteria are minimizing errors and lead to difficult interpretation and
inconsistent rankings. Like the maximum likelihood in estimation, R2 is in general
more consistent with the aggregated ranking results.

Table 3.2.1: 20 Best Forecasting Models Based On R2

Model Structure
Model Mean Equation Variance Distribution Log-L R2
Equation
Igarch11Skt ARMA (0,0) Igarch (1,1) Skewed-t 4222.01 0.021975
Igarch11N ARMA (0,0) Igarch (1,1) Normal 4163.80 0.021846
Igarch11GED ARMA (0,0) Igarch (1,1) GED 4212.15 0.021611
Aparch33Skt ARMA (0,0) Aparch (3,3) Skewed-t 4236.68 0.021544
Igarch11St -t ARMA (0,0) Igarch (1,1) Student-t 4221.30 0.021478
Gjr33Skt ARMA (0,0) GJR (3,3) Skewed-t 4236.68 0.021404
Ar1Igarch11Skt ARMA (1,0) Igarch (1,1) Skewed-t 4223.65 0.020980
ARMA11Igarch11GED ARMA (1,1) Igarch (1,1) GED 4214.04 0.020875
ARMA11Igarch11Skt ARMA (1,1) Igarch (1,1) Skewed-t 4225.31 0.020874
Aparch11Skt ARMA (0,0) Aparch (1,1) Skewed-t 4228.85 0.020827
Gjr11Skt ARMA (0,0) GJR (1,1) Skewed-t 4228.82 0.020781
ARMA11GJR33Skt ARMA(1,1) GJR (3,3) Skewed-t 4239.43 0.020429
Aparch11St-t ARMA (0,0) Aparch (1,1) Student-t 4227.98 0.020152
Gjr11St-t ARMA (0,0) GJR (1,1) Student-t 4227.97 0.020110
Aparch11N ARMA (0,0) Aparch (1,1) Normal 4170.24 0.020053
Gjr11N ARMA (0,0) GJR (1,1) Normal 4170.24 0.020052
Gjr11GED ARMA (0,0) GJR (1,1) GED 4218.37 0.019908
Aparch11Ged ARMA (0,0) Aparch (1,1) GED 4218.36 0.019904
AR1Aparch11Skt ARMA (1,0) Aparch (1,1) Skewed-t 4230.44 0.019801
Garch11Skt ARMA (0,0) GARCH (1,1) Skewed-t 4225.11 0.019484

IGARCH performance is worth noting. Riskmetrics process of J.P. Morgan is also

a kind of IGARCH. For details of the model see Mina and Xiao (2001).

Graph 3.2.1: IGARCH(1,1) Skt is one of the integrated models that proved to be a
good forecast model based on R2. MSE tells it is the third worse.
 The Use Of ARCH And GARCH Models…. 101

Graph 3.2.2: EGARCH(1,1) GED converged only weakly after 208 BFGS
iterations in 52 seconds. However it reached a record level R2 and MSE.

Graphically it seems to make a fairly good conditional variance forecast of

absolute returns. A closer look predicts that the peak points of forecasts follow
actual data with a small delay. Based on MAE and RMSE this model is the second
worse.

Graph 3.2.3: ARMA(1,1)FIAPARCH(2,d,1) Skt was the best estimator according

to Log-L, an average forecaster based on R2. RMSE ranks it to the bottom.

GED and skewed-t distributions performed well in predictions but it is not

possible to favor any distribution. While RMSE and MAE ranked the t-distributions
better, MSE favored GED and puts t distributions to the bottom. R2 made no
implications on the distribution. Normal distributions give consistently moderate
results.
102 Bahadtin Rüzgar, İsmet Kale

Table 3.2.2: First 20 And Last Ten Models With Constant Mean İn
Forecasting Performances According To Different Criteria

RMSE

MAE
Rank

Rank
MSE
R2

R2

R2

R2
EGarch11GED Garch11Skt 14 Gjr11GED 12 Garch11GED 17
5

5
Igarch11Skt FIAparchBBM11Skt 19 Aparch11Ged 13 FIEgarch11Skt 40
Igarch11N FIAparchCh11Skt 18 EGarch11GED 1 Garch11Skt 14
Igarch11GED FIGarchBBM11Skt 25 FIEgarch11N 38 FIAparchBBM11Sk 19
t
Igarch11St-t Igarch11Skt 2 Aparch11N 10 FIAparchCh11Skt 18
Aparch11Skt HYGarch11Skt 21 Gjr11N 11 FIGarchBBM11Skt 25
10

10
Gjr11Skt FIgarchCh11Skt 22 FIAparchBBM11G 20 Igarch11Skt 2
ED
Aparch11St-t Aparch11Skt 6 FIAparchCh11GED 36 HYGarch11Skt 21
Gjr11St-t Gjr11Skt 7 FIEgarch11GED 39 FIgarchCh11Skt 22
Aparch11N HYGarch11St-t 32 Garch11GED 17 FIAparchBBM11St-t 23
Gjr11N FIgarchCh11St-t 33 FIgarchCh11GED 28 FIAparchCh11St-t 37
15

15
Gjr11GED FIEgarch11Skt 40 FIAparchCh11N 35 Igarch11St-t 5
Aparch11Ged Igarch11N 3 Igarch11GED 4 Garch11St-t 16
Garch11Skt FIAparchBBM11St-t 23 FIAparchBBM11N 24 Aparch11Skt 6
Garch11N FIAparchCh11St-t 37 HYGarch11GED 26 Gjr11Skt 7
Garch11St-t Igarch11St-t 5 FI- 31 HYGarch11St-t 32
20

20
garchBBM11GED
Garch11GED Garch11St-t 16 Aparch11St-t 8 FIgarchCh11St-t 33
FIAparchCh11Skt FIgarchBBM11St-t 34 Gjr11St-t 9 FIgarchBBM11St-t 34
FIAparchBBM11Skt Gjr11GED 12 Garch11N 15 FIgarchCh11GED 28
FIAparchBBM11GED Aparch11Ged 13 FIgarchBBM11N 29 FIAparchCh11N 35
……….…………………………………………………………………………………
FIgarchBBM11GE HYGarch11GED 26 FIgarchCh11St-t 33 FIAparchCh11GED 36
10

10

HYGarch11St-t FIgarchBBM11GED 31 FIgarchBBM11St-t 34 FIEgarch11GED 39

FIgarchCh11St-t Aparch11St-t 8 FIEgarch11Skt 40 Igarch11GED 4
FIgarchBBM11St-t Gjr11St-t 9 Garch11Skt 14 FIAparchBBM11N 24
FIAparchCh11N Garch11N 15 FIAparchBBM11Sk19 HYGarch11GED 26
FIAparchCh11GED FIgarchBBM11N 29 FIAparchCh11Skt 18 FI- 31
5

garchBBM11GED
FIAparchCh11St-t HYGarch11N 30 FIGarchBBM11Skt 25 Gjr11GED 12
FIEgarch11N FIgarchCh11N 27 Igarch11Skt 2 Aparch11Ged 13
FIEgarch11GED EGarch11GED 1 HYGarch11Skt 21 EGarch11GED 1
FIEgarch11Skt FIEgarch11N 38 FIgarchCh11Skt 22 FIEgarch11N 38

While R2 ranks according to model specification, minimum error criteria seem to

give more importance to distributions. FI (1) models perform poor forecasts based
on R2 whereas other criteria do not allow for a conclusion.
 The Use Of ARCH And GARCH Models…. 103

It is also hard to draw conclusions from the model specification. Remarkable are the
forecasting performances of GJR, IGARCH and GARCH. R2 gave the worst
performances with FI (1) models while the best performers were I (1) and non-
integrated ones. Therefore we can conclude that either a complete integration or no
integration is preferred to a fractional integration. In general one obtains better R2
results the simpler a model is specified. Increased parameters through modifications in
mean or higher orders provide poor R2 results. Especially the order (2, 2) consistently
outputs very poor R2. The order (3, 3) can be either a good performer or a bad choice,
but it is worth trying. The results of FI processes of BMM and Chung are again very
similar. According to the evaluation criteria they are either among the first or among
the very last.

Table 3.2.3: Minimum Sum Of Rankings For Different IGARCH

Specifications
R2 RMSE MSE MAE TOTAL
Igarch11N 2 2 2 3 9
Igarch11GED 3 1 1 4 9
Igarch11Skt 1 4 5 1 11
Igarch11St-t 4 2 3 2 11
Igarch33Skt 10 3 4 1 18
Igarch22Skt 8 4 6 1 19
Ar1Igarch11Skt 5 5 9 3 22
ARMA11Igarch11GED 6 4 7 5 22
AR1Igarch22Skt 9 5 8 2 24
ARMA11Igarch11Skt 7 6 10 6 29

R2 rankings are consistent with aggregate rankings. In general simpler models

with less parameter perform better forecasts.

Table 3.2.4: Minimum Sum Of Rankings For Different GARCH Specifications

R2 RMSE MSE MAE TOTAL
Garch11GED 4 1 1 1 7
Garch11N 2 1 2 4 9
Garch11St-t 3 2 3 3 11
Garch11Skt 1 4 6 2 13
AR1Garch11N 6 3 4 5 18
AR1Garch11Skt 5 4 8 4 21
AR1Garch-m11Skt 7 3 5 6 21
Garch22 9 4 7 2 22
ARMA22Garch-m33Skt 8 3 6 7 24
AR1Garch22Skt 10 4 8 3 25
104 Bahadtin Rüzgar, İsmet Kale

RMSE, MSE and MAE give very different rankings in cross comparison.
However in general mean specifications restrict the flexibility of all models and
result in a general trend and can not capture the outliers.
Simple GARCH(1,1) performs generally well according to all criteria. GARCH
estimation outputs the sum of all coefficients very close to 1. This explains its
forecasting success close to IGARCH. The size of the sample is a crucial factor
affecting the forecasting performance. Therefore we believe that most models would
behave differently with different sample sizes which could be the topic of a separate
research. Test statistics for forecast evaluation measures of some models are given
in the following Table.

Table 3.2.5: Test Statistics Of Some Models İn Forecasting Performances

According To Different Criteria
Forecast Evaluation Measures
Model R² MSE(M) MSE(V) MAE(M) MAE(V) RMSE (M) RMSE (V)
ARMA22Garch-m33Skt 0,014413 0.0002767 1.705E-07 0.01307 0.000271 0.01663 0.000413
AR1Garch22Skt 0.008062 0.0002770 1.708E-07 0.01299 0.000284 0.01664 0.000413
Garch22Skt 0.008096 0.0002768 1.708E-07 0.01298 0.000283 0.01664 0.000413
AR1Garch-m11Skt 0.017135 0.0002766 1.689E-07 0.01306 0.000271 0.01663 0.000411
AR1Garch11Skt 0.018712 0.0002770 1.680E-07 0.01300 0.000271 0.01664 0.000410
Garch11Skt 0.019483 0.0002767 1.678E-07 0.01298 0.000270 0.01664 0.000410
Garch11St-t 0.018915 0.0002763 1.682E-07 0.01299 0.000270 0.01662 0.000410
EGarch11St-t 0.008337 0.0002758 1.900E-07 0.01301 0.000267 0.01661 0.000436
ARMA11Egarch22Skt 0.020393 0.0002801 2.616E-07 0.01305 0.000391 0.01674 0.000512
AR1Egarch11Skt 0.028130 0.0002779 2.713E-07 0.01299 0.000386 0.01667 0.000521
EGarch11Skt 0.028791 0.0002780 2.743E-07 0.01297 0.000388 0.01667 0.000524
ARMA11Egarch22GED 0.023115 0.0002769 2.227E-07 0.01307 0.000357 0.01664 0.000472
EGarch11GED 0.028731 0.0002757 2.290E-07 0.01304 0.000354 0.01660 0.000479
EGarch11N 0.028718 0.0002757 2.143E-07 0.01306 0.000349 0.01660 0.000463
ARMA11GJR33Skt 0.020429 0.0002777 1.690E-07 0.01304 0.000272 0.01667 0.000411
AR1Gjr22Skt 0.008727 0.0002767 1.729E-07 0.01299 0.000288 0.01663 0.000416
Gjr33Skt 0.021404 0.0002763 1.692E-07 0.01299 0.000271 0.01662 0.000411
Gjr11Skt 0.020781 0.0002764 1.689E-07 0.01299 0.000273 0.01663 0.000411
Gjr11St-t 0.020110 0.0002760 1.692E-07 0.01300 0.000273 0.01661 0.000411
AR1Gjr11GED 0.019125 0.0002759 1.688E-07 0.01302 0.000270 0.01661 0.000411
Gjr11GED 0.019908 0.0002757 1.685E-07 0.01302 0.000270 0.01661 0.000411
Gjr33N 0.018225 0.0002759 1.697E-07 0.01301 0.000271 0.01661 0.000412
ARMA11GJR22N 0.008162 0.0002776 1.706E-07 0.01309 0.000281 0.01666 0.000413
Gjr11N 0.020052 0.0002758 1.678E-07 0.01301 0.000269 0.01661 0.000410
ARMA11Aparch33Skt 0.008243 0.0002780 1.731E-07 0.01304 0.000289 0.01667 0.000416
AR1Aparch22Skt 0.008754 0.0002766 1.730E-07 0.01299 0.000288 0.01663 0.000416
Aparch22Skt 0.008964 0.0002765 1.727E-07 0.01298 0.000287 0.01663 0.000416
Aparch33Skt 0.021544 0.0002763 1.691E-07 0.01299 0.000271 0.01662 0.000411
AR1Aparch11Skt 0.019801 0.0002766 1.692E-07 0.01300 0.000274 0.01663 0.000411
 The Use Of ARCH And GARCH Models…. 105

Table 3.2.5. Devamı

Aparch11Skt 0.020827 0.0002764 1.688E-07 0.01299 0.000272 0.01663 0.000411
Aparch11St-t 0.020152 0.0002760 1.691E-07 0.01300 0.000272 0.01661 0.000411
ARMA11Aparch-m11GED 0.017656 0.0002786 1.693E-07 0.01325 0.000267 0.01669 0.000411
Aparch11Ged 0.019904 0.0002757 1.685E-07 0.01302 0.000270 0.01661 0.000411
Aparch11N 0.020053 0.0002758 1.678E-07 0.01301 0.000269 0.01661 0.000410
Igarch33Skt 0.007316 0.0002767 1.773E-07 0.01298 0.000304 0.01663 0.000421
AR1Igarch22Skt 0.008869 0.0002771 1.759E-07 0.01299 0.000303 0.01665 0.000419
Igarch22Skt 0.008889 0.0002769 1.760E-07 0.01298 0.000303 0.01664 0.000420
ARMA11Igarch11Skt 0.020874 0.0002789 1.714E-07 0.01305 0.000294 0.01670 0.000400
Ar1Igarch11Skt 0.020980 0.0002772 1.714E-07 0.01300 0.000293 0.01665 0.000414
Igarch11Skt 0.021975 0.0002769 1.710E-07 0.01298 0.000293 0.01664 0.000414
Igarch11St-t 0.021478 0.0002763 1.715E-07 0.01299 0.000293 0.01662 0.000414
ARMA11Igarch11GED 0.020875 0.0002770 1.709E-07 0.01305 0.000293 0.01664 0.000413
Igarch11GED 0.021611 0.0002759 1.706E-07 0.01301 0.000292 0.01661 0.000413
Igarch11N 0.021846 0.0002761 1.701E-07 0.01300 0.000292 0.01662 0.000412
AR1FIgarchBBM22St-t 0.012166 0.0002765 1.716E-07 0.01299 0.000303 0.01663 0.000414
FIGarchBBM11Skt 0.012040 0.0002768 1.711E-07 0.01298 0.000302 0.01664 0.000414
FIgarchBBM11St-t 0.011605 0.0002764 1.713E-07 0.01299 0.000302 0.01662 0.000414
FIgarchBBM11GED 0.011802 0.0002759 1.707E-07 0.01301 0.000301 0.01661 0.000413
FIgarchBBM11N 0.011863 0.000276 1.708E-07 0.01300 0.000302 0.01661 0.000413

4. Conclusion

Most linear time series models for prediction of returns descend from the AutoRe-
gressiv Moving Average (ARMA) and Generalized Autoregressiv Conditional Het-
eroskedastic (GARCH) models. Both concepts are useful in volatility modeling, but
less useful in return prediction.
Scientific prediction involves the spotting of past patterns or regularities and test-
ing them on recent observations. The data used to spot the patterns can therefore be
called the training data. Parametric models like GARCH make use of the training
data to modify the parameters in such a way that it fits best to the data. As a conse-
quence well structured models are able to model the data almost precisely. However
in the attempt to predict the future values with the same model one actually assumes
that the future results will follow the same characteristics, same patterns. It is also
assumed that the reactions to factors not included in the model are similar in both
the past and the future. This is the reason why GARCH models as parametric speci-
fications operate best under relatively stable market conditions. Although GARCH
is explicitly designed to model time-varying conditional variances, GARCH models
can fail to predict highly irregular phenomena, including wild market fluctuations
(e.g., crashes and subsequent rebounds), and other highly unanticipated events that
106 Bahadtin Rüzgar, İsmet Kale

can lead to significant structural change. The choice of the optimum sample size, the
window size for p and q are still highly an art based on experience.
In this study, in order to the estimation and forecasting performances of different
GARCH processes, Ox 3.40 and its G@RCH 3.0 packages dedicated to GARCH
models and many of its extensions by Laurent and Peters is used for analyzing the
models. With open-source code are able to add or modify specifications, processes
or graphics in the future.
The (p,q) = (1,1) variant of all models are systematically tested with four different
distributions, namely, Gaussian Normal, Student-t, Generalized Error Distribution
(GED) and Skewed-t distribution. 94 models of higher order are unsystematically
experimented to examine the effects on estimation and forecasting performances
with an emphasis on Maximum Likelihood.
Estimation results are evaluated on the basis of ML, Akaike, Schwarz, Shibata
and Hannan-Quinn values, whereas forecasting results are ranked according to Min-
cer Zarnowitz regression R2, Root Mean Square Error, Mean Square Error and Mean
Absolute Error value criteria. In comparing the estimation powers of models, we re-
strict our comments to student and skewed-t distributions. For optimizing maximum
likelihood, skewed-t performs better than student-t for all models. On the other hand
if we evaluate according to the other four criteria, by which more complicated mod-
els are penalized for the inclusion of additional parameters, skewed-t looses its ap-
parent advantage, because it requires an additional skewness parameter. Especially
Hannan-Quinn test seems to judge according to the distribution rather than the
model spesification and prefer student-t. We found that the choice of models is at
least as important as the choice of distributions, because best performing models
combined with both distributions found place in the front ranks, mostly succes-
sively.
Log Likelihood results are consistent with aggregate results. Normal distribution
estimates worse for all models. Higher orders alone improve results more than
ARMA specifications alone. Together they improve more but the marginal benefit
decreases. As a result we can conclude that, based on the maximum likelihood, the
more complex a model is the better it fits to the data.
Combined rankings allow for the conclusion that maximum likelihood is a
consistent evaluation criterion. While Akaike, Shibata, Schwarz and Hannan-Quinn
may result in complete different rankings, the aggregate results are consistent with
that of the maximum likelihood.
If skewed-t or student-t distributions are used, it is always possible to increase the
likelihood by adding more tailored processes, implying a better fit to the data on the
 The Use Of ARCH And GARCH Models…. 107

basis of numbers. GARCH models in general succeed in reproducing volatility

clustering, persistence, leverage effect and fat tail behavior of real world data.
For forecasting, the best models for estimation are not necessarily the best ones.
The same thing is also true for the distributions. The specification of the model has a
more clear and predictable effect on Mincer-Zarnowitz regression R2. As explained
by Laurent and Peters (2002) the Mincer-Zarnowitz regression has been largely used
to evaluate forecasts in the conditional mean. For the conditional variance, it is
computed by regressing the forecasted variances on the actual variances.
GED and skewed-t distributions performed well in predictions but it is not
possible to favor any distribution. While RMSE and MAE ranked the t-distributions
better, MSE favored GED and puts t distributions to the bottom. R2 made no
implications on the distribution. Normal distributions give consistently moderate
results. In general simpler models with less parameter perform better forecasts.
RMSE, MSE and MAE give very different rankings in cross comparison. However
in general mean specifications restrict the flexibility of all models and result in a ge-
neral trend and can not capture the outliers.
The size of the sample is a crucial factor affecting the forecasting performance.
Therefore we believe that most models would behave differently with different sam-
ple sizes which could be the topic of a separate research.

Volatilite Değerleme ve Tahmini Için ARCH ve GARCH Modellerinin Kullanımı

Özet: Bu çalışma, 9 yıllık günlük verilere dayanarak IMKB 100 endeksinin vola-
tilitesini değerlendirmek ve tahmin etmek için, her biri dört ayrı dağılımla denenen,
ARMA özellikleri eklenebilen 11 değişik ARCH modelinin performansını sunmaktadır.
Elde edilen sonuçlara göre, aynı dağılım kullanılırsa, kısmi entegre edilmiş asimetrik
modeller bu özelliğe sahip olmayan orjinal versiyonlarından daha iyi volatilite değer-
lemesi yapabilmektedir. Çarpık-t ve Student-t dağılımlarının kullanılması modelin
veriye daha uyumlu olmasını sağlamaktadır. Sonuç olarak, belirli bir model veya da-
ğılımın kullanılmasının volatilite tahmininde açık bir iyileşmeye yol açmadığı gözlen-
miştir.
Anahtar Kelimeler: GARCH; EGARCH; GJR; APARCH; IGARCH; FIGARCH;
FIAPARCH; FIEGARCH; HYGARCH; ARMA; GED; Skewed-t; Ox; G@RCH
108 Bahadtin Rüzgar, İsmet Kale

References

Baillie, R., T. Bollerslev, and H. Mikkelsen (1996): “Fractionally Integrated Generalized

Autoregressive Conditional Heteroskedasticity,” Journal of Econometrics, 74, 3–30.,
December 17, 2004. http://www.sciencedirect.com
Bhardwaj, G. and Swanson, N. R. (2004): “An Empirical Investigation of the Usefulness of
ARFIMA Models for Predicting Macroeconomic and Financial Time Series” Special issue
of the Journal of Econometrics on “Empirical Methods in Macroeconomics and Finance”.,
November 15, 2004. ftp://snde.rutgers.edu/Rutgers/wp/2004-22.pdf
Bollerslev, T. (1986): “Generalized autoregressive condtional heteroskedasticity,” Journal
of Econometrics, 31, 307–327., September 20, 2004. http://www.sciencedirect.com
Bollerslev, T., and H. O. Mikkelsen (1996): “Modeling and Pricing Long-Memory in Stock
Market Volatility,” Journal of Econometrics, 73, 151–184., December 7, 2004.
http://www.sciencedirect.com
Chung, C. F. (1999): “Estimating the fractionnally intergrated GARCH model,” National Tai-
wan University working paper, December 17, 2004.
http://gate.sinica.edu.tw/~metrics/Pdf_Papers/Figarch.pdf
Davidson, J. (2001): “Moment and Memory Properties of Linear Conditional Heteroscedas-
ticity Models,” Manuscript, Cardiff University, January 4, 2005.
http://www.ex.ac.uk/~jehd201/hygarch4.pdf
Ding, Z., C. W. J. Granger, and R. F. Engle (1993): “A Long Memory Property of Stock Mar-
ket Returns and a New Model,” Journal of Empirical Finance, 1, 83–106., September 20,
2004. http://www.sciencedirect.com
Doornik, J. A. (1999): An Object Oriented Matrix Programming Language. Timberlake Con-
sultant Ltd., third edn., Ox Documentation.
Doornik, J., and M. Ooms (1999): “A Package for Estimating, Forecasting and Simulating
ARFIMA Models: Arfima Package 1.0 for Ox,” Nuffield College (Oxford) discussion pa-
per, January 4, 2005. http://www.doornik.com/download/arfima.pdf
Engle, R. (1982): “Autoregressive conditional heteroscedasticity with estimates of the
variance of United Kingdom inflation,” Econometrica, 50, 987–1007., December 7, 2004.
http://www.sciencedirect.com
Engle, R., and T. Bollerslev (1986): “Modeling the Persistence of Conditional Variances,”
Econometric Reviews, 5, 1–50., December 17, 2004. http://www.sciencedirect.com
Glosten, L., R. Jagannathan, and D. Runkle (1993): “On the relation between expected re-
turn on stocks,” Journal of Finance, 48, 1779–1801., November 15, 2004.
http://www.sciencedirect.com
Holton, G. (1996): “Contingency Analysis” family of websites including a riskglossary, No-
vember 10, 2004 http://www.contingencyanalysis.com, www.riskglossary.com
 The Use Of ARCH And GARCH Models…. 109

Laurent, S., and J. P. Peters (2002): “A tutorial for G@RCH 2.3: a complete Ox Package
for Estimating and Forecasting ARCH Models,” G@RCH 3.0 documentation.
Lambert, P., and S. Laurent (2001): “Modelling Financial Time Series Using GARCH-Type
Models and a Skewed Student Density,” Mimeo, Universite de Liege., November 10,
2004. http://www.core.ucl.ac.be/~laurent/pdf/Lambert-Laurent.pdf
McKenzie, M. and Mitchell, H. (2001): “Generalised Asymetric Power ARCH Modelling of
Exchange Rate Volatility” Royal Melbourne Institute of Technology discussion paper,
January 4, 2005. http://mams.rmit.edu.au/mmt5alsrzfd2.pdf
Mina, J. and Xiao, J..Y. (2001): “Return to RiskMetrics: The Evolution of a Standard” Copy-
right © 2001 RiskMetrics Group. Update and restatement of the mathematical models
in the 1996 RiskMetrics Technical Document, November 2, 2004.
http://www.riskmetrics.com
Nelson, D. (1991): “Conditional heteroskedasticity in asset returns: a new approach,”
Econometrica, 59, 349–370., December 7, 2004. http://www.sciencedirect.com
Peters, J.P. (2001) “Estimating and Forecasting Volatility of Stock Indices Using Asym-
metric GARCH Models and Skewed Student-t Densities,” Working Paper, École d'Ad-
ministration des Affaires, University of Liège, Belgium, January 3, 2005.
http://www.panagora.com/2001crowell/2001cp_50.pdf
Tse, Y. (1998): “The Conditional Heteroscedasticity of the Yen-Dollar Exchange Rate,”
Journal of Applied Econometrics, 193, 49–55., November 10, 2004.
http://qed.econ.queensu.ca:80/jae/1998-v13.1/ by http://ideas.repec.org
Zakoian, J. M. (1994): “Threshold Heteroscedastic Models,” Journal of Economic Dynamics
and Control, 18, 931–955., November 10, 2004. http://papers.ssrn.com
Zemke, S. (2003): “Data Mining for Prediction. Financial Series Case” The Royal Institute
of Technology Department of Computer and Systems Sciences Doctoral Thesis, Decem-
ber 4, 2004. http://szemke.math.univ.gda.pl/zemke2003PhD.pdf