Asteriou - Series de Tiempo

Applied Econometrics
Second Edition
Dimitrios Asteriou
Associate Professor at the Department of Business Administration,
Hellenic Open University, Greece
Stephen G. Hall
Professor of Economics, University of Leicester
Part
V Time Series Econometrics
13 ARIMA Models and the Box–Jenkins Methodology 265

14 Modelling the Variance: ARCH–GARCH Models 287
15 Vector Autoregressive (VAR) Models and Causality Tests 319
16 Non-Stationarity and Unit-Root Tests 334
17 Cointegration and Error-Correction Models 355
18 Identification in Standard and Cointegrated Systems 391
19 Solving Models 402
13 ARIMA Models and the
Box–Jenkins Methodology
CHAPTER CONTENTS
An introduction to time series econometrics 266
ARIMA models 266
Stationarity 267
Autoregressive time series models 267
Moving average models 272
ARMA models 275
Integrated processes and the ARIMA models 275
Box–Jenkins model selection 276
Example: the Box–Jenkins approach 279
Questions and exercises 285
LEARNING OBJECTIVES
After studying this chapter you should be able to:
1. Understand the concept of ARIMA models.
2. Differentiate between univariate and multivariate time series models.
3. Understand the Box–Jenkins approach for model selection in the univariate time
series framework.
4. Know how to estimate ARIMA(p, d, q) models using econometric software.
265
266 Time Series Econometrics
An introduction to time series econometrics

In this section we discuss single equation estimation techniques in a different way
from Parts II and III of the text. In those parts we were explaining how to analyse the
behaviour and variability of a dependent variable by regressing it using a number of
different regressors or explanatory variables. In the time series econometrics framework,
the starting point is to exploit the information that can be obtained from a variable
that is accessible through the variable itself. An analysis of a single time series is called
a univariate time series, and this is the topic of this chapter. In general, the purpose of
time series analysis is to capture and examine the dynamics of the data. In time series
econometrics we can also have multivariate time series models, which will be discussed
in later chapters.
As has been mentioned before, traditional econometricians have emphasized the use
of economic theory and the study of contemporaneous relationships in order to explain
relationships among dependent and explanatory variables. (From here onwards we use
the term traditional econometrics to differentiate the econometric analysis examined in
Parts II and III from the new (‘modern’) developments of time series econometrics.)
Lagged variables were introduced occasionally, but not in any systematic way, or at
least not in a way that attempted to analyse the dynamics or the temporal structure of
the data. There are various aspects to time series analysis but one common theme to
them all is full use of the dynamic structure of the data; by this we mean that we extract
as much information as possible from the past history of the series. The two principal
types of time series analysis are time series forecasting and dynamic modelling. Time
series forecasting is unlike most other econometrics in that it is not concerned with
building structural models, understanding the economy or testing hypotheses. It is
only concerned with building efficient forecasting models, usually done by exploiting
the dynamic inter-relationship that exists over time for any single variable. Dynamic
modelling, on the other hand, is concerned only with understanding the structure
of the economy and testing hypotheses; however, it starts from the view that most
economic series are slow to adjust to any shock, and so to understand the process we
must fully capture the adjustment process, which may be long and complex. Since
the early 1980s, the techniques developed in the time series forecasting literature have
become increasingly useful in econometrics generally. Hence we begin this chapter
with an account of the basic ‘work horse’ of time series forecasting, the ARIMA model.
ARIMA models
Box and Jenkins (1976) first introduced ARIMA models, the term deriving from:
AR = autoregressive;
I = integrated; and
MA = moving average.
The following sections will present the different versions of ARIMA models and intro-
duce the concept of stationarity, which will be analysed extensively. After defining
stationarity, we will begin by examining the simplest model – the autoregressive model
ARIMA Models and the Box–Jenkins Methodology 267
of order one then continue with the survey of ARIMA models. Finally, the Box–Jenkins
approach for model selection and forecasting will be presented briefly.
Stationarity
A key concept underlying time series processes is that of stationarity. A time series is
covariance stationary when it has the following three characteristics:
(a) exhibits mean reversion in that it fluctuates around a constant long-run mean;
(b) has a finite variance that is time-invariant; and
(c) has a theoretical correlogram that diminishes as the lag length increases.
In its simplest terms a time series Yt is said to be stationary if:
(a) E(Yt ) = constant for all t;

(b) Var(Yt ) = constant for all t; and
(c) Cov(Yt , Yt+k ) = constant for all t and all k = 0,
or if its mean, variance and covariances remain constant over time.
Thus these quantities would remain the same whether observations for the time series
were, for example, from 1975 to 1985 or from 1985 to 1995. Stationarity is important
because, if the series is non-stationary, all the typical results of the classical regression
analysis are not valid. Regressions with non-stationary series may have no meaning and
are therefore called ‘spurious’. (The concepts of spurious regressions will be examined
and analysed further in Chapter 16.)
Shocks to a stationary time series are necessarily temporary; over time, the effects of
the shocks will dissipate and the series will revert to its long-run mean level. As such,
long-term forecasts of a stationary series will converge to the unconditional mean of
the series.
Autoregressive time series models

The AR(1) model
The simplest, purely statistical time series model is the autoregressive of order one
model, or AR(1) model
Yt = φYt−1 + ut (13.1)
where, for simplicity, we do not include a constant and |φ| < 1 and ut is a Gaussian
(white noise) error term. The assumption behind the AR(1) model is that the time series
behaviour of Yt is largely determined by its own value in the preceding period. So what
will happen in t is largely dependent on what happened in t − 1. Alternatively, what

will happen in t + 1 will be determined by the behaviour of the series in the current
time t.
Condition for stationarity

Equation (13.1) introduces the constraint |φ| < 1 in order to guarantee stationarity as
defined in the previous section. If we have |φ| > 1, then Yt will tend to get larger in each
period, so we would have an explosive series. To illustrate this, consider the following
example in EViews.
Example of stationarity in the AR(1) model

Open EViews and create a new workfile by choosing File/New Workfile. In the workfile
range choose undated or irregular and define the start observation as 1 and the
end observation as 500. To create a stationary time series process, type the following
commands in the EViews command line (the bracketed comments provide a description
of each command):
smpl 1 1 [sets the sample to be the first observation

only]
genr yt=0 [generates a new variable yt with the value
of 0]
smpl 2 500 [sets the sample to range from the second to the
five-hundredth observation]
genr yt = 0.4∗ yt(-1) +nrnd [creates yt as an AR(1) model with
φ = 0.4]
smpl 1 500 [sets the sample back to the full sample]
plot yt [provides a plot of the yt series]
The plot of the Yt series will look like that shown in Figure 13.1. It is clear that this series
has a constant mean and a constant variance, which are the first two characteristics of
a stationary series.
If we obtain the correlogram of the series we shall see that it indeed diminishes as
the lag length increases. To do this in EViews, first double-click on yt to open it in a
new window and then go to View/Correlogram and click OK.
Continuing, to create a time series (say Xt ) which has |φ| > 1, type in the
following commands:
smpl 1 1
genr xt=1
smpl 2 500
genr xt = 1.2∗ xt(−1) + nrnd
smpl 1 200
plot xt
With the final command Figure 13.2 is produced, where it can be seen that the series is
exploding. Note that we specified the sample to range from 1 to 200. This is because the
explosive behaviour is so great that EViews cannot plot all 500 data values in one graph.
–2
–4
50 100 150 200 250 300 350 400 450 500
Figure 13.1 Plot of an AR(1) model
8.E + 12
6.E + 12
4.E + 12
2.E + 12
–2.E + 12
20 40 60 80 100 120 140 160 180 200
Figure 13.2 A non-stationary, exploding AR(1) model
The AR(p) model

A generalization of the AR(1) model is the AR(p) model; the number in parentheses
denotes the order of the autoregressive process and therefore the number of lagged
dependent variables the model will have. For example, the AR(2) model will be an
autoregressive model of order two, and will have the form:
Yt = φ1 Yt−1 + φ2 Yt−2 + ut (13.2)
Similarly, the AR(p) model will be an autoregressive model of order p, and will have
p lagged terms, as in the following:
Yt = φ1 Yt−1 + φ2 Yt−2 + · · · + φp Yt−p + ut (13.3)
or, using the summation symbol:
p

Yt = φi Yt−i + ut (13.4)
i=1
Finally, using the lag operator L (which has the property Ln Yt = Yt−n ) we can write the
AR(p) model as:
Yt (1 − φ1 L − φ2 L2 − · · · − φp Lp ) = ut (13.5)
(L)Yt = ut (13.6)
where (L)Yt is a polynomial function of Yt .
Stationarity in the AR(p) model

The condition for stationarity of an AR(p) process is guaranteed only if the p roots of
the polynomial equation (z) = 0 are greater than 1 in absolute value, where z is a
real variable. (Alternatively, this can be expressed with the following terminology: the
solutions of the polynomial equation (z) = 0 should lie outside the unit circle.) To
see this, consider the AR(1) process. The condition for the AR(1) process according to
the polynomial notation reduces to:
(1 − φz) = 0 (13.7)
with its roots being greater than 1 in absolute value. If this is so, and if the first root is
equal to λ, then the condition is:

1
|λ| = > 1 (13.8)
φ
|φ| < 1 (13.9)
A necessary but not sufficient requirement for the AR(p) model to be stationary is that
the summation of the p autoregressive coefficients should be less than 1:
p

φi < 1 (13.10)
i=1
Properties of the AR models

We start by defining the unconditional mean and the variance of the AR(1) process,
which are given by:
E(Yt ) = E(Yt−1 ) = E(Yt+1 ) = 0
where Yt+1 = φYt + ut+1 . Substituting repeatedly for lagged Yt we have:

Yt+1 = φ t Y0 + φ t u1 + φ t−1 u2 + · · · + φ 0 ut+1
since |φ| < 1, φ t will be close to zero for large t. Thus we have that:
E(Yt+1 ) = 0 (13.11)
and:
σu2
Var(Yt ) = Var(φYt−1 + ut ) = φ 2 σY2 + σu2 = (13.12)
1 − φ 2 σY2
Time series are also characterized by the autocovariance and autocorrelation func-
tions. The covariance between two random variables Xt and Zt is defined as:
Cov(Xt , Zt ) = E{[Xt − E(Xt )][Zt − E(Zt )]} (13.13)
Thus for two elements of the Yt process, say Yt and Yt−1 , we have:
Cov(Yt , Yt−1 ) = E{[Yt − E(Yt )][Yt−1 − E(Yt−1 )]} (13.14)
which is called the autocovariance function. For the AR(1) model the autocovariance
function will be given by:
Cov(Yt , Yt−1 ) = E{[Yt Yt−1 ] − [Yt E(Yt−1 )] − [E(Yt )Yt−1 ]

+ [E(Yt )E(Yt−1 )]}
= E[Yt Yt−1 ]
where E(Yt ) = E(Yt−1 ) = E(Yt+1 ) = 0. This leads to:
Cov(Yt , Yt−1 ) = E[(φYt−1 + ut )Yt−1 ]

= E(φYt−1 Yt−1 ) + E(ut Yt−1 )
= φσY2 (13.15)
We can easily show that:
Cov(Yt , Yt−2 ) = E(Yt Yt−2 )

= E[(φYt−1 + ut )Yt−2 ]
= E[(φ(φYt−2 + ut−1 ) + ut )Yt−2 ]
= E(φ 2 Yt−2 Yt−2 )
= φ 2 σY2 (13.16)
and in general:
Cov(Yt , Yt−k ) = φ k σY2 (13.17)
The autocorrelation function will be given by:
Cov(Yt , Yt−k ) φ k σY2

Cor(Yt , Yt−k ) = = = φk (13.18)
Var(Yt )Var(Yt−k ) σY2
So, for an AR(1) series, the autocorrelation function (ACF) (and the graph of it which
plots the values of Cor(Yt , Yt−k ) against k and is called a correlogram) will decay
exponentially as k increases.
Finally, the partial autocorrelation function (PACF) involves plotting the estimated
coefficient Yt−k from an OLS estimate of an AR(k) process, against k. If the observations
are generated by an AR(p) process then the theoretical partial autocorrelations will be
high and significant for up to p lags and zero for lags beyond p.
Moving average models

The MA(1) model
The simplest moving average model is that of order one, or the MA(1) model, which
has the form:
Yt = ut + θ ut−1 (13.19)
Thus the implication behind the MA(1) model is that Yt depends on the value of the
immediate past error, which is known at time t.
The MA(q) model

The general form of the MA model is an MA(q) model of the form:
Yt = ut + θ1 ut−1 + θ2 ut−2 + · · · + θq ut−q (13.20)

which can be rewritten as:

q

Yt = ut + θj ut−j (13.21)
j=1
or, using the lag operator:

Yt = 1 + θ1 L + θ2 L2 + · · · + θq Lq ut (13.22)
= (L)ut (13.23)
Because any MA(q) process is, by definition, an average of q stationary white-noise

processes it follows that every moving average model is stationary, as long as q is finite.
Invertibility in MA models
A property often discussed in connection with the moving average processes is that
of invertibility. A time series Yt is invertible if it can be represented by a finite-order
MA or convergent autoregressive process. Invertibility is important because the use of
the ACF and PACF for identification assumes implicitly that the Yt sequence can be
approximated well by an autoregressive model. As an example, consider the simple
MA(1) model:
Yt = ut + θ ut−1 (13.24)
Using the lag operator, this can be rewritten as:
Yt = (1 + θ L)ut
Yt
ut = (13.25)
(1 + θ L)
If |θ | < 1, then the left-hand side of Equation (13.25) can be considered as the sum of
an infinite geometric progression:
ut = Yt (1 − θ L + θ 2 L2 − θ 3 L3 + · · · ) (13.26)
To understand this, consider the MA(1) process:
Yt = ut − θ ut−1
Lagging this relationship one period and solving for ut we have:
ut−1 = Yt−1 − θ ut−2
Substituting this into the original expression we have:
Yt = ut − θ (Yt−1 − θ ut−2 ) = ut − θ Yt−1 + θ 2 ut−2

Lagging the above expression one period and solving for ut−2 and resubstituting we get:
Yt = ut − θ Yt−1 + θ 2 Yt−2 − θ 3 ut−3
and repeating this an infinite number of times we finally get the expression Equa-
tion (13.26). Thus the MA(1) process has been inverted into an infinite order AR process
with geometrically declining weights. Note that for the MA(1) process to be invertible
it is necessary that |θ| < 1.
In general, MA(q) processes are invertible if the roots of the polynomial:
(z) = 0 (13.27)
are greater than 1 in absolute value.
Properties of the MA models

The mean of the MA process will clearly be equal to zero as it is the mean of white-noise
error terms. The variance will be (for the MA(1) model) given by:
Var(Yt ) = Var(ut + θ ut−1 ) = σu2 + θ 2 σu2 = σu2 (1 + θ 2 ) (13.28)
The autocovariance will be given by:
Cov(Yt , Yt−1 ) = E[(ut + θ ut−1 )(ut−1 + θ ut−2 )] (13.29)

2 ) + θ 2 E(u
= E(ut ut−1 ) + θ E(ut−1 t−1 ut−2 ) (13.30)
= θ σu2 (13.31)
And since ut is serially uncorrelated it is easy to see that:
Cov(Yt , Yt−k ) = 0 for k > 1 (13.32)
From this we can understand that for the MA(1) process the autocorrelation function
will be:


 θσu2 σu2
Cov(Yt , Yt−k ) = for k = 1
Cor(Yt , Yt−k ) = = σu2 (1 + θ 2 ) 1 + θ2 (13.33)
Var(Yt )Var(Yt−k ) 

0 for k > 1
So, with an MA(q) model the correlogram (the graph of the ACF) is expected to have
q spikes for k = q, and then go down to zero immediately. Also, since any MA process
can be represented as an AR process with geometrically declining coefficients, the PACF
for an MA process should decay slowly.
ARMA models
After presenting the AR(p) and the MA(q) processes, it should be clear that there can
be combinations of the two processes to give a new series of models called ARMA(p, q)
models.
The general form of the ARMA model is an ARMA(p, q) model of the form:
Yt = φ1 Yt−1 + φ2 Yt−1 + · · · + φp Yt−p + ut
+ θ1 ut−1 + θ2 ut−2 + · · · + θq ut−q (13.34)
which can be rewritten, using the summations, as:
p q

Yt = φi Yt−i + ut + θj ut−j (13.35)
i=1 j=1
or, using the lag operator:
Yt (1 − φ1 L − φ2 L2 − · · · − φp Lp ) = (1 + θ1 L + θ2 L2 + · · · + θq Lq )ut (13.36)
(L)Yt = (L)ut (13.37)
In the ARMA(p, q) models the condition for stationarity deals only with the AR(p) part
of the specification. Therefore the p roots of the polynomial equation (z) = 0 should
lie outside the unit circle. Similarly, the property of invertibility for the ARMA(p, q)
models will relate only with the MA(q) part of the specification and the roots of the
(z) polynomial should also lie outside the unit circle. The next section will deal with
integrated processes and explain the ‘I’ part of ARIMA models. Here it is useful to note
that the ARMA(p, q) model can also be denoted as an ARIMA(p,0,q) model. To give
an example, consider the ARMA(2,3) model, which is equivalent to the ARIMA(2,0,3)
model and is:
Yt = φ1 Yt−1 + φ2 Yt−2 + ut
+ θ1 ut−1 + θ2 ut−2 + θ3 ut−3 (13.38)
Integrated processes and the ARIMA models

An integrated series
ARMA models can only be made with time series Yt that are stationary. This means that
the mean, variance and covariance of the series are all constant over time. However,
most economic and financial time series show trends over time, and so the mean of Yt
during one year will be different from its mean in another year. Thus the mean of most
economic and financial time series is not constant over time, which indicates that the
series are non-stationary. To avoid this problem, and to induce stationarity, we need to
de-trend the raw data through a process called differencing. The first differences of a
series Yt are given by the equation:
Yt = Yt − Yt−1 (13.39)
As most economic and financial time series show trends to some degree, we nearly
always take the first differences of the input series. If, after first differencing, a series
is stationary, then the series is also called integrated to order one, and denoted I(1) –
which completes the abbreviation ARIMA. If the series, even after first differencing, is
not stationary, second differences need to be taken, using the equation:
Yt = 2 Yt = Yt − Yt−1 (13.40)
If the series becomes stationary after second differencing it is integrated of order two
and denoted by I(2). In general, if a series d times is differenced in order to induce
stationarity, the series is called integrated of order d and denoted by I(d). Thus the
general ARIMA model is called an ARIMA(p, d, q), with p being the number of lags of
the dependent variable (the AR terms), d being the number of differences required to
take in order to make the series stationary, and q being the number of lagged terms of
the error term (the MA terms).
Example of an ARIMA model

To give an example of an ARIMA(p, d, q) model, we can say that in general an inte-
grated series of order d must be differenced d times before it can be represented by a
stationary and invertible ARMA process. If this ARMA representation is of order (p, q),
then the original, undifferenced series is following an ARIMA(p, d, q) representation.
Alternatively, if a process Yt has an ARIMA(p, d, q) representation, then the d Yt has
an ARMA(p, q) representation, as presented by this equation:
d Yt (1 − φ1 L − φ2 L2 − · · · − φp Lp ) = (1 + θ1 L + θ2 L2 + · · · + θq Lq )ut (13.41)
Box–Jenkins model selection

A fundamental principle in the Box–Jenkins approach is parsimony. Parsimony (mean-
ing sparseness or stinginess) should come as second nature to economists and financial
analysts. Incorporating additional coefficients will necessarily increase the fit of the
regression equation (that is the value of the R2 will increase), but the cost will be a
reduction of the degrees of freedom. Box and Jenkins argue that parsimonious models
produce better forecasts than do overparametrized models. In general, Box and Jenk-
ins popularized a three-stage method aimed at selecting an appropriate (parsimonious)
ARIMA model for the purposes of estimating and forecasting a univariate time series.
The three stages are: (a) identification; (b) estimation; and (c) diagnostic checking.
These are presented below.
We have already seen that a low-order MA model is equivalent to a high-order AR
model, and similarly a low-order AR model is equivalent to a high-order MA model.
This gives rise to the main difficulty in using ARIMA models, called the identification
problem. The essence of this is that any model may be given more than one (and
in most cases many) different representations, which are essentially equivalent. How,
then, should we choose the best one and how should it be estimated? Defining the ’best’
representation is fairly easy, and here we use the principle of parsimony. This simply
means that we pick the form of the model with the smallest number of parameters to
be estimated. The trick is to find this model. You might think it is possible to start with
a high-order ARMA model and simply remove the insignificant coefficients. But this
does not work, because within this high-order model will be many equivalent ways of
representing the same model and the estimation process is unable to choose between
them. We therefore have to know the form of the model before we can estimate it. In
this context this is known as the identification problem and it represents the first stage
of the Box–Jenkins procedure.
Identification
In the identification stage (this identification should not be confused with the
identification procedure explained in the simultaneous equations chapter), the
researcher visually examines the time plot of the series ACF and PACF. Plotting
each observation of the Yt sequence against t provides useful information con-
cerning outliers, missing values and structural breaks in the data. It was men-
tioned earlier that most economic and financial time series are trended and
therefore non-stationary. Typically, non-stationary variables have a pronounced
trend (increasing or declining) or appear to meander without a constant long-run
mean or variance. Missing values and outliers can be corrected at this point. At
one time, the standard practice was to first-difference any series deemed to be
non-stationary.
A comparison of the sample ACF and PACF to those of various theoretical ARIMA
processes may suggest several plausible models. In theory, if the series is non-stationary,
the ACF of the series will not die down or show signs of decay at all. If this is the case,
the series needs to be transformed to make it stationary. As was noted above, a common
stationarity-inducing transformation is to take logarithms and then first differences of
the series.
Once stationarity has been achieved, the next step is to identify the p and q orders of
the ARIMA model. For a pure MA(q) process, the ACF will tend to show estimates that
are significantly different from zero up to lag q and then die down immediately after
the qth lag. The PACF for MA(q) will tend to die down quickly, either by an exponential
decay or by a damped sine wave.
In contrast to the MA processes, the pure AR(p) process will have an ACF that will
tend to die down quickly, either by an exponential decay or by a damped sine wave,
while the PACF will tend to show spikes (significant autocorrelations) for lags up to p
and then will die down immediately.
If neither the ACF nor the PACF show a definite cut-off, a mixed process is suggested.
In this case it is difficult, but not impossible, to identify the AR and MA orders. We
should think of the ACF and PACF of pure AR and MA processes as being superimposed
Table 13.1 ACF and PACF patterns for possible ARMA(p, q) models
Model ACF PACF
Pure white noise All autocorrelations are zero All partial autocorrelations are
zero
MA(1) Single positive spike at lag 1 Damped sine wave or exponential
decay
AR(1) Damped sine wave or exponential Single positive spike at lag 1
decay
ARMA(1,1) Decay (exp. or sine wave) Decay (exp. or sine wave)
beginning at lag 1 beginning at lag 1
ARMA(p, q) Decay (exp. or sine wave) Decay (exp. or sine wave)
beginning at lag q beginning at lag p
onto one another. For example, if both ACF and PACF show signs of slow exponen-
tial decay, an ARMA(1,1) process may be identified. Similarly, if the ACF shows three
significant spikes at lags one, two and three and then an exponential decay, and the
PACF spikes at the first lag and then shows an exponential decay, an ARMA(3,1) pro-
cess should be considered. Table 13.1 reports some possible combinations of ACF and
PACF forms that allow us the detection of the order of ARMA processes. In general, it
is difficult to identify mixed processes, so sometimes more than one ARMA(p, q) model
might be estimated, which is why the estimation and diagnostic checking stages are
both important and necessary.
Estimation
In the estimation stage, each of the tentative models is estimated and the various coef-
ficients are examined. In this second stage, the estimated models are compared using
the Akaike information criterion (AIC) and the Schwarz Bayesian criterion (SBC). We
want a parsimonious model, so we choose the model with the smallest AIC and SBC
values. Of the two criteria, the SBC is preferable. Also at this stage we have to be aware
of the common factor problem. The Box–Jenkins approach necessitates that the series
is stationary and the model invertible.
Diagnostic checking
In the diagnostic checking stage we examine the goodness of fit of the model. The
standard practice at this stage is to plot the residuals and look for outliers and evidence
of periods in which the model does not fit the data well. Care must be taken here to
avoid overfitting (the procedure of adding another coefficient in an appropriate model).
The special statistics we use here are the Box–Pierce statistic (BP) and the Ljung–Box
(LB) Q-statistic (see Ljung and Box, 1979), which serve to test for autocorrelations of
the residuals.
The Box–Jenkins approach step by step

The Box–Jenkins approach involves the following steps:
Step 1 Calculate the ACF and PACF of the raw data, and check whether the series is
stationary or not. If the series is stationary, go to step 3; if not, go to step 2.
Step 2 Take the logarithm and the first differences of the raw data and calculate the
ACF and PACF for the first logarithmic differenced series.
Step 3 Examine the graphs of the ACF and PACF and determine which models would
be good starting points.
Step 4 Estimate those models.
Step 5 For each of the estimated models:
(a) check to see if the parameter of the longest lag is significant. If not, there
are probably too many parameters and you should decrease the order of p
and/or q.
(b) check the ACF and PACF of the errors. If the model has at least enough
parameters, then all error ACFs and PACFs will be insignificant.
(c) check the AIC and SBC together with the adj-R2 of the estimated models to
detect which model is the parsimonious one (that is the one that minimizes
AIC and SBC and has the highest adj-R2 ).
Step 6 If changes in the original model are needed, go back to step 4.
Example: the Box–Jenkins approach

The Box–Jenkins approach in EViews
The file ARIMA.wf1 contains quarterly data observations for the consumer price index
(cpi) and gross domestic product (gdp) of the UK economy. We shall try to identify the
underlying ARMA model for the gdp variable.
Step 1 As a first step we need to calculate the ACF and PACF of the raw data. To do this
we need to double-click on the cpi variable to open the variable in a new EViews
window. We can then calculate the ACF and PACF and view their respective
graphs by clicking on View/Correlogram in the window that contains the gdp
variable. This will give us Figure 13.3.
From Figure 13.3 we can see that the ACF does not die down at all for all lags
(see also the plot of gdp to notice that it is clearly trended), which suggests that
the series is integrated and we need to proceed with taking logarithms and first
differences of the series.
Step 2 We take logs and then first differences of the gdp series by typing the following
commands into the EViews command line:
genr lgdp = log(gdp)
genr dlgdp = lgdp − lgdp(−1)
Date: 02/26/04 Time: 15:31

Sample: 1980:1 1998:2
Included observations: 74
Autocorrelation Partial correlation AC PAC Q-stat Prob
. |*******| . |*******| 1 0.963 0.963 71.464 0.000

. |*******| .*| . | 2 0.922 –0.079 137.85 0.000
. |*******| .|. | 3 0.878 –0.049 198.98 0.000
. |****** | .|. | 4 0.833 –0.047 254.74 0.000
. |****** | .|. | 5 0.787 –0.038 305.16 0.000
. |****** | .|. | 6 0.740 –0.021 350.47 0.000
. |***** | .|. | 7 0.695 –0.002 391.06 0.000
. |***** | .|. | 8 0.650 –0.040 427.05 0.000
. |***** | .|. | 9 0.604 –0.029 458.63 0.000
. |**** | .|. | 10 0.559 –0.026 486.05 0.000
Figure 13.3 ACF and PACF of gdp
Date: 02/26/04 Time: 15:43

Sample: 1980:1 1998:2
Autocorrelation Partial correlation AC PAC Q-stat Prob
. |*** | . |*** | 1 0.454 0.454 15.645 0.000

. |** | . |*. | 2 0.288 0.104 22.062 0.000
. |** | . |*. | 3 0.312 0.187 29.661 0.000
. |** | .|. | 4 0.242 0.037 34.303 0.000
. |*. | .|. | 5 0.130 –0.049 35.664 0.000
. |** | . |*. | 6 0.238 0.174 40.287 0.000
.|. | .*| . | 7 0.055 –0.187 40.536 0.000
.*| . | .*| . | 8 –0.085 –0.141 41.149 0.000
.|. | .|. | 9 –0.010 –0.032 41.158 0.000
.|. | .|. | 10 –0.020 –0.026 41.193 0.000
Figure 13.4 ACF and PACF of dlgdp
and then double-click on the newly created dlgdp (log-differenced series) and
click again on View/Correlogram to obtain the correlogram of the dlgdp series.
Step 3 From step 2 above we obtain the ACF and PACF of the dlgdp series, provided
in Figure 13.4. From this correlogram we can see that there are 2 to 3 spikes
on the ACF, and then all are zero, while there is also one spike in the PACF
which then dies down to zero quickly. This suggests that we might have up
to MA(3) and AR(1) specifications. So, the possible models are the ARMA(1,3),
ARMA(1,2) or ARMA(1,1) models.
Step 4 We then estimate the three possible models. The command for estimating the
ARMA(1,3) model is:
ls dlgdp c ar(1) ma(1) ma(2) ma(3)

similarly, for ARMA(1,2) it is:
ls dlgdp c ar(1) ma(1) ma(2)
and for ARMA(1,1) it is:
ls dlgdp c ar(1) ma(1)
The results are presented in Tables 13.2, 13.3 and 13.4, respectively.
Step 5 Finally, the diagnostics of the three alternative models need to be checked,
to see which model is the most appropriate. Summarized results of all three
specifications are provided in Table 13.5, from which we see that, in terms of
the significance of estimated coefficients, the model that is most appropriate
is probably ARMA(1,3). ARMA(1,2) has one insignificant term (the coefficient
of the MA(2) term, which should be dropped), but when we include both
MA(2) and MA(3), the MA(3) term is highly significant and the MA(2) term
is significant at the 90% level. In terms of AIC and SBC we have contradic-
tory results. The AIC suggests the ARMA(1,3) model, but the SBC suggests
the ARMA(1,1) model. The adj-R2 is also higher for the ARMA(1,3) model. So
evidence here suggests that the ARMA(1.3) model is probably the most appro-
priate one. Remembering that we need a parsimonious model, there might
be a problem of overfitting here. For this we also check the Q-statistics of
the correlograms of the residuals for lags 8, 16 and 24. We see that only the
ARMA(1,3) model has insignificant lags for all three cases, while the other two
Table 13.2 Regression results of an ARMA(1,3) model

Dependent variable: DLGDP
Method: least squares
Date: 02/26/04 Time: 15:50
Sample(adjusted): 1980:3 1998:2
Included observations: 72 after adjusting endpoints
Convergence achieved after 10 iterations
Backcast: 1979:4 1980:2
Variable Coefficient Std. error t-statistic Prob.
C 0.006817 0.001541 4.423742 0.0000

AR(1) 0.710190 0.100980 7.032979 0.0000
MA(1) −0.448048 0.146908 −3.049866 0.0033
MA(2) −0.220783 0.123783 −1.783625 0.0790
MA(3) 0.323663 0.113301 2.856665 0.0057
R-squared 0.340617 Mean dependent var. 0.005942

Adjusted R-squared 0.301251 S.D. dependent var. 0.006687
S.E. of regression 0.005590 Akaike info criterion −7.468887
Sum squared resid. 0.002093 Schwarz criterion −7.310785
Log likelihood 273.8799 F -statistic 8.652523
Durbin–Watson stat. 1.892645 Prob(F -statistic) 0.000011
Inverted AR Roots 0.71

Inverted MA Roots 0.55+0.44i 0.55−0.44i −0.65

Date: 02/26/04 Time: 16:00
Backcast: 1980:1 1980:2
C 0.006782 0.001387 4.890638 0.0000

AR(1) 0.722203 0.114627 6.300451 0.0000
MA(1) −0.342970 0.171047 −2.005128 0.0489
MA(2) −0.124164 0.130236 −0.953374 0.3438


Inverted MA Roots 0.56 −0.22

Date: 02/26/04 Time: 16:03
Backcast: 1980:2
C 0.006809 0.001464 4.651455 0.0000

AR(1) 0.742291 0.101186 7.335927 0.0000
MA(1) −0.471431 0.161407 −2.920758 0.0047


Inverted MA Roots 0.47
models have significant (for 90%) lags for the eighth and the 16th lag, sug-
gesting that the residuals are serially correlated. So, again, here the ARMA(1,3)
model seems to be the most appropriate. As an alternative specification, as an
exercise for the reader, go back to step 4 (as step 6 suggests) and re-estimate a
model with an AR(1) term and MA(1) and MA(3) terms, to see what happens
to the diagnostics.
Table 13.5 Summary results of alternative ARMA(p, q) models

ARMA(1,3) ARMA(1,2) ARMA(1,1)
Degrees of freedom 68 69 70
SSR 0.002093 0.002266 0.002288
φ(t -stat in parentheses) 0.71 (7.03) 0.72 (6.3) 0.74 (7.33)
θ1 (t -stat in parentheses) −0.44 (−3.04) −0.34 (−2.0) −0.47 (−2.92)
θ2 (t -stat in parentheses) −0.22 (−1.78) −0.12 (0.9) —
θ3 (t -stat in parentheses) 0.32 (2.85) — —
AIC/SBC −7.4688/−7.3107 −7.4173/−7.2908 −7.4356/−7.3407
Adj R 2 0.301 0.254 0.258
Ljung–Box statistics Q(8) = 5.65(0.22) Q(8) = 9.84(0.08) Q(8) = 11.17(0.08)
for residuals (sig Q(16) = 14.15(0.29) Q(16) = 20.66(0.08) Q(16) = 19.81(0.07)
levels in parentheses) Q(24) = 19.48(0.49) Q(24) = 24.87(0.25) Q(24) = 28.58(0.15)
The Box–Jenkins approach in Stata

The file ARIMA.dat contains quarterly data observations for the consumer price index
(cpi) and gross domestic product (gdp) of the UK economy. In this example we shall
give the commands for the identification of the best ARMA model for the gdp variable.
The analysis is the same as in the EViews example presented earlier.
Step 1 To calculate the ACF and PACF, the command in Stata is:
corrgram gdp
The results obtained are shown in Figure 13.5. Additionally, Stata calculates
the ACF and the PACF with graphs that show the 95% confidence limit. The
commands for these are:
ac gdp
pac gdp
The graphs of these commands are shown in Figures 13.6 and 13.7, respectively.
Step 2 To take logs and first differences of the gdp series the following commands
should be executed:
g lgdp = log(gdp)
g dlgdp = D.lgdp
Then again, for the correlograms, the commands are:
corrgram dlgdp
ac dlgdp
pac dlgdp
Step 3–5 We proceed with the estimation of the various possible ARMA models. The
command for estimating ARIMA(p, d, q) models in Stata is the following:
arima depvarname , arima(#p,#d,#q)

284
−1 0 1 −1 0 1
LAG AC PAC Q Prob > Q [Autocorrelation] [Partial autocor]
1 0.9584 1.0062 68.932 0.0000

2 0.9584 –0.4796 132.39 0.0000
3 0.8655 –0.0349 190.23 0.0000
4 0.8173 –0.2830 242.57 0.0000
5 0.7701 –0.0471 289.73 0.0000
6 0.7226 –0.0778 331.88 0.0000
7 0.6753 –0.0674 369.26 0.0000
8 0.6285 0.2121 402.15 0.0000
9 0.5817 0.1550 430.77 0.0000
10 0.5344 0.0570 455.31 0.0000
11 0.4904 –0.0105 476.31 0.0000
12 0.4463 0.0612 494 0.0000
13 0.4034 0.2093 508.69 0.0000
14 0.3618 –0.0505 520.72 0.0000
15 0.3210 –0.1443 530.34 0.0000
16 0.2802 0.0415 537.81 0.0000
17 0.2415 0.1475 543.46 0.0000
18 0.2061 0.0301 547.65 0.0000
19 0.1742 –0.0824 550.7 0.0000
20 0.1458 0.0461 552.88 0.0000
21 0.1182 0.0243 554.34 0.0000
22 0.0918 0.3626 555.24 0.0000
23 0.0680 0.0783 555.74 0.0000
24 0.0461 0.0034 555.98 0.0000
25 0.0258 0.1899 556.05 0.0000
26 0.0060 0.0019 556.06 0.0000
27 0.0143 0.1298 556.08 0.0000
28 0.0332 0.0009 556.22 0.0000
29 0.0502 0.1807 556.53 0.0000
30 0.0675 0.1939 557.11 0.0000
31 0.0837 0.2127 558.02 0.0000
32 0.1011 0.0757 559.38 0.0000
33 0.1197 0.1165 561.33 0.0000
34 0.1371 0.0255 563.97 0.0000
Figure 13.5 ACF and PACF for gdp
1.00
Autocorrelations of gdp
0.50
0.00
−0.50
−1.00
0 10 20 30 40
Lag
Bartlett's formula for MA(q), 95% confidence bands
Figure 13.6 ACF for gdp with 95% confidence bands

1.00
Partial autocorrelations of gdp

0.50
0.00
−0.50
0 10 20 30 40
Lag
95% Confidence bands [se = 1/sqrt(n)]
Figure 13.7 PACF for gdp with 95% confidence bands
where for #p we put the number of lagged AR terms (that is, if we want
AR(4) we simply put 4) and so on. If we want to estimate an ARMA model,
then the middle term is always defined as zero (that is for ARMA(2,3) we put
arima(2,0,3)).
Therefore, the commands for the gdp variable are:
arima dlgdp , arima(1,0,3)

The results are similar to those presented in Tables 13.2, 13.3 and 13.4,
respectively.
Questions and exercises

Questions
1 Explain the implication behind the AR and MA models by giving examples of each.
2 Define the concepts of stationarity and invertibility and state the conditions for
stationarity in the AR models and invertibility for the MA models.
3 Define and explain the concepts of stationarity and invertibility. Why are they
important in the analysis of time series data? Present examples of stationary and
non-stationary, invertible and non-invertible processes.
4 Discuss analytically the three stages involved in the Box–Jenkins process for ARIMA
model selection.
Exercise 13.1
Show that an MA(1) process can be expressed as an infinite AR process.
Exercise 13.2
The file ARIMA.wf1 contains quarterly data for the consumer price index (cpi) and gross
domestic product (gdp) of the UK economy. Follow the steps described in the example
for the Box–Jenkins approach regarding gdp for the cpi variable.
14 Modelling the Variance:
ARCH–GARCH Models
CHAPTER CONTENTS
Introduction 288
The ARCH model 289
The GARCH model 299
Alternative specifications 301
Empirical illustrations of ARCH/GARCH models 313
Questions and exercises 317
LEARNING OBJECTIVES
After studying this chapter you should be able to:
1. Understand the concept of conditional variance.
2. Detect ‘calm’ and ‘wild’ periods in a stationary time series.
3. Understand the autoregressive conditional heteroskedasticity (ARCH) model.
4. Perform a test for ARCH effects.
5. Estimate an ARCH model.
6. Understand the GARCH model and the difference between the GARCH and ARCH
specifications.
7. Understand the distinctive features of the ARCH-M and GARCH-M models.
8. Understand the distinctive features of the TGARCH and EGARCH models.
9. Estimate all ARCH-type models using appropriate econometric software.
287
Introduction
Recent developments in financial econometrics have led to the use of models and tech-
niques that can model the attitude of investors not only towards expected returns but
also towards risk (or uncertainty). These require models that are capable of dealing
with the volatility (variance) of the series. Typical are the autoregressive conditional
heteroskedasticity (ARCH) family of models, which are presented and analysed in
this chapter.
Conventional econometric analysis views the variance of the disturbance terms as
being constant over time (the homoskedasticity assumption that was analysed in Chap-
ter 7). However, often financial and economic time series exhibit periods of unusually
high volatility followed by more tranquil periods of low volatility (‘wild’ and ‘calm’
periods, as some financial analysts like to call them).
Even from a quick look at financial data (see, for example, Figure 14.1, which plots the
daily returns of the FTSE-100 index from 1 January 1990 to 31 December 1999) we
can see that there are certain periods that have a higher volatility (and are therefore
riskier) than others. This means that the expected value of the magnitude of the dis-
turbance terms may be greater at certain periods compared with others. In addition,
these riskier times seem to be concentrated and followed by periods of lower risk (lower
volatility) that again are concentrated. In other words, we observe that large changes in
stock returns seem to be followed by further large changes. This phenomenon is what
financial analysts call volatility clustering. In terms of the graph in Figure 14.1, it is
clear that there are subperiods of higher volatility; it is also clear that after 1997 the
volatility of the series is much higher than it used to be.
Therefore, in such cases, it is clear that the assumption of homoskedasticity (or con-
stant variance) is very limiting, and in such instances it is preferable to examine patterns
that allow the variance to depend on its history. Or, to use more appropriate terminol-
ogy, it is preferable to examine not the unconditional variance (which is the long-run
forecast of the variance and can be still treated as constant) but the conditional variance,
based on our best model of the variable under consideration.
0.06
0.04
0.02
0.00
– 0.02
– 0.04
– 0.06
1/01/90 12/02/91 11/01/93 10/02/95 9/01/97 8/02/99
R_FTSE
Figure 14.1 Plot of the returns of FTSE-100, 1 January 1990 to 31 December 1999
Modelling the Variance: ARCH–GARCH Models 289
To understand this better, consider an investor who is planning to buy an asset at

time t and sell it at time t + 1. For this investor, the forecast of the rate of return on
this asset alone will not be enough; she/he would be interested in what the variance
of the return over the holding period would be. Therefore, the unconditional variance
is of no use either; the investor will want to examine the behaviour of the condi-
tional variance of the series to estimate the riskiness of the asset at a certain period
of time.
This chapter will focus on the modelling of the behaviour of conditional vari-
ance, or more appropriately, of conditional heteroskedasticity (from which comes
the CH part of the ARCH models). The next section presents the first model
that proposed the concept of ARCH, developed by Robert F. Engle in his sem-
inal paper ‘Autoregressive Conditional Heteroskedasticity with Estimates of the
Variance of United Kingdom Inflation’ published in Econometrica in 1982, and
which began a whole new era in applied econometrics with many ARCH varia-
tions, extensions and applications. We shall then present the generalized ARCH
(GARCH) model, followed by an alternative specification. Finally, illustrations of
ARCH/GARCH models are presented using examples from financial and economic
time series.
The ARCH model

Engle’s model suggests that the variance of the residuals at time t depends on the
squared error terms from past periods. Engle simply suggested that it is better to model
simultaneously the mean and the variance of a series when it is suspected that the
conditional variance is not constant.
Let’s examine this in a more detailed way. Consider the simple model:
Yt = a + β Xt + ut (14.1)
where Xt is a k × 1 vector of explanatory variables and β is a k × 1 vector of coefficients.

Normally, we assume that ut is independently distributed with a zero mean and a
constant variance σ 2 , or, in mathematical notation:
ut ∼ iid N(0, σ 2 ) (14.2)
Engle’s idea begins by allowing the variance of the residuals (σ 2 ) to depend on history,
or to have heteroskedasticity because the variance will change over time. One way to
allow for this is to have the variance depend on one lagged period of the squared error
terms, as follows:
σt2 = γ0 + γ1 ut−1
2 (14.3)
which is the basic ARCH(1) process.

The ARCH(1) model

Following on, the ARCH(1) model will simultaneously model the mean and the
variance of the series with the following specification:
Yt = a + β Xt + ut (14.4)
ut |t ∼ iid N(0, ht )
2
ht = γ0 + γ1 ut−1 (14.5)
where t is the information set. Here Equation (14.4) is called the mean equation
and Equation (14.5) the variance equation. Note that we have changed the nota-
tion of the variance from σt2 to ht . This is to keep the same notation from now
on, throughout this chapter. (The reason it is better to use ht rather than σt2 will
become clear through the more mathematical explanation provided later in the
chapter.)
The ARCH(1) model says that when a big shock happens in period t − 1, it is
more likely that the value of ut (in absolute terms because of the squares) will
2
also be bigger. That is, when ut−1 is large/small, the variance of the next inno-
vation ut is also large/small. The estimated coefficient of γ1 has to be positive for
positive variance.
The ARCH(q) model

In fact, the conditional variance can depend not just on one lagged realization but on
more than one, for each case producing a different ARCH process. For example, the
ARCH(2) process will be:
2
ht = γ0 + γ1 ut−1 2
+ γ2 ut−2 (14.6)
the ARCH(3) will be given by:
2
ht = γ0 + γ1 ut−1 2
+ γ2 ut−2 2
+ γ3 ut−3 (14.7)
and in general the ARCH(q) process will be given by:
2
ht = γ0 + γ1 ut−1 2
+ γ2 ut−2 2
+ · · · + γq ut−q
q

= γ0 + 2
γj ut−j (14.8)
j=1
Therefore, the ARCH(q) model will simultaneously examine the mean and the variance
of a series according to the following specification:
Yt = a + β Xt + ut (14.9)
q

ht = γ0 + 2
γj ut−j (14.10)
j=1
Again, the estimated coefficients of the γ s have to be positive for positive variance.
Testing for ARCH effects

Before estimating ARCH(q) models it is important to check for the possible presence
of ARCH effects in order to know which models require the ARCH estimation method
instead of OLS. Testing for ARCH effects was examined extensively in Chapter 7, but a
short version of the test for qth order autoregressive heteroskedasticity is also provided
here. The test can be done along the lines of the Breusch–Pagan test, which entails
estimation of the mean equation:
Yt = a + β Xt + ut (14.11)
by OLS as usual (note that the mean equation can also have, as explanatory variables in
the xt vector, autoregressive terms of the dependent variable), to obtain the residuals
ût , and then run an auxiliary regression of the squared residuals (ût2 ) on the lagged
2 , . . . , û2 ) and a constant as in:
squared terms (ût−1 t−q
ût2 = γ0 + γ1 ût−1
2 2 +w
+ · · · + γq ût−q t (14.12)
and then compute R2 × T . Under the null hypothesis of homoskedasticity (0 = γ1 =

· · · = γq ) the resulting test statistic follows a χ 2 distribution with q degrees of freedom.
Rejection of the null suggests evidence of ARCH(q) effects.
Estimation of ARCH models by iteration

The presence of ARCH effects in a regression model does not invalidate completely
the use of OLS estimation: the coefficients will still be consistent estimates, but
they will not be fully efficient and the estimate of the covariance matrix of the
parameters will be biased, leading to invalid t-statistics. A fully efficient estima-
tor with a valid covariance matrix can, however, be calculated by setting up a
model that explicitly recognizes the presence of ARCH effects. This model can no
longer be estimated using a simple technique such as OLS, which has an analyti-
cal solution, but instead a non-linear maximization problem must be solved, which
requires an iterative computer algorithm to search for the solution. The method used
to estimate ARCH models is a special case of a general estimation strategy known as

the maximum-likelihood approach. A formal exposition of this approach is beyond
the scope of this book (see Cuthbertson et al., 1992), but an intuitive account of how
this is done is given here. Approaching the task, we assume we have the correct model
and know the distribution of the error process; we select a set of values for the param-
eters to be estimated and can then in principle calculate the probability that the set
of endogenous variables we have noted in our dataset would actually occur. We then
select a set of parameters for our model that maximize this probability. These parameters
are then called the maximum-likelihood parameters and they have the general prop-
erty of being consistent and efficient (under the full set of CLRM assumptions, OLS
is a maximum-likelihood estimator). Except in certain rare cases, finding the parame-
ters which maximize this likelihood function requires the computer to search over the
parameter space, and hence the computer will perform a number of steps (or iterations)
as it searches for the best set of parameters. Packages such as EViews or Microfit include
routines that do this very efficiently, though if the problem becomes too complex the
program may sometimes fail to find a true maximum, and there are switches within the
software to help convergence by adjusting a range of options. The next section explains
step by step how to use EViews to estimate ARCH models, and provides a range of
examples.
Estimating ARCH models in EViews

The file ARCH.wf1 contains daily data for the logarithmic returns FTSE-100 (named
r_ftse) and three more stocks of the UK stock market (named r_stock1, r_stock2 and
r_stock3, respectively). We first consider the behaviour of r_ftse alone, by checking
whether the series is characterized by ARCH effects. From the time plot of the series in
Figure 14.1, it can be seen clearly that there are periods of greater and lesser volatility
in the sample, so the possibility of ARCH effects is quite high.
The first step in the analysis is to estimate an AR(1) model (having this as the mean
equation for simplicity) for r_ftse using simple OLS. To do this, click Quick/Estimate
Equation, to open the Equation Specification window. In this window we need to
specify the equation to be estimated (by typing it in the white box of the Equation
Specification window). The equation for an AR(1) model will be:
r_ftse c r_ftse(−1)
Next click OK to obtain the results shown in Table 14.1.

These results are of no interest in themselves. What we want to know is whether
there are ARCH effects in the residuals of this model. To test for such effects we use
the Breusch–Pagan ARCH test. In EViews, from the equation results window click on
View/Residuals Tests/ARCH-LM Test. EViews asks for the number of lagged terms to
include, which is simply the q term in the ARCH(q) process. To test for an ARCH(1)
process, type 1, and for higher orders the value of q. Testing for ARCH(1) (by typing 1
and pressing OK), we get the results shown in Table 14.2.
The T ∗ R2 statistic (or Obs*R-squared, as EViews presents it) is 46.05 and has a
probability value of 0.000. This clearly suggests that we reject the null hypothesis of
Table 14.1 A simple AR(1) model for the FTSE-100

Dependent variable: R_FTSE
Date: 12/26/03 Time: 15:16
Sample: 1/01/1990 12/31/1999
C 0.000363 0.000184 1.975016 0.0484

R_FTSE(−1) 0.070612 0.019538 3.614090 0.0003
R-squared 0.004983 Mean dependent var 0.000391

Adjusted R-squared 0.004602 S.D. dependent var 0.009398
Sum squared resid 0.229287 Schwarz criterion −6.495981
Durbin–Watson stat 1.993272 Prob(F -statistic) 0.000307
Table 14.2 Testing for ARCH(1) effects in the FTSE-100

ARCH test :
F -statistic 46.84671 Probability 0.000000

Obs*R-squared 46.05506 Probability 0.000000
Test equation:
Dependent variable: RESID∧ 2
Date: 12/26/03 Time: 15:27
Sample(adjusted): 1/02/1990 12/31/1999
C 7.62E−05 3.76E−06 20.27023 0.0000

RESID∧ 2(−1) 0.132858 0.019411 6.844466 0.0000
R-squared 0.017652 Mean dependent var 8.79E − 05

Sum squared resid 7.64E−05 Schwarz criterion −14.50260
homoskedasticity, and conclude that ARCH(1) effects are present. Testing for higher-
order ARCH effects (for example order 6) the results appear as shown in Table 14.3.
This time the T ∗ R2 statistic is even higher (205.24), suggesting a massive rejection
of the null hypothesis. Observe also that the lagged squared residuals are all highly
statistically significant. It is therefore clear for this equation specification that an ARCH
model will provide better results.
To estimate an ARCH model, click on Estimate in the equation results window to
go back to the Equation Specification window (or in a new workfile, by clicking on
Quick/Estimate Equation to open the Equation Specification window) and this time
change the estimation method by clicking on the down arrow in the method setting
Table 14.3 Testing for ARCH(6) effects in the FTSE-100

ARCH test:
F -statistic 37.03529 Probability 0.000000

Obs*R-squared 205.2486 Probability 0.000000
Test equation:
Dependent variable: RESID∧ 2
Date: 12/26/03 Time: 15:31
Sample(adjusted): 1/09/1990 12/31/1999
C 4.30E−05 4.46E−06 9.633006 0.0000

RESID∧ 2(−1) 0.066499 0.019551 3.401305 0.0007
RESID∧ 2(−2) 0.125443 0.019538 6.420328 0.0000
RESID∧ 2(−3) 0.097259 0.019657 4.947847 0.0000
RESID∧ 2(−4) 0.060954 0.019658 3.100789 0.0020
RESID∧ 2(−5) 0.074990 0.019539 3.837926 0.0001
RESID∧ 2(−6) 0.085838 0.019551 4.390579 0.0000
R-squared 0.078821 Mean dependent var 8.79E − 05

Sum squared resid 7.16E−05 Schwarz criterion −14.55004
and choosing the ARCH-Autoregressive Conditional Heteroskedasticity option. In

this new window, the upper part is devoted to the mean equation specification and the
lower part to the ARCH specification, or the variance equation specification. In this
window some things will appear that are unfamiliar, but they will become clear after
the rest of this chapter has been worked through. To estimate a simple ARCH(1) model,
assuming that the mean equation, as before, follows an AR(1) process, type in the mean
equation specification:
r_ftse c rftse(−1)
making sure that the ARCH-M part selects None, which is the default EViews case. For
the ARCH specification choose GARCH/TARCH from the drop-down Model: menu,
which is again the default EViews case, and in the small boxes type 1 for the Order
ARCH and 0 for the GARCH. The Threshold Order should remain at zero (which is
the default setting). By clicking OK the results shown in Table 14.4 will appear.
Note that it took ten iterations to reach convergence in estimating this model. The
model can be written as:
Yt = 0.0004 + 0.0751Yt−1 + ut (14.13)

(2.25) (3.91)
2
ht = 0.000007 + 0.1613ut−1 (14.14)
(35.97) (7.97)
Table 14.4 An ARCH(1) model for the FTSE-100

Method: ML–ARCH
Date: 12/26/03 Time: 15:34
Sample: 1/01/1990 12/31/1999
Coefficient Std. error z-statistic Prob.
C 0.000401 0.000178 2.257832 0.0240

R_FTSE(−1) 0.075192 0.019208 3.914538 0.0001
Variance equation
C 7.39E−05 2.11E−06 35.07178 0.0000

ARCH(1) 0.161312 0.020232 7.973288 0.0000

with values of z-statistics in parentheses. Note that the estimate of γ1 is highly signifi-
cant and positive, which is consistent with the finding from the ARCH test above. The
estimates of a and β from the simple OLS model have changed slightly and become
more significant.
To estimate a higher-order ARCH model, such as the ARCH(6) examined above, again
click on Estimate and this time change the Order ARCH to 6 (by typing 6 in the small
box) leaving 0 for the GARCH. The results for this model are presented in Table 14.5.
Again, all the γ s are statistically significant and positive, which is consistent
with the findings above. After estimating ARCH models in EViews you can view
the conditional standard deviation or the conditional variance series by clicking on
the estimation window View/Garch Graphs/Conditional SD Graph or View/Garch
Graphs/Conditional Variance Graph, respectively. The conditional standard devia-
tion graph for the ARCH(6) model is shown in Figure 14.2.
You can also obtain the variance series from EViews by clicking on Procs/Make
GARCH Variance Series. EViews automatically gives names such as GARCH01,
GARCH02 and so on for each of the series. We renamed our obtained variance series
as ARCH1 for the ARCH(1) series model and ARCH6 for the ARCH(6) model. A plot of
these two series together is presented in Figure 14.3.
From this graph we can see that the ARCH(6) model provides a conditional vari-
ance series that is much smoother than that obtained from the ARCH(1) model.
This will be discussed more fully later. To obtain the conditional standard deviation
series plotted above, take the square root of the conditional variance series with the
following command:
genr sd_arch1=arch1ˆ(1/2) [for the series of the ARCH(1) model]

genr sd_arch6=arch6ˆ(1/2) [for the series of the ARCH(6) model]
Table 14.5 An ARCH(6) model for the FTSE-100

Method: ML–ARCH
Date: 12/26/03 Time: 15:34
Sample: 1/01/1990 12/31/1999
C 0.000399 0.000162 2.455417 0.0141

R_FTSE(−1) 0.069691 0.019756 3.527551 0.0004
Variance equation
C 3.52E−05 2.58E−06 13.64890 0.0000

ARCH(1) 0.080571 0.014874 5.416946 0.0000
ARCH(2) 0.131245 0.024882 5.274708 0.0000
ARCH(3) 0.107555 0.022741 4.729525 0.0000
ARCH(4) 0.081088 0.022652 3.579805 0.0003
ARCH(5) 0.089852 0.022991 3.908142 0.0001
ARCH(6) 0.123537 0.023890 5.171034 0.0000

A plot of the conditional standard deviation series for both models is presented in
Figure 14.4.
A more mathematical approach

Consider the simple stationary model of the conditional mean of a series Yt :
Yt = a + β Xt + ut (14.15)
It is usual to treat the variance of the error term Var(ut ) = σ 2 as a constant, but the vari-
ance can be allowed to change over time. To explain this more fully, let us decompose
the ut term into a systematic component and a random component, as:

ut = zt ht (14.16)
where zt follows a standard normal distribution with zero mean and variance one, and
ht is a scaling factor.
In the basic ARCH(1) model we assume that:
2
ht = γ0 + γ1 ut−1 (14.17)
0.025
0.020
0.015
0.010
0.005
1/01/90 11/01/93 9/01/97
Figure 14.2 Conditional standard deviation graph for an ARCH(6) model of the FTSE-100
0.0006
0.0005
0.0004
0.0003
0.0002
0.0001
0.0000
1/01/90 11/01/93 9/01/97
ARCH1 ARCH6
Figure 14.3 Plot of the conditional variance series

0.025
0.020
0.015
0.010
0.005
1/01/90 11/01/93 9/01/97
SD_ARCH1 SD_ARCH6
Figure 14.4 Plot of the conditional standard deviation series
The process for yt is now given by:

yt = a + β xt + zt γ0 + γ1 ut−1
2 (14.18)
and from this expression it is easy to see that the mean of the residuals will be zero
(E(ut ) = 0), because E(zt ) = 0. Additionally, the unconditional (long-run) variance of
the residuals is given by:
γ0
Var(ut ) = E zt2 E(ht ) = (14.19)
1 − γ1
which means that we simply need to impose the constraints γ0 > 0 and 0 < γ1 < 1 to
obtain stationarity.
The intuition behind the ARCH(1) model is that the conditional (short-run) variance
(or volatility) of the series is a function of the immediate past values of the squared
error term. Therefore the effect of each new shock zt depends on the size of the shock
in one lagged period.
An easy way to extend the ARCH(1) process is to add additional, higher-order
lagged parameters as determinants of the variance of the residuals to change
Equation (14.17) to:
q

ht = γ0 + 2
γj ut−j (14.20)
j=1
which denotes an ARCH(q) process. ARCH(q) models are useful when the variability
of the series is expected to change more slowly than in the ARCH(1) model. However,
ARCH(q) models are quite often difficult to estimate, because they frequently yield
negative estimates of the γj s. To resolve this issue, Bollerslev (1986) developed the idea
of the GARCH model, which will be examined in the next section.
The GARCH model

One of the drawbacks of the ARCH specification, according to Engle (1995), was that
it looked more like a moving average specification than an autoregression. From this,
a new idea was born, which was to include the lagged conditional variance terms as
autoregressive terms. This idea was worked out by Tim Bollerslev, who in 1986 pub-
lished a paper entitled ‘Generalised Autoregressive Conditional Heteroskedasticity’ in
the Journal of Econometrics, introducing a new family of GARCH models.
The GARCH( p, q) model

The GARCH(p, q) model has the following form:
Yt = a + β Xt + ut (14.21)
p q

ht = γ0 + δi ht−i + 2
γj ut−j (14.22)
i=1 j=1
which says that the value of the variance scaling parameter ht now depends both on
past values of the shocks, which are captured by the lagged squared residual terms, and
on past values of itself, which are captured by lagged ht terms.
It should be clear by now that for p = 0 the model reduces to ARCH(q). The sim-
plest form of the GARCH(p,q) model is the GARCH(1,1) model, for which the variance
equation has the form:
2
ht = γ0 + δ1 ht−1 + γ1 ut−1 (14.23)
This model specification usually performs very well and is easy to estimate because it
has only three unknown parameters: γ0 , γ1 and δ1 .
The GARCH(1,1) model as an infinite ARCH process

To show that the GARCH(1,1) model is a parsimonious alternative to an infinite
ARCH(q) process, consider Equation (14.23). Successive substitution into the right-hand
side of Equation (14.23) gives:
2
ht = γ0 + δht−1 + γ1 ut−1

2
= γ0 + δ γ0 + δht−2 + γ1 ut−2 2
+ γ1 ut−1
2
= γ0 + γ1 ut−1 2
+ δγ0 + δ 2 ht−2 + δγ1 ut−2

2
= γ0 + γ1 ut−1 + δγ0 + δ 2 γ0 + δht−3 + γ1 ut−3
2 2
+ δγ1 ut−2
···
γ0
= 2
+ γ1 ut−1 2
+ δut−2 2
+ δ 2 γ1 ut−3 + ···
1−δ
∞
γ0
= + γ1 δ j−1 ut−j
2 (14.24)
1−δ
j=1
which shows that the GARCH(1,1) specification is equivalent to an infinite order ARCH
model with coefficients that decline geometrically. For this reason, it is essential to
estimate GARCH(1,1) models as alternatives to high-order ARCH models, because with
the GARCH(1,1) there are fewer parameters to estimate and therefore fewer degrees
of freedom are lost.
Estimating GARCH models in EViews

Consider again the r-ftse series from the ARCH.wf1 file. To estimate a GARCH model,
click on Quick/Estimate Equation, to open the Equation Specification window, and
again change the estimation method by clicking on the down arrow in the method set-
ting and choosing the ARCH-Autoregressive Conditional Heteroskedasticity option.
In this new Equation Specification window, the upper part is for the mean equation
specification while the lower part is for the ARCH/GARCH specification or the variance
equation. To estimate a simple GARCH(1,1) model, assuming that the mean equation
as before follows an AR(1) process, in the mean equation specification window, we type:
making sure that within the ARCH-M part None is selected, which is the default in
EViews. For the ARCH/GARCH specification choose GARCH/TARCH from the drop-
down Model: menu, which is again the default EViews case, and in the small boxes
type 1 for the Order ARCH and 1 for the GARCH. It is obvious that for higher orders,
for example a GARCH(4,2) model, you would have to change the number in the small
boxes by typing 2 for the Order ARCH and 4 for the GARCH. After specifying the
number of ARCH and GARCH and clicking OK the required results appear. Table 14.6
presents the results for a GARCH(1,1) model.
Note that it took only five iterations to reach convergence in estimating this model.
The model can be written as:
Yt = 0.0004 + 0.0644Yt−1 + ût (14.25)

(2.57) (3.05)
2
ht = 0.0000002 + 0.893ht−1 + 0.084ût−1 (14.26)
(4.049) (59.43) (7.29)
Table 14.6 A GARCH(1,1) model for the FTSE-100

Method: ML–ARCH
Date: 12/26/03 Time: 18:52
Sample: 1/01/1990 12/31/1999
C 0.000409 0.000158 2.578591 0.0099

R_FTSE(−1) 0.064483 0.021097 3.056426 0.0022
Variance equation
C 2.07E − 06 5.10E−07 4.049552 0.0001

ARCH(1) 0.084220 0.011546 7.294102 0.0000
GARCH(1) 0.893243 0.015028 59.43780 0.0000

with values of z-statistics in parentheses. Note that the estimate of δ is highly significant
and positive, as well as the coefficient of the γ1 term. Taking the variance series for the
GARCH(1,1) model (by clicking on Procs/Make GARCH Variance Series) it has been
renamed as GARCH11 and this series has been plotted together with the ARCH6 series
to obtain the results shown in Figure 14.5.
From this we observe that the two series are quite similar (if not identical), because
the GARCH term captures a high order of ARCH terms as was proved earlier. Therefore,
again, it is better to estimate a GARCH instead of a high order ARCH model because of
its easier estimation and the least possible loss of degrees of freedom.
Changing the values in the boxes of the ARCH/GARCH specification to 6 in order to
estimate a GARCH(6,6) model, the results shown in Table 14.7 are obtained, where the
insignificance of all the parameters apart from the ARCH(1) term suggests that it is not
an appropriate model.
Similarly, estimating a GARCH(1,6) model gives the results shown in Table 14.8,
where now only the ARCH(1) and the GARCH(1) terms are significant; also some of
the ARCH lagged terms have a negative sign. Comparing all the models from both the
ARCH and the GARCH alternative specifications, we conclude that the GARCH(1,1) is
preferred, for the reasons discussed above.
Alternative specifications
There are many alternative specifications that could be analysed to model conditional
volatility, and some of the more important variants are presented briefly in this section.
(Berra and Higgins (1993) and Bollerslev et al. (1994) provide very good reviews of these
0.0006
0.0005
0.0004
0.0003
0.0002
0.0001
0.0000
1/01/90 11/01/93 9/01/97
ARCH6 GARCH1,1
Figure 14.5 Plots of the conditional variance series for ARCH(6) and GARCH(1,1)
alternative specifications, while Engle (1995) collects some important papers in the
ARCH/GARCH literature.)
The GARCH in mean or GARCH-M model

GARCH-M models allow the conditional mean to depend on its own conditional vari-
ance. Consider, for example, investors who are risk-averse and therefore require a
premium as compensation for holding a risky asset. That premium is clearly a posi-
tive function of the risk (that is the higher the risk, the higher the premium should
be). If the risk is captured by the volatility or by the conditional variance, then the
conditional variance may enter the conditional mean function of Yt .
Therefore, the GARCH-M(p,q) model has the following form:
Yt = a + β Xt + θ ht + ut (14.27)
p q

ht = γ0 + δi ht−i + 2
γj ut−j (14.28)
i=1 j=1
Another variant of the GARCH-M type model is to capture risk not through the
variance series but by using the standard deviation of the series having the following

Method: ML–ARCH
Date: 12/26/03 Time: 19:05
Sample: 1/01/1990 12/31/1999
C 0.000433 0.000160 2.705934 0.0068

R_FTSE(−1) 0.065458 0.020774 3.150930 0.0016
Variance equation
C 1.70E−06 7.51E−06 0.227033 0.8204

ARCH(1) 0.038562 0.015717 2.453542 0.0141
ARCH(2) 0.070150 0.113938 0.615692 0.5381
ARCH(3) 0.022721 0.269736 0.084234 0.9329
ARCH(4) −0.017544 0.181646 −0.096585 0.9231
ARCH(5) 0.011091 0.077074 0.143905 0.8856
ARCH(6) −0.017064 0.063733 −0.267740 0.7889
GARCH(1) 0.367407 3.018202 0.121730 0.9031
GARCH(2) 0.116028 1.476857 0.078564 0.9374
GARCH(3) 0.036122 1.373348 0.026302 0.9790
GARCH(4) 0.228528 0.819494 0.278864 0.7803
GARCH(5) 0.217829 0.535338 0.406900 0.6841
GARCH(6) −0.092748 0.979198 −0.094719 0.9245

Adjusted R-squared −0.000465 S.D. dependent var 0.009398
specification for the mean and the variance equation:

Yt = a + β Xt + θ ht + ut (14.29)
p q

ht = γ0 + δi ht−i + 2
γj ut−j (14.30)
i=1 j=1
GARCH-M models can be linked with asset–pricing models such as the capital asset–
pricing models (CAPM) with many financial applications (for more, see Campbell et al.
1997; Hall et al. 1990).
Estimating GARCH-M models in EViews

To estimate a GARCH-M model in EViews, first click Quick/Estimate Equation to open
the Estimation Window, then change the estimation method by clicking on the down
arrow in the method setting and choosing the ARCH-Autoregressive Conditional

Method: ML–ARCH
Date: 12/26/03 Time: 19:34
Sample: 1/01/1990 12/31/1999
C 0.000439 0.000158 2.778912 0.0055

R_FTSE(−1) 0.064396 0.020724 3.107334 0.0019
Variance equation
C 9.12E−07 2.79E−07 3.266092 0.0011

ARCH(1) 0.040539 0.013234 3.063199 0.0022
ARCH(2) 0.048341 0.025188 1.919235 0.0550
ARCH(3) −0.027991 0.031262 −0.895354 0.3706
ARCH(4) −0.037356 0.028923 −1.291542 0.1965
ARCH(5) 0.016418 0.028394 0.578219 0.5631
ARCH(6) 0.015381 0.023587 0.652097 0.5143
GARCH(1) 0.934786 0.011269 82.95460 0.0000

Heteroskedasticity option. In this new Equation Specification window, the upper

part is again for the mean equation specification while the lower part is for the
ARCH/GARCH specification or the variance equation. To estimate a GARCH-M(1,1)
model, assuming that the mean equation (as before) follows an AR(1) process, type in
the mean equation specification:
and this time click on either Std.Dev or the Var selections from the ARCH-M part for
versions of the mean Equations (14.29) and (14.27), respectively.
For the ARCH/GARCH specification choose GARCH/TARCH from the drop-down
Model: menu, which is again the default EViews case, and in the small boxes specify
by typing the number of the q lags (1, 2, . . . , q) for the Order ARCH and the number of
p lags (1, 2, . . . , p) for the GARCH. Table 14.9 presents the results for a GARCH-M(1,1)
model based on the specification that uses the variance series to capture risk in the
mean equation, as given by Equation (14.27).
Note that the variance term (GARCH) in the mean equation is slightly significant
but its inclusion substantially increases the significance of the GARCH term in the
variance equation. Re-estimating the above model but this time clicking on the Std.Dev
from the ARCH-M part to include the conditional standard deviation in the mean
equation. The results are presented in Table 14.10, where this time the conditional
305
Table 14.9 A GARCH-M(1,1) model for the FTSE-100

Method: ML – ARCH
Date: 12/26/03 Time: 19:32
Sample: 1/01/1990 12/31/1999
GARCH 6.943460 4.069814 1.706088 0.0880

C −2.39E−05 0.000311 −0.076705 0.9389
R_FTSE(−1) 0.061006 0.020626 2.957754 0.0031
Variance equation
C 7.16E-07 2.22E−07 3.220052 0.0013

ARCH(1) 0.049419 0.006334 7.801997 0.0000
GARCH(1) 0.942851 0.007444 126.6613 0.0000

Table 14.10 A GARCH-M(1,1) for the FTSE-100 (using standard deviation)

Method: ML – ARCH
Date: 12/26/03 Time: 19:36
Sample: 1/01/1990 12/31/1999
SQR(GARCH) 0.099871 0.080397 1.242226 0.2142

C −0.000363 0.000656 −0.553837 0.5797
R_FTSE(−1) 0.063682 0.020771 3.065923 0.0022
Variance equation
C 9.23E-07 2.72E−07 3.394830 0.0007

ARCH(1) 0.055739 0.007288 7.647675 0.0000
GARCH(1) 0.934191 0.008832 105.7719 0.0000

standard deviation (or SQR(GARCH)) coefficient is not significant, suggesting that if

there is an effect of the risk on the mean return, this is captured better by the variance.
The threshold GARCH (TGARCH) model

A major restriction of the ARCH and GARCH specifications above is that they are sym-
metric. By this we mean that what matters is only the absolute value of the innovation
and not its sign (because the residual term is squared). Therefore, in ARCH/GARCH
models a large positive shock will have exactly the same effect in the volatility of the
series as a large negative shock of the same magnitude. However, for equities it has
been observed that negative shocks (or ‘bad news’) in the market have a larger impact
on volatility than do positive shocks (or ‘good news’) of the same magnitude.
The threshold GARCH model was introduced by the works of Zakoian (1990) and
Glosten et al. (1993). The main target of this model is to capture asymmetries in terms
of negative and positive shocks. To do this, simply add into the variance equation
a multiplicative dummy variable to check whether there is a statistically significant
difference when shocks are negative.
The specification of the conditional variance equation (for a TGARCH(1,1)) is
given by:
2
ht = γ0 + γ ut−1 2 d
+ θ ut−1 t−1 + δht−1 (14.31)
where dt takes the value of 1 for ut < 0, and 0 otherwise. So ‘good news’ and ‘bad
news’ have different impacts. Good news has an impact of γ , while bad news has an
impact of γ + θ. If θ > 0 we conclude that there is asymmetry, while if θ = 0 the news
impact is symmetric. TGARCH models can be extended to higher order specifications
by including more lagged terms, as follows:
q q

ht = γ0 + 2 +
(γi + θi dt−i )ut−i δj ht−j (14.32)
i=1 j=1
Estimating TGARCH models in EViews

To estimate a TGARCH model in EViews, first click Quick/Estimate Equation to open
the Estimation Window. Then change the estimation method by clicking on the down
arrow in the method setting, to choose the ARCH-Autoregressive Conditional Het-
eroskedasticity option. In this new Equation Specification window we again have the
upper part for the mean equation specification and the lower part for the ARCH/GARCH
specification or the variance equation. To estimate a TGARCH(p,q) model, assuming
that the mean equation follows an AR(1) process as before, type in the mean equation
specification:
Table 14.11 A TGARCH(1,1) model for the FTSE-100

Method: ML–ARCH
Date: 12/27/03 Time: 15:04
Sample: 1/01/1990 12/31/1999
C 0.000317 0.000159 1.999794 0.0455

R_FTSE(−1) 0.059909 0.020585 2.910336 0.0036
Variance equation
C 7.06E − 07 1.90E − 07 3.724265 0.0002

ARCH(1) 0.015227 0.006862 2.218989 0.0265
(RESID<0)*ARCH(1) 0.053676 0.009651 5.561657 0.0000
GARCH(1) 0.950500 0.006841 138.9473 0.0000

ensuring also that None was clicked on in the ARCH-M part of the mean equation
specification.
For the ARCH/GARCH specification, choose GARCH/TARCH from the drop-down
Model: menu, and specify the number of q lags (1, 2, . . . , q) for the Order ARCH, the
number of p lags (1, 2, . . . , p) for the Order GARCH and the Threshold Order by chang-
ing the value in the box from 0 to 1 to have the TARCH model in action. Table 14.11
presents the results for a TGARCH(1,1) model.
Note that, because the coefficient of the (RESID < 0)∗ ARCH(1) term is positive and
statistically significant, indeed for the FTSE-100 there are asymmetries in the news.
Specifically, bad news has larger effects on the volatility of the series than good news.
The exponential GARCH (EGARCH) model

The exponential GARCH (EGARCH) model was first developed by Nelson (1991),
and the variance equation for this model is given by:

q
u q p
t−j ut−j
log(ht ) = γ + ζj + ξj + δi log(ht−i ) (14.33)

j=1 ht−j j=1 ht−j i=1
where γ , the ζ s, ξ s and δs are parameters to be estimated. Note that the left-hand side
is the log of the variance series. This makes the leverage effect exponential rather than
quadratic, and therefore the estimates of the conditional variance are guaranteed to
be non-negative. The EGARCH model allows for the testing of asymmetries as well
as the TGARCH. To test for asymmetries, the parameters of importance are the ξ s. If
ξ1 = ξ2 = · · · = 0, then the model is symmetric. When ξj < 0, then positive shocks
(good news) generate less volatility than negative shocks (bad news).
Estimating EGARCH models in EViews

To estimate an EGARCH model in EViews, first click Quick/Estimate Equation to open
the Estimation Window. Then change the estimation method by clicking the down
arrow in the method setting to choose the ARCH-Autoregressive Conditional Het-
eroskedasticity option. In this new Equation Specification window we again have
the upper part for the mean equation specification, while the lower part is for the
ARCH/GARCH specification or the variance equation. To estimate an EGARCH(p,q)
model, assuming that the mean equation follows an AR(1) process, as before type in
the mean equation specification:
again making sure that None is clicked on in the ARCH-M part of the mean
equation specification.
For the ARCH/GARCH specification now choose EGARCH from the drop-down
Model: menu, and in the small boxes specify the number of the q lags (1, 2, . . . , q)
for the Order ARCH and the number of p lags (1, 2, . . . , p) for the GARCH. Table 14.12
presents the results for an EGARCH(1,1) model.
Table 14.12 An EGARCH(1,1) model for the FTSE-100

Method: ML–ARCH
Date: 12/26/03 Time: 20:19
Sample: 1/01/1990 12/31/1999
C 0.000306 0.000156 1.959191 0.0501

R_FTSE(−1) 0.055502 0.020192 2.748659 0.0060
Variance equation
C −0.154833 0.028461 −5.440077 0.0000

|RES|/SQR[GARCH](1) 0.086190 0.012964 6.648602 0.0000
RES/SQR[GARCH](1) −0.044276 0.007395 −5.987227 0.0000
EGARCH(1) 0.990779 0.002395 413.7002 0.0000

Note that, because the coefficient of the RES/SQR[GARCH](1) term is negative and
statistically significant, indeed for the FTSE-100 bad news has larger effects on the
volatility of the series than good news.
Adding explanatory variables in the mean equation

ARCH/GARCH models may be quite sensitive to the specification of the mean equation.
Consider again for example, the FTSE-100 return series examined above. In all our
analyses it was assumed (quite restrictively and without prior information) that a good
specification for the mean equation would be a simple AR(1) model. It is obvious that,
using daily data, AR models of a higher order would be more appropriate. Also, it
might be more appropriate to use MA terms alongside the AR terms. Estimating an
ARCH(1) and a GARCH(1,1) model for the FTSE-100 returns, assuming that it follows
an ARMA(1,1) specification, in both cases gives results for the mean equation that are
statistically insignificant. (We leave this as an exercise for the reader. To the mean
equation specification, type in: r_ftse c AR(1) MA(1), and then arrange the number of
ARCH(q) and GARCH(p) terms.) It should be clear that results, or even convergence
of iterations, might be highly affected by wrong specifications of the mean equation,
and if research using GARCH models is to be undertaken, the researcher has to be very
careful first to identify the correct specification.
Adding explanatory variables in the variance equation

GARCH models also allow us to add explanatory variables in the specification of the
conditional variance equation. We can have an augmented GARCH(q,p) specification
such as the following:
p q

m
ht = γ0 + δi ht−i + 2 +
γj ut−j µk Xk (14.34)
i=1 j=1 k=1
where xk is a set of explanatory variables that might help to explain the variance. As an
example, consider the case of the FTSE-100 returns once again, and test the assumption
that the Gulf War (which took place in 1994) affected the FTSE-100 returns, making
them more volatile. This can be tested by constructing a dummy variable, named
Gulf, which will take the value of 1 for observations during 1994 and 0 for the rest
of the period. Then in the estimation of the GARCH model, apart from specifying
as always the mean equation and the order of q and p in the variance equation, add
the dummy variable in the box where EViews allows the entry of variance regressors,
by typing the name of the variable there. Estimation of a GARCH(1,1) model with
the dummy variable in the variance regression gave the results shown in Table 14.13,
where it can be seen that the dummy variable is statistically insignificant, so the
hypothesis that the Gulf War affected the volatility of the FTSE-100 returns can be
rejected. Other examples with dummy and regular explanatory variables are given in
Table 14.13 A GARCH(1,1) model with an explanatory variable in the variance equation
Method: ML–ARCH
Date: 12/27/03 Time: 17:25
Sample: 1/01/1990 12/31/1999
C 0.000400 0.000160 2.503562 0.0123

R_FTSE(−1) 0.068514 0.021208 3.230557 0.0012
Variance equation
C 2.22E−06 6.02E−07 3.687964 0.0002

ARCH(1) 0.083656 0.013516 6.189428 0.0000
GARCH(1) 0.891518 0.016476 54.11098 0.0000
GULF −4.94E−07 5.96E−07 −0.829246 0.4070

the empirical illustration section below for the GARCH model of UK GDP and the effect
of socio-political instability.
Estimating ARCH/GARCH-type models in Stata

All the analyses performed in the previous sections using Eviews can be performed with
Stata, using the following commands. The data are given in the file named ARCH.dat.
First, to obtain simple OLS results for the r_ftse daily time series regressed to a lag of the
same series (r_ftset−1 ) the command is:
regress r_ftse L.rftse
where L. denotes the lag operator. The results are similar to those in Table 14.1.
To test for ARCH effects, the command is:
estat archlm, lags(1)
The results are similar to those reported in Table 14.2 and suggest that there are ARCH
effects in the series. To test for ARCH effects of a higher order (order 6 in the example
reported in Table 14.3), the command is:
estat archlm, lags(6)
Then, to estimate the ARCH model, the command syntax is:

arch depvar indepvars , options
where depvar is replaced with the name of the dependent variable and indepvars
with the names of the independent variables you want to include in the mean equation,
and after the comma choose from the options which type of ARCH/GARCH model you
wish to estimate (that is you specify the variance equation). Thus, for a simple ARCH(1)
model of regressing r_ftse to r_ftset−1 , in the mean equation the command is:
arch r_ftse L.r_ftse , arch(1)
Then, to obtain the ht variance series of this ARCH(1) model, the command is:
predict htgarch1 , variance
(Here, htgarch1 is a name that helps us remember that the series is a variance series
for the ARCH(1) model; any other name the reader might want to give to the series will
work just as well): while the command:
tsline htgarch1
provides a time plot of the variance series.

Continuing, the commands for an ARCH(6) model are:
arch r_ftse L.r_ftse , arch(6)

tsline htgarch6
For an ARCH-M(1) model:
arch r_ftse L.r_ftse , archm arch(1)

predict htgarchm1 , variance
tsline htgarchm1
For a GARCH(1,1) model:
arch r_ftse L.r_ftse , arch(1) garch(1)

tsline htgarch11
while for higher orders (that is for GARCH(3,4)) only the values in the parentheses
should change:
arch r_ftse L.rftse , arch(1/3) garch(1/4)
The TGARCH(1,1,1) model is given by:
arch r_ftse L.r_ftse , arch(1) garch(1) tarch(1)

and, finally, the EGARCH(1,1,1) model is estimated by:
arch r_ftse L.r_ftse , arch(1) garch(1) earch(1)
All these commands are left as an exercise for the reader. The analysis and interpretation
of the results are similar to those discussed previously in this chapter.
Estimating ARCH/GARCH models in Microfit

In Microfit, apart from the TGARCH model, all ARCH/GARCH-type models can be
estimated automatically using Microfit’s own menus. The data are given in the ARCH.fit
file. The first step is to go to the volatility estimation menu by clicking on the Volatility
Microfit button. This menu looks exactly like the single-equation menu, and here the
mean equation to be estimated is defined. In this case, type:
r_ftse c r_ftse(-1)
then click on Run, which brings up the GARCH estimation menu. Here, a set of
options is provided, and in each case you need to define which model you want to
estimate from six possible choices:
GARCH
GARCH-M
AGARCH
AGARCH-M
EGARCH
EGARCH-M
Leaving aside cases 3 and 4 of absolute GARCH models, all the rest of the options are
familiar to us from the theory in this chapter. So, to estimate a GARCH-M(1,1) model,
choose option 2 from this list and click OK. Then Microfit requires you to specify the
underlying distribution. This is left as the default case, which is the z-distribution. After
clicking OK again a new window appears, where the orders of ARCH and GARCH terms
in our model must be specified. First, type the number of the GARCH terms and then,
separated by “;”, the number of ARCH terms. Therefore, for GARCH-M(1,1) type:
1 ; 1
Then click Run again, which takes you to the window where you can specify the
number of additional variables to be included in the Variance equation (we can leave
this blank for this example). After clicking Run again the results appear, after a number
of iterations that are shown on the screen while Microfit executes the calculations. The
analysis and interpretation are similar to the cases that have been examined above. The
rest of the ARCH/GARCH models have been left as exercises for the reader.
Empirical illustrations of ARCH/GARCH models

A GARCH model of UK GDP and the effect of
socio-political instability
Asteriou and Price (2001) used GARCH models to capture the effects of socio-political
instability in UK GDP. To approximate and quantify socio-political instability, they
constructed indices that summarized various variables capturing phenomena of social
unrest for the UK over the period 1960–97 using quarterly time series data. Specifically,
their indices were constructed by applying the method of principal components to
the following variables: TERROR, the number of terrorist activities that caused mass
violence; STRIKES, the number of strikes that were caused by political reasons; ELECT,
the number of elections; REGIME, a dummy variable that takes the value of one for
government changes to different political parties, zero otherwise; FALKL, a dummy
variable that takes the value of 1 for the period of the Falklands War (1982; q1–q4),
zero otherwise; and finally GULF, a dummy variable which takes the value of 1 for
the period of the first Gulf War (1991; q1–q4), zero otherwise. Their main results are
presented below.
Results from GARCH models

Asteriou and Price (2001) estimated the following model:

4
4
6

ln(Yt ) = a0 + a1i
ln(Yt−i ) + a2i
ln(It−i ) + dj Xjt + ut (14.35)
i=0 i=0 j=1
ut ∼ N(0, ht ) (14.36)
2 +b h
ht = b1 et−1 (14.37)
2 t−1
That is, the growth rate of GDP (denoted by

ln(Yt )) is modelled as an AR(4) process,
including the growth and four lags of investments (denoted by
ln(It )) plus the polit-
ical instability proxies (Xjt ), where the variance is conditioned on the lagged variance
and lagged squared residuals.
Table 14.14, model 1 presents the results of a GARCH(1,1) model for GDP growth or
reference without including political dummies. (In each case the model has first been
estimated with four lagged terms of GDP per capita and four lagged terms of the rate of
growth of investment, and subsequently reduced to a parsimonious model, including
only the significant regressors.) Despite the low R2 , the variance part of the model
fits well.
Continuing, Asteriou and Price re-estimated the above model, including in Equation
(14.35) the political dummies. All the dummies entered the equation with the expected
negative sign and three of them were statistically significant. The results of the parsi-
monious model are shown in Table 14.14, model 2, and from these we observe that
REGIME, TERROR and STRIKES are highly significant and negative. The variance equa-
tion is improved and R2 , while it remains relatively low, is increased compared to the
previous specification.
Table 14.14 GARCH estimates of GDP growth with political uncertainty proxies
Dependent variable:
ln(Yt ); Sample: 1961q2–1997q4
Parameter 1 2 3 4
Constant 0.003 (3.49) 0.005 (3.78) 0.004 (3.80) 0.006 (5.66)
ln(Yt −3 ) 0.135 (1.36) 0.194 (1.99) 0.186 (1.87) 0.270 (3.42)
ln(Yt −4 ) 0.131 (1.23) 0.129 (1.22) 0.122 (1.48) 0.131 (1.29)
ln(It −2 ) 0.180 (2.25) 0.132 (1.48) 0.162 (1.92)

REGIME −0.012 (−4.91) −0.012 (−5.63)
TERROR −0.004 (−2.72) −0.005 (−2.66)
STRIKES −0.011 (−2.58) −0.015 (−3.44)
PC1 −0.005 (−4.33)
PC2 −0.003 (−2.02)
Variance equation
Constant 0.00001 (1.83) 0.00001 (1.66) 0.000006 (1.16) 0.00006 (1.71)
ARCH (1) 0.387 (3.27) 0.314 (2.44) 0.491 (4.18) 0.491 (4.46)
GARCH (1) 0.485 (2.95) 0.543 (3.14) 0.566 (6.21) 0.566 (3.36)
R2 0.006 0.099 0.030 0.104
S.E . of d .v . 0.010 0.010 0.010 0.010
S.E . of Reg. 0.010 0.010 0.010 0.010
The results from the alternative specification, with the inclusion of the PCs in place
of the political instability variables (Table 14.14, model 3) are similar to the previ-
ous model. Negative and significant coefficients were obtained for the first and the
third components.
Asteriou and Price (2001) also estimated all the above specifications without including
the investment terms. The results for the case of the political uncertainty dummies are
presented in the same table in model 4, and show clearly that the strong negative
direct impact remains. Thus, the impact of political uncertainty on growth does not
appear to operate through investment growth, leaving open the possibility of political
uncertainty affecting the level of investment.
Results from GARCH-M models

Asteriou and Price (2001) argued that it is mainly political instability that affects
uncertainty and thereby growth. So it was of considerable interest for them to allow
uncertainty to affect growth directly. To do this they used the GARCH-M class of mod-
els, first to test whether uncertainty in GDP (conditioned by the ‘in mean’ term of the
GARCH-M model) affects GDP growth, and second whether political instability (condi-
tioned by the political dummies and by the PCs in the variance equation) affects GDP
growth separately.
The GARCH-M model they estimated may be presented as follows:

4
4

ln(Yt ) = a0 + a1i
ln(Yt−i ) + a2i
ln(It−i ) + γ ht + ut (14.38)
i=0 i=0
ut ∼ N(0, ht ) (14.39)

6
2
ht = b1 ut−1 + b2 ht−1 + b3i Xit (14.40)
i=1
Table 14.15 GARCH-M(1,1) estimates with political uncertainty proxies

Dependent variable:
Parameter 1 2 3
Constant 0.008 (2.67) 0.009 (4.22) 0.007 (4.33)
ln(Yt −3 ) 0.154 (1.59) 0.175 (1.15) 0.161 (2.10)
ln(Yt −4 ) 0.128 (1.24) 0.089 (0.81) 0.141 (1.84)
ln(Ivt −2 ) 0.136 (1.69) 0.132 (1.33) 0.126 (1.84)

SQR(GARCH ) −0.498 (−1.40) −0.674 (−3.07) −0.444 (−2.42)
Variance equation
Constant 0.00001 (1.68) 0.00005 (1.21) 0.000002 (0.80)
ARCH (1) 0.335 (3.07) 0.133 (1.33) 0.460 (4.05)
GARCH (1) 0.554 (3.53) 0.650 (4.00) 0.580 (6.64)
ELECT 0.007 (3.11)
REGIME 0.006 (2.84)
FAUKL 0.002 (5.11)
STRIKES 0.066 (2.91)
PC1 0.000047 (1.45)
PC2 0.000002 (0.09)
PC3 0.000031 (3.20)
R2 0.054 0.053 0.064
S.E . of d .v . 0.010 0.0106 0.0106
S.E . of Reg. 0.010 0.0108 0.0107
That is, the growth rate of GDP is modelled as an AR process, including four lags of
the growth rate of investments and the variance of the error term. Equation (14.39)
defines ht as the variance of the error term in Equation (14.38), and Equation (14.40)
states that the variance of the error term is in turn a function of the lagged variance
and lagged squared residuals as well as the political instability proxies Xit . To accept the
first hypothesis it would be necessary for γ to be non-zero, while to accept the second
hypothesis there should be evidence of positive statistically significant estimates for
the coefficients of the political instability proxies (b3i ).
Table 14.15, model 1 reports the results of estimating a GARCH-M(1,1) model without
political instability proxies. (Again, as in the previous section, the reported results are
only from the parsimonious models.) The model is satisfactory given that the parame-
ters (b1 , b2 ) are strongly significant. The inclusion of the ‘in mean’ specification turns
out to be redundant as γ is insignificant, suggesting that GDP uncertainty does not
itself affect GDP growth. However, this turns out to be misleading and follows from
the fact that political factors are ignored.
In estimating a GARCH-M(1,1) model including the political dummies in the variance
equation (see Table 14.15, model 2), Asteriou and Price observed that all the political
instability variables – with the exception of REGIME – entered the equation with the
expected positive sign, indicating that political uncertainty increases the variance of
GDP growth. All variables were statistically significant. The ‘in mean’ term is in this
case highly significant and negative. The results from the alternative specification, with
the inclusion of the PCs in the place of the political instability variables (Table 14.15,
model 3) are similar to the previous one, with the exception that positive and significant
coefficients were obtained only for the fifth component.
Continuing, Asteriou and Price estimated more general GARCH-M(1,1) models,
first including the political dummies and the PCs in the growth equation, and then
including political dummies and PCs in both the growth and the variance equation.
Table 14.16 GARCH-M(1,1) estimates with political proxies

Dependent variable:
Parameter Estimate Std. error t-statistic
Constant 0.009 0.003 2.964
ln(Yt −3 ) 0.206 0.093 2.203
ln(Yt −4 ) 0.123 0.102 1.213
ln(It −4 ) 0.109 0.088 1.241

SQR(GARCH ) −0.447 0.365 −1.304
REGIME −0.012 0.002 −5.084
TERROR −0.005 0.001 −3.018
STRIKES −0.012 0.004 −2.753
Variance equation
Constant 0.00001 0.000008 1.648
ARCH (1) 0.285 0.120 2.380
GARCH (1) 0.575 0.161 3.553
R2 0.124
S.E . of d .v . 0.0106
S.E . of Reg. 0.0103
Table 14.17 GARCH-M(1,1) estimates with political proxies

Dependent variable:
Parameter Estimate Std. error t-statistic
Constant 0.005 0.001 3.611
ln(Yt −3 ) 0.172 0.095 1.799
ln(Yt −4 ) 0.123 0.090 1.353
ln(It −4 ) 0.181 0.089 2.023

SQR(GARCH ) −0.169 0.254 −0.667
REGIME −0.013 0.006 −1.925
GULF −0.007 0.003 −1.899
STRIKES −0.020 0.006 −3.356
Variance equation
Constant 0.00002 0.00001 2.013
ARCH (1) 0.265 0.126 2.091
GARCH (1) 0.527 0.171 3.076
ELECT 0.00004 0.00001 2.608
REGIME 0.0001 0.0001 1.131
FALKL 0.00002 0.00002 1.326
R2 0.141
S.E . of d .v . 0.0106
S.E . of Reg. 0.0103
With the first version of the model they wanted to test whether the inclusion of
the dummies in the growth equation would affect the significance of the ‘in mean’
term which captures the uncertainty of GDP. Their results, presented in Table 14.16,
showed that GDP growth was significantly affected only by political uncertainty, cap-
tured either by the dummies or by the PCs, denoting the importance of political factors
other than the GARCH process. (We report here only the results from the model with
the political uncertainty dummies. The results with the PCs are similar but are not
presented for economy of space. Tables and results are available from the authors
on request.)
The final and most general specification was used to capture both effects stemming
from political uncertainty, namely the effect of political uncertainty on GDP growth,
and its effect on the variance of GDP. Asteriou and Price’s results are presented in
Table 14.17. After the inclusion of the political dummies in the variance equation,
the model was improved (the political dummies significantly altered the variance of
GDP), but the effect on GDP growth came only from the political uncertainty prox-
ies that were included in the growth equation. The ‘in mean’ term was negative and
insignificant.
The final conclusion of Asteriou and Price (2001) was that political instability has
two identifiable effects. Some measures impact on the variance of GDP growth; others
directly affect the growth itself. Instability has a direct impact on growth and does not
operate indirectly via the conditional variance of growth.
Questions and exercises

Questions
1 Explain the meaning of ARCH and GARCH models, showing how each is a form
of heteroskedasticity.
2 Explain how one can test for the presence of ARCH(q) effects in a simple OLS
estimation framework.
3 Explain how one may estimate models with ARCH and GARCH effects.
4 What is meant by the comment that ‘GARCH(1,1) is an alternative parsimonious
process for an infinite ARCH(q) process’. Prove this mathematically.
5 Explain the meaning of asymmetries in news, and provide appropriate specifications
for GARCH models that can capture these effects.
6 What should researchers be very careful of in estimating ARCH/GARCH models?
7 Provide a GARCH-M(q,p) model and explain the intuition behind this model.
8 Explain the effect of the dummy variable in the TGARCH model. Why does it enter
the variance equation in a multiplicative form, and what is the rationale behind this?
Exercise 14.1
The file arch.wf1 contains daily data for the logarithmic returns FTSE-100 (named r_ftse)
and three more stocks of the UK stock market (named r_stock1, r_stock2 and r_stock3,
respectively). For each of the stock series do the following:
(a) Estimate an AR(1) up to AR(15) model and test the individual and joint significance
of the estimated coefficients.
(b) Compare AIC and SBC values of the above models and, along with the results for
the significance of the coefficients, conclude which will be the most appropriate
specification.
(c) Re-estimate this specification using OLS and test for the presence of ARCH(p) effects.
Choose several alternative values for p.
(d) For the preferred specification of the mean equation, estimate an ARCH(p) model
and compare your results with the previous OLS results.
(e) Obtain the conditional variance and conditional standard deviations series and
rename them with names that will show from which model they were obtained
(for example SD_ARCH6 for the conditional standard deviation of an ARCH(6)
process).
(f ) Estimate a GARCH(q,p) model, obtain the conditional variance and standard devia-
tion series (rename them again appropriately) and plot them against the series you
have already obtained. What do you observe?
(g) Estimate a TGARCH(q,p) model. Test the significance of the TGARCH coefficient.
Is there any evidence of asymmetric effects?
(h) Estimate an EGARCH(q,p) model. How does this affect your results?
(i) Summarize all models in one table and comment on your results.
Exercise 14.2
You are working in a financial institution and your boss proposes to upgrade the finan-
cial risk-management methodology the company uses. In particular, to model the
FTSE-100 index your boss suggests estimation using an ARCH(1) process. You disagree
and wish to convince your boss that a GARCH(1,1) process is better.
(a) Explain, intuitively first, why a GARCH(1,1) process will fit the returns of FTSE-100
better than an ARCH(1) process. (Hint: You will need to refer to the stylized facts
of the behaviour of stock indices.)
(b) Prove your point with the use of mathematics. (Hint: You will need to mention
ARCH(q) processes here.)
(c) Estimate both models and try to analyse them in such a way that you can convince
your boss about the preferability of the model you are proposing. Check the condi-
tional standard deviation and conditional variance series as well. (Hint: Check the
number of iterations and talk about computational efficiency.)

Asteriou - Series de Tiempo

Uploaded by

Copyright:

Available Formats

Asteriou - Series de Tiempo

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Asteriou - Series de Tiempo

Uploaded by

Copyright:

Available Formats

Applied Econometrics

V Time Series Econometrics

13 ARIMA Models and the Box–Jenkins Methodology 265

An introduction to time series econometrics

In its simplest terms a time series Yt is said to be stationary if:

(a) E(Yt ) = constant for all t;

Autoregressive time series models

will happen in t is largely dependent on what happened in t − 1. Alternatively, what

Condition for stationarity

Example of stationarity in the AR(1) model

smpl 1 1 [sets the sample to be the ﬁrst observation

Figure 13.1 Plot of an AR(1) model

Figure 13.2 A non-stationary, exploding AR(1) model

The AR(p) model

autoregressive model of order two, and will have the form:

Yt = φ1 Yt−1 + φ2 Yt−2 + ut (13.2)

Yt = φ1 Yt−1 + φ2 Yt−2 + · · · + φp Yt−p + ut (13.3)

or, using the summation symbol:

where (L)Yt is a polynomial function of Yt .

Stationarity in the AR(p) model

Properties of the AR models

E(Yt ) = E(Yt−1 ) = E(Yt+1 ) = 0

where Yt+1 = φYt + ut+1 . Substituting repeatedly for lagged Yt we have:

Cov(Xt , Zt ) = E{[Xt − E(Xt )][Zt − E(Zt )]} (13.13)

Cov(Yt , Yt−1 ) = E{[Yt − E(Yt )][Yt−1 − E(Yt−1 )]} (13.14)

Cov(Yt , Yt−1 ) = E{[Yt Yt−1 ] − [Yt E(Yt−1 )] − [E(Yt )Yt−1 ]

where E(Yt ) = E(Yt−1 ) = E(Yt+1 ) = 0. This leads to:

Cov(Yt , Yt−1 ) = E[(φYt−1 + ut )Yt−1 ]

We can easily show that:

Cov(Yt , Yt−2 ) = E(Yt Yt−2 )

= E(φ 2 Yt−2 Yt−2 )

Cov(Yt , Yt−k ) = φ k σY2 (13.17)

The autocorrelation function will be given by:

Cov(Yt , Yt−k ) φ k σY2

Moving average models

The MA(q) model

Yt = ut + θ1 ut−1 + θ2 ut−2 + · · · + θq ut−q (13.20)

which can be rewritten as:

or, using the lag operator:

Because any MA(q) process is, by deﬁnition, an average of q stationary white-noise

Using the lag operator, this can be rewritten as:

To understand this, consider the MA(1) process:

Lagging this relationship one period and solving for ut we have:

ut−1 = Yt−1 − θ ut−2

Substituting this into the original expression we have:

Yt = ut − θ (Yt−1 − θ ut−2 ) = ut − θ Yt−1 + θ 2 ut−2

Yt = ut − θ Yt−1 + θ 2 Yt−2 − θ 3 ut−3

are greater than 1 in absolute value.

Properties of the MA models

Var(Yt ) = Var(ut + θ ut−1 ) = σu2 + θ 2 σu2 = σu2 (1 + θ 2 ) (13.28)

The autocovariance will be given by:

Cov(Yt , Yt−1 ) = E[(ut + θ ut−1 )(ut−1 + θ ut−2 )] (13.29)

And since ut is serially uncorrelated it is easy to see that:

Cov(Yt , Yt−k ) = 0 for k > 1 (13.32)

Yt = φ1 Yt−1 + φ2 Yt−1 + · · · + φp Yt−p + ut

+ θ1 ut−1 + θ2 ut−2 + · · · + θq ut−q (13.34)

which can be rewritten, using the summations, as:

or, using the lag operator:

Integrated processes and the ARIMA models

Yt = Yt − Yt−1 (13.39)

Yt = Yt − Yt−1 (13.39)

Yt = 2 Yt = Yt − Yt−1 (13.40)

. || . || 1 0.963 0.963 71.464 0.000

. |* | . |* | 1 0.454 0.454 15.645 0.000