MODELLING VOLATILITY

IN FINANCIAL TIME SERIES:

AN INTRODUCTION TO ARCH
Econometrics C Lecture Note 7
Heino Bohn Nielsen
February 6, 2012
S
o far in this course, we have mainly discussed models for the conditional mean
and we have focussed particularly on the structure of time dependence. In
many applications within nance, however, there is a major interest also in
the conditional variance of a time series, vaguely interpreted as the risk of holding
certain assets. In this note we introduce a particular class of so-called autoregres-
sive conditional heteroskedasticity (ARCH) models for the conditional variance. In
2003 Robert Engle was awarded the Nobel prize for the ARCH model, and since its
introduction in Engle (1982) many interesting renements have been proposed. In
this note we focus on the main ideas, and, although interesting from a theoretical
and empirical point of view, we only briey mention some of the possible extensions.
To give you a feeling for the behavior of nancial data and the ARCH model we
emphasize an empirical example.
Outline
1 Changing Volatility in Financial Time Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2 A Misspecication Test for ARCH . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
3 The ARCH Model Dened. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
4 Generalized ARCH (GARCH) Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
5 Extensions to the Basic Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
6 Concluding Remarks. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1
1 Changing Volatility in Financial Time Series
Applications in nancial economics are often interested in both the mean and the variance
of investments, and sometimes in the entire return distribution. An intuitive reason is that
an investor typically faces a trade o between the return and risk.
One application where the mean-variance trade-o is obvious is within the area of
portfolio management, i.e. how to combine bonds, stocks, etc. in a portfolio, j. A simple
setup used in some theoretical work is the so-called mean variance utility function, where
the expected utility of the portfolio return, 1
j
, is dened as being
l(1
j
) = 1[1
j
]
c
2
\ (1
j
),
where c is the degree of risk aversion, measuring the explicit trade-o between the expected
return and the associated variance.
Another application where the whole distribution of returns is needed is for calculation
the so-called Value-at-Risk, VaR. The VaR is used as a stress test and measures the
maximum loss of a given portfolio calculated at a given probability level. In otheer words,
the 5% VaR of portfolio j is a value so that the probalitity of a larger loss is exactly 5%,
and it corresponds to the 5% quantile of the distribution of 1
j
.
A stylized fact for many nancial time series is a tendency for the variance, or volatility,
to be non-constant. In particular, as noted by Mandelbrot (1963):
...large changes tend to be followed by large changes, of either sign, and small
changes tend to be followed by small changes.
In terms of economics we may think of a large shock to the return at time t1, after which
there is a high probability of another large shock at time t. The shock has upset the market
and a period of larger uncertainty on the future direction of j
t
follows. This idea, often
referred to as volatility clustering, is just a reection of high and low market uncertainty.
The non-constant variance is known as heteroskedasticity, and this particular form of
time varying heteroskedasticity is known as autoregressive conditional heteroskedasticity
or ARCH. One main insight of the ARCH model is the distinction between conditional and
unconditional variance; where the conditional variance, i.e. the part of the variance that
cannot be predicted based on the information set, can be seen as a measure of risk. This is
parallel to the distinction in a linear regression between the conditional and unconditional
mean.
In many applications the autocorrelation of nancial returns is quite low. Note, how-
ever, that the ARCH eects will make the squared returns highly autocorrelated, and an
informal way to examine a data set for ARCH eects is therefore to look at the squared
returns and to calculate the autocorrelation function (ACF) for squared returns.
Example 1 (volatility of ibm stock returns): As an example, let j
t
denote the
percentage return from month to month on the IBMstock on the New York Stock Exchange
2
for the period 1926:1 to 1999:12. The returns are reported in Figure 1 (A). We note that
the variance is high is some periods, e.g. in the early 1930ties and in the 1990ties,
while movements are much smaller in other periods. Notice that the periods of large
variation entail both positive and negative returns, so whereas the mean seems by-and-
large constant, the variance is non-constant. We say that there are clear ARCH eects.
Figure 1 (B) depicts the squared returns for the IBM stock. In this graph the time
dependence is clearly visible reecting the clusters of high variance. Figure 1 (C) shows
the ACF for the returns, j
t
, and the squared returns, j
2
t
. Whereas the autocorrelations are
generally small for the returns they are much larger for squared returns: Again indicating
ARCH eects.
We could estimate the variance of returns for the whole sample using the normal
formula to estimate the unconditional variance,
b :
2
=
1
T
T
X
t=1
c
2
t
, (1)
where c
t
= j
t
b j is the deviation from the estimated mean of j
t
. To illustrate the
expected variation in c
t
we could draw approximate 95% condence bands as 1.96 b :.
These constant condence bands are reported in Figure 1 (D). Notice that due to the
non-constant variance, the constant bands are very poor measures of the uncertainty at a
given point in time, and the observations outside the band seem to cluster. As a measure
of the uncertainty at a given point in time, t, we could instead calculate the variance using
a small window of observations, e.g. the ve previous observations:
b :
2
t
=
1
5
5
X
i=1
c
2
ti
. (2)
This conditional variance measure from (2) is also depicted in graph (D), and seems to be
a better measure of the variance at a given point in time. The descriptive measure could
be rened by changing the weights in the moving average, and the class of ARCH models
we will introduce below is a way of formally embedding this kind of changing condence
bands into the statistical model.
2 A Misspecification Test for ARCH
First, let j
t
denote the variable of interest and we think of j
t
as a (stationary) time series
for e.g. an asset return or an interest rate. We consider the following linear model for the
conditional mean,
1[j
t
| r
t
] = r
0
t
0, t = 1, 2, ..., T
which corresponds to the linear regression model,
j
t
= r
0
t
0 +c
t
, (3)
3
1940 1960 1980 2000
25
0
25
(A) IBM stock returns, percent
1940 1960 1980 2000
250
500
750
(B) Squared returns
0 5 10 15 20
0.2
0.1
0.0
0.1
0.2
(C) Autocorrelation functions
ACF: Returns
ACF: Squared returns
1940 1960 1980 2000
50
25
0
25
(D) Measures of the volatility
Returns, deviation from mean
Unconditional volatility
Conditional volatility
Figure 1: Returns and squared returns for IBM stock.
with 1[c
t
| r
t
] = 0. In the literature of nance, there is a long history of seeing returns as
approximately uncorrelated over time, and it is dicult to predict the conditional mean
of an asset returns. This fact is often seen as a weak form of market eciency, because
predictability would typically allow easy speculative gains. It implies that the equation
in (3) is often dicult to specify, and r
0
t
0 only accounts for a small proportion of the
variation in j
t
.
In the presence of ARCH eects, OLS estimation of the equation in (3) is consistent
but inecient. The sucient condition for consistency of the OLS estimator in (3) is that
1[r
t
c
t
] = 0, and ARCH eects will not in general violate that. OLS is no-longer ecient,
however, and there exists a non-linear model that takes the ARCH eects into account,
and the estimator in this model has a smaller variance.
To test the hypothesis of no-ARCH eects, we can use the standard Breusch-Pagan LM
test for no-heteroskedasticity and apply it to this particular form of heteroskedasticity. If
the null hypothesis is no-ARCH eects up to order j, we consider the auxiliary regression
model
c
2
t
=
0
+
1
c
2
t1
+
2
c
2
t2
+... +
j
c
2
tj
+ error, (4)
where c
t
is the estimate residual from the regression of interest. The null hypothesis of no
ARCH is given by
H
0
:
1
=
2
= ... =
j
= 0,
4
so that the expected value of c
2
t
is a constant for all t. We can use the familiar LM statistic,

1C1
= T 1
2
,
which is asymptotically distributed as a
2
(j) if the null is true.
It is important to note that the ARCH test has also power against residual autocor-
relation. This is because autocorrelation in c
t
will imply autocorrelation in c
2
t
(while the
opposite is not true in general). Before the ARCH test is applied it is therefore important
always to test for no-autocorrelation rst. If the residual are not autocorrelated but the
squared residuals are, that is interpreted as an indication of ARCH eects.
Example 2 (misspecification test, ibm stock return): First consider a linear re-
gression for the conditional mean
j
t
= 1.149
(5.01)
+ 0.076
(2.26)
j
t1
+bc
t
.
The lagged dependent variable seems borderline signicant, indicating some predictability
of returns. The coecient of determination is low, however, 1
2
= 0.00574, meaning that
the statistical model explains under one percent of the variation in returns. There are no
signs of autocorrelation in the residuals, and a test for no-autocorrelation of order 1-2 is
accepted with a LM test statistic of 0.228, corresponding to a jvalue of 0.89 in a
2
(2).
To test for the presence of ARCH eects, we consider the auxiliary regression of squared
residuals on j = 5 lagged squared residuals:
bc
2
t
= 25.804
(6.74)
+ 0.153
(4.53)
bc
2
t1
+ 0.100
(2.93)
bc
2
t2
+ 0.048
(1.39)
bc
2
t3
+ 0.110
(3.18)
bc
2
t4
+ 0.017
(0.484)
bc
2
t5
+ error.
We note that many of the lags are signicant. The coecient of determination is 1
2
=
0.06939 and the LM test statistic for no-ARCH is given by

1C1
= T1
2
= 882 0.06939 = 61.2,
which is highly signicant in the asymptotic
2
(5) distribution. This is a formal indication
of ARCH eects.
3 The ARCH Model Defined
To analyze the conditional heteroskedasticity in more details we need a statistical model
that allows for ARCH eects. We therefore dene the conditional variance:
o
2
t
= 1[c
2
t
| I
t1
],
where I
t1
= {j
t1
, j
t2
, ..., r
t1
, r
t2
, ...} is the information set available at time t 1.
The class of ARCH models consists of an equation for the conditional mean (3) augmented
with an equation for the conditional variance, o
2
t
.
5
Engle (1982) suggests a statistical model for o
2
t
which follows directly for the descrip-
tive measure in (2). In particular he suggests the ARCH(p) model for the conditional
variance:
o
2
t
= c +c
1
c
2
t1
+c
2
c
2
t2
+... +c
j
c
2
tj
. (5)
Notice that the variance, o
2
t
, is a simple function of j lagged squared residuals. To ensure
a consistent model that generates a positive variance, we need to constrain the parameters,
c 0, c
i
0. The economic interpretation is straightforward: If there is a large shock
at time t 1, i.e. if c
2
t1
is high, then the variance of the following shocks will be large.
Graphically, the width of the condence bands depends on the magnitude of the past
shocks, and the ARCH(p) process has a memory of j periods.
A common way to write the full model is the following
j
t
= r
0
t
0 +c
t
c
t
= o
t
.
t
,
where o
t
is dened in (5) and .
t
is an IID error term with mean zero and unit variance. At
each point in time there is a new independent underlying shock to the system, .
t
, which is
scaled by o
t
so that the ARCH innovation have a conditional variance of 1[c
2
t
| I
t1
] = o
2
t
,
and the conditionally heteroskedastic shock c
t
drives the observed return process j
t
.
3.1 Interpretation
Another way to understand the model is to decompose c
2
t
into the conditional expectation
and the surprise in the squared innovations:
c
2
t
= 1[c
2
t
| I
t1
] +
t
= o
2
t
+
t
,
where it follows that 1[
t
| I
t1
] = 0, so that
t
is an uncorrelated (but not necessarily
homoskedastic) sequence. Inserting o
2
t
= c
2
t

t
into the ARCH(p) equation in (5) yields
the equation
c
2
t
= c +c
1
c
2
t1
+c
2
c
2
t2
+... +c
j
c
2
tj
+
t
, (6)
which shows that the squared innovation, c
2
t
, follows an AR(p) process. Note that (6) is
exactly the implication we use in the auxiliary regression (4) in the misspecication test.
Note that c is not the unconditional variance. We can illustrate by taking expectations
to obtain
1[c
2
t
] = c +c
1
1[c
2
t1
] +... +c
j
1[c
2
tj
],
which denes a constant unconditional variance of
o
2
=
c
1 c
1
... c
j
,
provided that the sum of the coecients is less than one,
P
j
i=1
c
i
< 1. Note that whereas
the ARCH model exhibit conditional heteroskedasticity, the ARCH process is uncondi-
tionally homoskedastic.
6
3.2 Maximum Likelihood Estimation
Before we continue let us briey discuss the estimation of ARCH models. To make the
notation simple we consider an ARCH(1) model with
j
t
= r
0
t
0 +c
t
o
2
t
= c +cc
2
t1
.
To perform a likelihood analysis we have to specify a distributional shape for c
t
. First
consider conditional normality:
c
t
= o
t
.
t
, .
t
(0, 1),
or alternatively that c
t
| I
t1
(0, o
2
t
). We can write the likelihood contribution as a
function of observed data as
1
t
(0, c, c | j
t
, r
t
, j
t1
, r
t1
, ...)
=
1
p
2o
2
t
exp

1
2
c
2
t
o
2
t

=
1
q
2(c +c(j
t1
r
0
t1
0)
2
)
exp

1
2
(j
t
r
0
t
0)
2
c +c(j
t1
r
0
t1
0)
2

,
and maximize the sum of the contributions with respect to the parameters 0, c and c.
The analytical analysis of the likelihood function is somewhat complicated and we cannot
solve the likelihood equations analytically. Instead we use numerical optimization to nd
the ML estimators.
In place of the normal distribution, other forms of distributions can be used. Many
applications use a more fat-tailed distribution to allow a larger proportion of extreme
observations. A particularly convenient solution is to use a student t() distribution
where the degree of freedom, , determines how fat the tails should be. Here can be
treated as a parameter in the likelihood function and estimated jointly with the remaining
parameters.
Example 3 (arch(p) model, ibm stock return): To illustrate the estimation of
ARCH(p) models we consider a model given by the two equations
j
t
= 0
0
+0
1
j
t1
+c
t
o
2
t
= c +c
1
c
2
t1
+c
2
c
2
t2
+... +c
j
c
2
tj
,
where we assume that the error term is normal, c
t
| I
t1
(0, o
2
t
). The estimation
results for dierent lag lengths, j = 1, 2, 3, 5, 7, are reported in Table 1. The results suggest
that the ARCH eects are clearly signicant. Unfortunately, it is quite hard to determine
the appropriate number of lags in the conditional variance. Many of the coecients have
similar magnitudes and comparable signicance. It is an observed weakness of the ARCH
model that it is hard to precisely pin down the shape of the memory structure, and the
estimated coecients are often relatively unstable between models.
7
ARCH(1) ARCH(2) ARCH(3) ARCH(5) ARCH(7)
0
0
1.121
(4.94)
1.181
(5.88)
1.196
(6.11)
1.198
(6.12)
1.194
(6.11)
0
1
0.113
(2.81)
0.116
(3.05)
0.110
(2.98)
0.102
(2.83)
0.102
(2.85)
c 36.838
(12.2)
30.733
(10.5)
27.260
(9.36)
24.838
(7.80)
21.780
(6.37)
c
1
0.175
(3.35)
0.156
(2.83)
0.155
(2.96)
0.134
(2.61)
0.130
(2.48)
c
2
0.157
(2.32)
0.123
(1.92)
0.098
(1.52)
0.100
(1.47)
c
3
0.118
(2.42)
0.100
(2.13)
0.102
(2.10)
c
4
0.060
(1.26)
0.025
(0.595)
c
5
0.055
(1.67)
0.051
(1.38)
c
6
0.054
(1.26)
c
7
0.055
(1.41)
P
j
i=1
c
i
0.175 0.313 0.396 0.446 0.517
log-lik 2929.19 2916.92 2912.09 2909.10 2904.70
SC 6.635 6.615 6.612 6.621 6.626
HQ 6.622 6.599 6.592 6.594 6.593
AIC 6.614 6.588 6.580 6.577 6.572
ARCH 1-5 test: [0.00] [0.08] [0.63] [0.98] [0.98]
Table 1: Estimation of ARCH(p) models for IBM stock returns.
4 Generalized ARCH (GARCH) Models
To resolve the problem with many lags in ARCH models Bollerslev (1986) and Taylor
(1986) suggested a generalized version of the ARCH model which economize more with
the number of parameters. The simplest case is the very popular GARCH(1,1) model
dened by the equation
o
2
t
= c +cc
2
t1
+,o
2
t1
,
where the lagged variance is included along with the squared innovation. A simple inter-
pretation is that the lagged dependent variable allows for a more smooth development in
the variance and a longer memory without including many parameters.
To understand the relationship between ARCH and GARCH models we can again use
the denition o
2
t
= c
2
t

t
, and obtain that
o
2
t
= c +cc
2
t1
+,o
2
t1
c
2
t

t
= c +cc
2
t1
+,

c
2
t1

t1

c
2
t
= c + (c +,)c
2
t1
+
t
,
t1
.
8
This suggests that the GARCH(1,1) implies an ARMA(1,1) structure for the squared
innovation. We recall that that an ARMA model can be seen as a restricted and parsi-
monious representation of an innite AR model, and we can think of the GARCH model
as a restricted innite ARCH model. By repeated substitution we can write the GARCH
model as
o
2
t
= c +cc
2
t1
+,o
2
t1
= c +cc
2
t1
+,(c +cc
2
t2
+,o
2
t2
)
= c(1 +,) +cc
2
t1
+c,c
2
t2
+,
2
o
2
t2
.
.
.
= c(1 +, +,
2
+,
3
+...) +cc
2
t1
+c,c
2
t2
+c,
2
c
2
t2
+...
=
c
1 ,
+c

X
)=1
,
)1
c
2
t)
,
where we assume that the process started in the innite past.
Also for the GARCH model we need o
2
t
to be non-negative and constrain the coe-
cients to be non-negative. By looking at the ARMA representation we can also read of
the condition under which c
2
t
is covariance stationary, namely that c+, < 1. In this case
the unconditional variance is given by
o
2
= 1[c
2
t
] =
c
1 c ,
.
The GARCH(1,1) model is extremely popular in applied work, but it can of cause be
generalized with more lags as the GARCH(p,q) model:
o
2
t
= c +
j
X
)=1
c
)
c
2
t)
+
q
X
)=1
,
)
o
2
t)
.
The condition for covariance stationarity is
P
j
)=1
c
)
+
P
q
)=1
,
)
< 1.
4.1 Volatility Forecasts
One important application of ARCH and GARCH models is prediction of future volatility.
We use o
2
T+I|T
= 1[c
2
T+I
| I
T
] to denote the forecast of volatility at time T +/ given the
information set at time T. To construct the forecast from an ARCH(1) we rst use the
fact c = o
2
(1 c) to rewrite the equation in deviations from the mean
o
2
t
= c +cc
2
t1
o
2
t
= o
2
(1 c) +cc
2
t1
o
2
t
o
2
= c

c
2
t1
o
2

.
To forecast volatility for T + 1 we nd the best prediction
o
2
T+1|T
= 1[c
2
T+1
| I
T
] = 1

o
2
+c

c
2
T
o
2

| I
T

= o
2
+c

c
2
T
o
2

,
9
where we have used that c
T
is in the information set at time T, so that 1

c
2
T
| I
T

= c
2
T
.
Forecasts for the next periods are constructed as the recursion
o
2
T+2|T
= 1[c
2
T+2
| I
T
]
= 1

o
2
+c

c
2
T+1
o
2

| I
T

= o
2
+c

c
2
T+1
| I
T

o
2

= o
2
+c

o
2
T+1|T
o
2

,
o
2
T+3|T
= 1[c
2
T+3
| I
T
]
= 1

o
2
+c

c
2
T+2
o
2

| I
T

= o
2
+c

c
2
T+2
| I
T

o
2

= o
2
+c

o
2
T+2|T
o
2

,
etc. Note that the forecast will produce an exponential convergence towards the uncon-
ditional variance, o
2
.
For the GARCH(1,1) model we use a similar recursion. First write the model in
deviations from mean
o
2
t
= c +cc
2
t1
+,o
2
t1
o
2
t
= o
2
(1 c ,) +cc
2
t1
+,o
2
t1
o
2
t
o
2
= c

c
2
t1
o
2

+,

o
2
t1
o
2

.
The rst period forecast is given by
o
2
T+1|T
= 1[c
2
T+1
| I
T
]
= 1

o
2
+c

c
2
T
o
2

+,

o
2
T
o
2

| I
T

= o
2
+c

c
2
T
o
2

+,

o
2
T
o
2

,
because c
T
is in the information set at time T and o
2
T
can be calculated from the infor-
mation set. The next period forecast is
o
2
T+2|T
= 1[c
2
T+2
| I
T
]
= 1

o
2
+c

c
2
T+1
o
2

+,

o
2
T+1
o
2

| I
T

= o
2
+c

c
2
T+1
| I
T

o
2

+,

o
2
T+1
| I
T

o
2

= o
2
+c

o
2
T+1|T
o
2

+,

o
2
T+1|T
o
2

= o
2
+ (c +,)

o
2
T+1|T
o
2

,
where we have used that 1

c
2
T+1
| I
T

= o
2
T+1|T
and 1

o
2
T+1
| I
T

= o
2
T+1|T
. For longer
horizons we nd similarly that
o
2
T+I|T
= o
2
+ (c +,)

o
2
T+I1|T
o
2

,
which is an exponential convergence with speed c +,.
10
Example 4 (garch models, ibm stock return): To illustrate, we estimate the
AR(1)-GARCH(1,1), i.e.
j
t
= 0
0
+0
1
j
t1
+c
t
o
2
t
= c +cc
2
t1
+,o
2
t1
.
First column reports the ML estimates obtained by assuming normal innovations,
c
o

(0, 1). We note that the lagged variance is large and signicant, and c + , = 0.934
is quite close to one. The fact that the estimated GARCH models have roots close to
unity is a common observation which is discussed in more details in Box 1. To illustrate
the results we report in Figure 2 (A) the estimated residuals as well as the condence
bands constructed as 1.96 b o
t
. In Figure 2 (B) we compare the estimated standard
deviation from the GARCH(1,1) with the one obtained in the ARCH(5) model. They
share the same general tendencies, but the GARCH(1,1) model is much less erratic and
have a longer memory. The dierence becomes even more pronounced if we compare the
forecasts in Figure 2 (C). The GARCH model suggests a quite high variance for a number
of years while the ARCH model converges very fast towards the unconditional variance;
this reects the dierence in convergence speed between
P
5
i=1
c
i
= 0.45 for the ARCH(5)
and c + , = 0.93 for the GARCH(1,1). Also note that the volatility forecasts converge
towards the unconditional variance, which is an average of the low and high variance
periods in the past, and not towards the volatility of the low variance periods, although
they are often the most common.
We note that the null of normality is strongly rejected (jvalue of 0.00), due to a large
proportion of extreme observations. One renement is to estimate the model with the
assumption of a more general error distribution, and second column reports the results
for a studentt distribution,
c

t(i). The estimated degrees of freedom is = 8.4

which gives a much more fat-tailed distribution than the normal. It holds that kurtosis
for the normal distribution is K = 3 while it is K = 3 +
6
4
= 3 +
6
8.44
= 4.4 for the
t(8.4)distribution. The latter is also close to the kurtosis of the residuals from the basic
GARCH(1,1) model. There are also small dierences in the estimated parameters, but in
a graph the conditional standard errors are almost identical (not shown).
5 Extensions to the Basic Model
Thee are many possible extensions of these basic model. Here we only consider two
important cases.
5.1 Asymmetric ARCH Models and the News Impact Curve
The basic models have the feature that the sign of a shock does not matter and positive
and negative shocks have the same eects on the conditional variance. The eect of a
11
Box 1: Persistence in Volatility and the IGARCH
It is a stylized feature of estimated GARCH models that the sum of the coecients is close to
one,

X
=1
c

+

X
=1

1,
known as an integrated GARCH or IGARCH model. In this case the squared residuals follow
a unit root ARMA model, and the unit root has strange implications for the behavior of the
model. First it violates what we called the stationarity condition. Secondly it implies that the
unconditional variance is not dened, and, nally, a forecasts for the conditional variance will
replicate the forecast of a random walk (with drift), i.e. a linear trend!
The latter implication may be a valid characterization of the data, but it is hard to maintain as
a behavioral model for investors. As a consequence the IGARCH phenomenon is often discussed
as a sign of misspecication of the class of GARCH models. One problem could be that there
are structural shifts in the unconditional variance, i.e. in the constant term of the GARCH
equation. If this is not modelled it will bias the roots of the ARMA model towards unity, as
we have seen for the analysis of unit root data.
The asymptotic analysis of the IGARCH model is given in Nelson (1990). He demonstrates
that the IGARCH model is actually strictly stationary, but the variance is innite (and is not
weakly stationary!). A result is that although the analysis looks strange, a test for a unit root
in the variance, i.e. a likelihood ratio test for the IGARCH model against the GARCH follows
a standard
2
distribution.
To illustrate we have replicated the estimates in Table 2 imposing the IGARCH restriction
c+, = 1. Based on a likelihood ratio test, the IGARCH model is clearly rejected in this case.
For the model with normal errors we get a test statistic of
11(c +, = 1) = 2 (2908.03 + 2901.02) = 14.02,
which is clearly signicant in a
2
(1).
An additional complication in the near-integrated GARCH case is that the unconditional
variance is poorly estimated. A simple estimator of the unconditional variance is :
2
from (1).
The unconditional variance in the GARCH model is o
2
= c,(1 c ,) and for c + , 1
the denominator is close to zero and the estimate of the constant term, c, also becomes very
small. As a consequence the estimator o
2
= c,(1 c

,) can be very poor, and in the limiting

case, c+, 1, it does not exist. In practice, the estimate o
2
can be far from the estimate :
2
.
Since the forecasts for the conditional variance describe a convergence back to the unconditional
variance, a poor estimate of o
2
may render the forecasts useless. A simple solution is to insert
c = o
2
(1 c ,) and write the GARCH equation as
o
2

= o
2
(1 c ,) +cc
2
1
+,o
2
1
.
Instead of estimating

o
2
, c, ,

(which is equivalent to estimating (c, c, ,)), we can x o

2
to
some robustly estimated value (e.g. :
2
) and only estimate the parameters (c, ,) in the GARCH
model. This is known as variance targeting.
12
GARCH(1,1) GARCH(1,1) IGARCH(1,1) IGARCH(1,1)
c
o

(0, 1)
c
o

t(i)
c
o

(0, 1)
c
o

t(i)
0
0
1.179
(6.00)
1.226
(6.17)
1.292
(6.001)
1.312
(6.411)
0
1
0.104
(2.86)
0.071
(1.93)
0.105
(3.123)
0.070
(1.974)
c 2.932
(2.10)
2.535
(2.60)
0.622
(0.910)
0.822
(1.349)
c 0.097
(3.12)
0.093
(3.88)
0.094
(1.515)
0.107
(2.412)
, 0.837
(16.4)
0.850
(22.8)
0.906
()
0.893
()
i 8.409
(4.31)
6.708
(4.267)
c +, 0.934 0.944 1.000 1.000
log-lik -2901.02 -2890.04 -2908.03 -2893.96
SC 6.580 6.562 6.588 6.564
HQ 6.563 6.542 6.574 6.547
AIC 6.553 6.530 6.566 6.537
ARCH 1-5 test: [0.91] [0.91] [0.75] [0.83]
Normality [0.00] ... [0.00] ...
Table 2: Estimation of GARCH(p,q) models.
shock c
t1
on the conditional variance, o
2
t
, is known as the news impact curve; and for
the basic model the news impact curve is obviously symmetric.
In some cases it is reasonable to believe that negative shocks have a dierent impact
from positive shocks, and the models can easily be adapted to this situation. In a famous
paper, Glosten, Jagannathan, and Runkle (1993) suggest to model the asymmetric eects
in the following simple way
o
2
t
= c +cc
2
t1
+ic
2
t1
1(c
t1
) +,o
2
t1
,
where
1(r) =
(
1 if r < 0
0 Otherwise
is the indicator function. We follow PcGive and call this a threshold asymmetric model. We
note that the news impact curve is now asymmetric, with an impact of squared residuals
of c for positive shocks and c +i for negative values.
Example 5 (threshold model, imb stock return): For the IBM stock return we
estimate a threshold model given by
j
t
= 1.114
(5.35)
+ 0.107
(3.07)
j
t1
+c
t
o
2
t
= 3.228
(2.20)
+ 0.064
(2.48)
c
2
t1
+ 0.063
(1.16)
c
2
t1
1(c
t1
) + 0.831
(16.2)
o
2
t1
13
1940 1960 1980 2000
25
0
25
(A) Residual and conditional confidence bands
Residual
95% confidence band
1940 1960 1980 2000
5
10
15
(B) Comparison of Conditional Standard Deviation (B) Comparison of Conditional Standard Deviation
Unconditional standard deviation
GARCH(1,1)
ARCH(5)
1990 1995 2000
5
10
15
(C) Comparison of Forecasts (C) Comparison of Forecasts
Unconditional standard deviation
GARCH(1,1)
ARCH(5)
30 20 10 0 10 20
50
100
(D) News Impact Curve
Basic GARCH model
Asymmetric model
Threshold model
Figure 2: Results for the analysis of the IBM stock returns.
with tvalues in parentheses and a log-likelihood of 2899.308. At face value the estimates
suggest an asymmetric eect, with impacts of c = 0.064 and c+i = 0.064+0.063 = 0.127
for positive and negative shocks respectively. This is compared to the basic symmetric
news impact curve in Figure 2 (D), and suggests that the market is more upset by negative
than positive shocks. The threshold parameter is not statistically signicant, however,
with a ttest statistic of t
i=0
= 1.16. We can also test the signicant by a likelihood ratio
test, where the statistic 11(i = 0) = 2 (2899.308 + 2901.02) = 3.424 corresponds to a
jvalue of 0.064.
An alternative form of asymmetry, suggested by Engle and Ng (1993), is
o
2
t
= c +c(c
t1
)
2
+,o
2
t1
,
where the news impact curve has the same slope for small and large values, but zero is
no-longer the neutral shock. We follow PcGive and refer to this as the asymmetric model.
Example 6 (asymmetric model, ibm stock return): To illustrate the asymmetric
model, we consider the IBM stock return and obtain
j
t
= 1.125
(5.32)
+ 0.107
(3.01)
j
t1
+c
t
o
2
t
= 2.988
(2.30)
+ 0.096
(3.58)
(c
t1
1.332
(0.298)
)
2
+ 0.832
(17.0)
o
2
t1
.
14
with tvalues in parentheses. Based on this specication there is very little evidence for
asymmetry. Again we have depicted the news impact curve in Figure 2 (D) and note that
the eect is minor. The log-likelihood of this specication is 2899.928.
The two models can easily be combined into a more general specication and a fully
asymmetric GARCH(p,q) model could have the form
o
2
t
= c +
j
X
i=1
c
i
(c
ti
)
2
+ic
2
t1
1(c
t1
) +
q
X
i=1
,
i
o
2
ti
,
which allows for a very elaborate specication of the news impact curve.
5.2 ARCH in Mean
Most theoretical models in nance suggest a trade-o between risk and return; and within
a given time period investors require a larger expected return from an investment which
is riskier, often referred to as a risk premium. Similar eects may also work over time, so
that investors require a higher return in time periods where the volatility is asserted to
be higher. There may also be eects in the opposite direction, see Glosten, Jagannathan,
and Runkle (1993) for a discussion.
The argument above suggests that there should be a relationship between the condi-
tional variance and the conditional mean. One way to test the hypothesis is to let the
conditional mean depend on the variance, o
2
t
. It is not obvious how the relationship should
be, but some suggested examples replaces the simple equation for the conditional mean
(3) with
j
t
= r
0
t
0 +co
2
t
+c
t
j
t
= r
0
t
0 +co
t
+c
t
j
t
= r
0
t
0 +c log

o
2
t

+c
t
.
A positive estimate for c corresponds to positive risk premium.
Example 7 (garch-in-mean, ibm stock return): We estimate the GARCH-in-mean
model for the IBM stock return and obtain
j
t
= 0.948
(1.49)
+ 0.104
(2.92)
j
t1
+ 0.0060
(0.377)
o
2
t
+c
t
o
2
t
= 2.894
(2.24)
+ 0.096
(3.26)
c
2
t1
+ 0.839
(16.9)
o
2
t1
.
The log-likelihood of this specication is 2900.930. The estimated GARCH-in-mean
eect is very small and certainly not statistically signicant. This seems to suggest that
the return has not been driven by changes in the variance.
15
6 Concluding Remarks
This note has introduced the main ideas of ARCH and GARCH models and some exten-
sions. There is an enormous literature in this eld and the number of specic ARCH-type
models is exploding. Excellent reviews of the early ARCH literature are given in Boller-
slev, Chou, and Kroner (1992), Bollerslev, Engle, and Nelson (1992), and Bera and Higgins
(1995); and they also contain many references to specic models and extensions.
References
Bera, A. K., and M. L. Higgins (1995): On ARCH Models: Properties, Estimation
and Testing, in Surveys in Econometrics, chapter 8, pp. 215272. Blackwell, Oxford.
Bollerslev, T. (1986): Generalized Autoregressive Conditional Heteroskedasticity,
Journal of Econometrics, 31(3), 307327.
Bollerslev, T., R. Y. Chou, and K. F. Kroner (1992): ARCH Modelling in
Finance - A Review of the Theory and Empirical Evidence, Journal of Econometrics,
52, 559.
Bollerslev, T., R. F. Engle, and D. B. Nelson (1992): ARCH Models, in Hand-
book of Econometrics, Volume IV, ed. by R. Engle, and D. McFadden, chapter 49, pp.
29593038. North-Holland, Amsterdam.
Engle, R. F. (1982): Autoregressive Conditional Heteroscedasticity with Estimates of
the Variance of United Kingdom Ination, Econometrica, 50(4), 9871008.
Engle, R. F., and V. Ng (1993): Measuring and Testing the Impact og News on
Volatility, Journal of Finance, 48, 17491777.
Glosten, L., R. Jagannathan, and D. Runkle (1993): Relationship Between the
Expected Value and the Volatility of the Nominal Excess Return on Stocks, Journal
of Finance, 48(5), 17791802.
Mandelbrot, B. (1963): The Variation of Certain Speculative Prices, The Journal of
Business, 36(4), 394419.
Nelson, D. B. (1990): Stationarity and Persistence in the GARCH(1,1) Model, Econo-
metric Theory, 6(3), 318334.
Taylor, S. J. (1986): Modelling Financial Time Series. Wiley, New York.
16