UiMlvtKbi.

, OF
ILLINOIS LIBRARY
AT URBANA-CHA:
BOOKSTACKS
 \J> >*-t £->

CENTRAL CIRCULATION BOOKSTACKS

The person charging this material is re-
sponsible for its renewal or its return to
the library from which it was borrowed
on or before the Latest Date stamped
below. You may be charged a minimum
fee of $75.00 for each lost book.
Theft, mutilation, and underlining of books art reasons
for disciplinary action and may result In dismissal from
the University.
TO RENEW CALL TELEPHONE CENTER, 333-8400
UNIVERSITY OF ILLINOIS LIBRARY AT URBANA-CHAMPAIGN

APR 2 9 1396
AUG 3 1998

When renewing by phone, write new due date below

previous due date, L162
 Digitized by the Internet Archive
in 2011 with funding from
University of Illinois Urbana-Champaign

[Link]
 BEBR
FACULTY WORKING
PAPER NO. 90-1638

A Test for Conditional Hetcroskedasticity

In Time Series Models

Anil K. Bera
M. L. Higgins

y or th
e

College of Commerce and Business Administration

Bureau of Economic and Business Research
University of Illinois Urbana-Champaign
 BEBR

FACULTY WORKING PAPER NO. 90-1638

College of Commerce and Business Administration

University of Illinois at Urbana- Champaign

March 1990

A Test for Conditional Heteroskedasticity

in Time Series Models

A. K. Bera
University of Illinois, Champaign, IL 61820

and

University of Western Ontario, London, ONT N6A 5C2 , Canada

M. L. Higgins
University of Wisconsin, Milwaukee, WI 53201
 .

ABSTRACT

While testing for conditional heteroskedasticity and nonlinearity the power of

,

the test in general depends on the functional forms of conditional

heteroskedasticity and nonlinearity that are allowed under the alternative
hypothesis. In this paper, we suggest a test for conditional
heteroskedasticity/nonlinearity with the nonlinear autoregressive conditional
heteroskedasticity (NARCH) model of Higgins and Bera (1989) as the nonlinear ARCH
parameter is not identified under the null hypothesis. To resolve this problem,
we apply the procedure recently proposed by Davies (1987) Power and size of the
.

suggested test are investigated through simulation and an empirical application

of testing for ARCH in exchange rates is also discussed.

Keywords. ARCH; bilinear; Davies' test; NARCH; Nonlinearity; Nonlinear time

series models
 1. INTRODUCTION

Autoregressive conditional heteroskedasticity (ARCH) , introduced by

Engle (1982) , is frequently used to model the changing volatility of economic

time series. Such applications of the ARCH model can be found in Weiss

(1984) and papers surveyed in Engle and Bollerslev (1986) . When testing for

conditional heteroskedasticity, the form of the test statistic, and hence the

power of the test, in general depends on the functional form of the

conditional variance specified under the alternative hypothesis. In this

paper, we suggest a test in which the alternative is the nonlinear ARCH

(NARCH) model proposed by Higgins and Bera (1989b) . Consider the dynamic

linear regression model

y t = x't /3 + e
t
(t = 1.....T) (1.1)

where x is a vector of k predetermined variables which may include lagged

values of the dependent variable y . Let $ denote the information set at

time t which includes current and all lagged

bb values of e . The error e is
t t

generated by the ARCH process

- N(0
<J* t -i 'V

where h is a function of the elements of $ Engle (1982) proposes

several functional forms for h but emphasizes the linear ARCH model

h Q + + ••• +
t - o Vt-i Vt- P (1 - 2)
for its analytic convenience and ease of interpretation. Other functional

forms, however, have been found to be useful [See Engle and Bollerslev

(1986), Higgins and Bera (1989b), and Nelson (1989)]. For example, the

linear ARCH model (1.2) requires a > and a. > (i = l,...,p) to insure

that the conditional variance is strictly positive. Geweke (1986) and

Pantula (1986) suggest the logarithmic specification

log(h ) = a + o log(e ) + ... + a log(e ) (1.3)

t Q 1 t_1

for which the conditional variance is positive for all parameter values.

The Lagrange multiplier (LM) or score test principle provides an easily

computed test for the presence of linear ARCH. Engle (1982) shows that the

LM statistic for the null hypothesis H *

a = . . . = a = in (1.2) is

2 2
equivalent to T-R , where R is the squared uncentered multiple correlation
~2 -2
coefficient of the regression of e on an intercept and e . (i = 1, . .
. ,p)

and the e 's are the least squares residuals of (1.1). In general, however,

the form of the LM statistic depends on the functional form of the ARCH

process assumed under the alternative hypothesis. Conducting the above test

for ARCH when the true alternative is, for example, the logarithmic model

(1.3), may result in a significant loss of power. Furthermore, there is a

general view that linear ARCH models do not provide a rich enough class of

nonlinearities , and this necessitates a need for a more flexible parametric

specification for the conditional heteroskedas'ticity [see Pagan and Wickens

(1989, p. 983)]. In this paper we propose a test for ARCH in which the
alternative is the nonlinear ARCH (NARCH) model suggested by Higgins and Bera

(1989b). The NARCH model of order p, NARCH (p ) , specifies the conditional

variance function

5ll/5
h
t
=
V*
2
>
5
+
Vv/ + --- +
VVp> (1.4)

with the parameter space restricted to

1
a > 0; 4>. > 0, (i = 1 p); S > 0; £ <f>
= 1
1
i=0

Rearranging (1.4)

— ' *o—T— +
h 1
+ • •
+ *
P s

the NARCH model is seen to be a Box-Cox power transformation of the terms of

the linear ARCH model (1.2). The Box-Cox transformation is widely used in

the selection of functional form of the mean of a regression model. In the

present context, the NARCH model encompasses many of the functional forms

used for ARCH. For example, when 5=1 the model is identical to the linear

ARCH model (1.2). As S approaches from above, the model is equivalent to

Geweke's logarithmic model (1.3). Higgins and Bera (1989b) discuss other

ARCH specifications which NARCH encompasses. Furthermore, by adding Box-Cox

transformations of lagged values of h to the right hand side of (1.4), the

model can easily be generalized to include the GARCH model of Bollerslev

(1986, 1988).
 The null hypothesis
yr to be tested is H
o
: i.
1
=...= i =0. When these
p
conditions are imposed, the conditional variance function (1.4) reduces to a

constant and the model becomes the standard normal regression model. It is

immediately noticed, however, that when the null hypothesis is imposed, 5

drops out of the conditional variance function h . In other words, the

nuisance parameter 5 is identified only under the alternative hypothesis. It

can be shown that under H_, the information matrix is singular; thus

invalidating the standard formulation of the LM test. Watson and Engle

(1985) encounter the same problem in testing the constancy of a regression

coefficient against the alternative that the parameter follows a first order

autoregressive process. They consider the varying coefficient model

y
J = x'7 + z 8 + e
t t' r t t

where x is a vector of exogenous variables, 7 is vector of fixed

parameters,
K z is an exogenous
& scalar and e is a random disturbance. The
t t

stochastic parameter 8 is generated by

(B
t
- fi) = *<0
t-1
- ~P) + ut M < 1.

where <f>
and 8 are fixed parameters and u is a random disturbance with
2
variance q. Since the unconditional variance of 8 is q/(l-$ ), constancy of

the parameter 8 can be examined by testing q = 0. When q = 0, however, <f>

is

not identified. In order to proceed, they follow the suggestion of Davies

(1977) and base a test on Roy's union- intersection principle. Like Watson

and Engle (1985) , we follow Davies (1977) , and in addition make use of
 .

approximations provided by Davies (1987) to simplify the computation of

p -values for the test. In Section 2, we briefly review the work of Davies

(1977, 1987) and discuss its application to our testing problem. In Section

3 we derive the LM test required to implement Davies' procedure and in

Section 4 we report Monte Carlo results on the finite sample null

distribution and power of the proposed test. To illustrate the usefulness of

the test, in Section 5 we present an application of testing for nonlinear

ARCH in foreign exchange rates. Section 6 contains a few concluding remarks.

2. DAVIES' TEST

Davies (1977) considers a situation in which the density of the sample

depends on two parameters a and 8. It is desired to test the hypothesis a =

against the alternative a > 0. It is assumed that, for a given value of 8,

an appropriate Gaussian test statistic Z is available to test a = 0. When

the null hypothesis is true, however, it is assumed that the model is free of

8. In such a situation, in which the nuisance parameter 8 is not identified

under the null hypothesis, the asymptotic distribution theory of Z is

invalid. Since the distribution of Z is correct for any arbitrarily assigned

value of 8, Davies appeals to the "union- intersection principle" of Roy

(1953) and suggests basing the test on a critical region of the form

{sup Z(0) > c}

8
 .

Although the distribution of this test is unknown, Davies provides

approximations for computing the p-values. The approximations require

performing numerical integration of the continuous time autocorrelation

function of Z(9) . Davies (1987) extends his results to statistics which are
2
asymptotically x an d provides a much simpler approximation for the

p-values. It is this approximation which we make use of below.

To apply the results of Davies (1987) , we first fix 8 at an arbitrary

value 8 and derive a test for A>- = . . . = 6 = 0. In Section 3. we show that

1 p
given 8 = 8, the LM test statistic is

* 2
S(5 ) = T-R

2
where R is the squared multiple correlation coefficient of the regression of
~2
e on an intercept and

5 (i=l,...,p),
8

where e is the least squares residual. The actual test statistic is defined

as

S = sup S(5 ) (2.1)

8

However, unlike S(5 ), S will not have an asymptotic x distribution

under the null hypothesis. Of course, it will be very difficult to find the

exact critical values or p-values for S. Davies (1987) suggests an upper

bound of the p-values which is described below.

For each value of 5 G A C R , we can express S(5) as

S(5) = I Z AS) = Z'(5)Z(5)

i-1

where Z(<5) = (Z (5), Z_(<5), ..., Z (5))' is a p x 1 vector. Under certain

regularity conditions and the null hypothesis H •

4> = ... <f>
= 0, Z.(5)'s
p l

are asymptotically i.i.d. N(0,1). Define Y(5) = dZ{S)/dS and denote

Var[Y(5)] = B(5) and Cov[Z(5) Y(5)

the eigenvalues of B(5) -

,

A'(5)A(5) and let

] = A(5). Let XAS),

r}(6)
11
- N(0,A),
X n (5) ,

where A =
..., A
p
(5) be

diagCA^S), A (5), .... A (5)).

2

Under the above setup, Davies (1987, p. 35) shows that

Pr[{Sup S(5); <5 G A) > u] < Pr(v > u) + V>(5)d5

P

where

-u/2 (p-l ) /2
1/Z
1/2 e u
"
tf(S) = E[r
L
'(&)r (g)]J
7 ?
-

74 (2.2)
l/2
' oPp//2'|(
, 2 P+ l)/2

As proved in Davies (1987, Theorem A.l)

xp(6)dS

is the expected number of upcrossing of the level u by S(5) for 8 e A. This

can also be viewed as the correction factor to the standard x 2 p-value due to

the scanning across a range of values of 6 e A. Theorem A. 2 of Davies (1987)

further shows that

1/2
X
7T "|(p+l)/2
1/2 1/2
E[r?'(5)r,(5)] = E[| dS (6)/d6 |]
.
(2.3)
[?72

Combining (2.2) and (2.3), the upper bound of the significance level is given

by

-u/2 (p-l)/2
,1/2
?r(xl > u) + '
E[|3S '
(S)/d6\]dS. (2.4)
P p/2
2 fp72 A

Davies proposes to estimate

l,2
E[\dS (5)/d5\]dS
J
A

from the total variation

1/2
V = \dS (5)/d6\d8

1/2 1/2 1/2 1/2

= |S C5 ) - S (L)| + |S (5 ) - S (5 )| + ...
1 2 1

1/2 1/2
+ |S (U) - S (5 (2.5)
M )|

where L and U are the lower and upper bounds for 5 and 5, , 5~. .... 6 W are
1 2 M
1/2
the turning points of S (5). Therefore, from (2.4), the significance level

of our test based on S will be approximately

-S/2 (p-l)/2
Pr(;/ > S) + V-^— .
(2.6)
P p/2
2 [p72
Although (2.6) is only an approximation, we expect it to perform better than

basing the test on just the first term in (2.6). In the second term, one

2
part is essentially the x density function and the other part, V, reflects
1/2
the variation in S (5) over values of 5 corresponding to different

alternative hypotheses. Davies (1987) presents numerical results which show

that this type of approximation performs very well.

Here we should note that the set A need not coincide with the

theoretical range for 5; it could be any subset of that range. The only

constraint is that the same set should be used for maximizing S(5) in (2.1)

and in calculating V in (2.5). Also, Davies (1977, p. 253) mentions that for

the the procedure to be useful, S(6) cannot have spurious peaks. To see that

spurious peaks of S(5) is not very likely in our case, let us write S(<5) as

_ 2 g'WCW'WrVg
p
T-R - T ^_

-2
where £ is the vector of e and W includes an unit vector and Box-Cox
-2 -2 -2
transformations of t ., f e . Therefore, this is a standard
t-1 t-/ t-p
Box-Cox regression with transformation only in the non-constant independent

variables. Later, in the simulation study, we present graphs of realizations

of S(5) under both the null and alternative hypotheses which indicate that

S(5) is in fact very smooth.

To see the behavior of the test under H and the alternative hypothesis

H : at least one . * 0, let us first note that

A <f>
l

P ' fi W'W
plim -=- < co, plim —=- - a finite non-null matrix
T—>oo T^co
 10

p im
and under H rt that W'£/T = for any value of 8. While under the
t—>«
p im
alternative hypothesis H , W'£ 5* for any 6. Therefore, the test is

consistent. However, we cannot claim that our test has any optimality

property. For a weak optimality property of this kind of test, see Davies

(1977, p. 252).

3. LM TEST FOR FIXED 5

We now derive the LM test with 5 fixed at a pre-assigned value 5. The

conditional variance function (1.4) becomes

*iV«
h = r/2.0 ,.2 .0
t p t-p

Let 0' = (<£.,,..., ), i/' = (a ,0') and 9' = (B',v'). The log- likelihood for
1 P
the NARCH regression model can be written, omitting a constant,

£(&) = -\ I log(h ) -
I 2h (3.1)
t

where Che summations are over t. Higgins and Bera (1989b) show chat the

information matrix is block diagonal between the regression parameters /3 and

the variance parameters u . Furthermore, since H n does not impose

restrictions on j3 , the LM test reduces to [see e.g. Bruesch and Pagan (1980)]

1
LM = d(?)'I(?)" d(?) (3.2)
V 1/1/ V
 11

where d(9) and 1(0) are the score function and information matrix with
u uu
respect to the variance parameters and "~" denotes quantities evaluated at

the restricted maximum likelihood estimators (MLE's). Differentiating (3.1)

with respect to the variance parameters, the score function is

3h
d£
du L 2h
1
du
t
4"
and the hessian of the log- likelihood is

r 2
8h dh
d 1 y
L
1_
e
t __t _t + y
*•
€
t
- 1
1
3
\
)udu' ''

, 2*h 'du ' du' du' 2h 8u

h
t

The information matrix with respect to the variance parameters is then given

by

(
2
a £
uu dudu'

2
a i
dudu t-1

>
. E(e ) 3h ati E(e*) ' . ah
i t
a^' 2h du
t
t

C. ah ah 1
t
ilH l
2 du du'
t

which can be consistently estimated by

 12

«g i
ah
J
-I
2 L du h 3i/'
t

Therefore, the LM statistic (3.2) is

-l
r-2
ah. € 3h 3h aK
LM = £ I y
A du - 1 (3.3)
3i/ a j/ du' -2
V
a

-2
where e is the least squares residual and a is the usual MLE of the

variance of the error in the standard normal linear regression model. Now
-2 -2
define f to be a column vector whose elements are (e /a ) - 1 (t = 1 p)

and let z = ah /du and Z' = (z . ..,z ). The LM statistic (3.3) can then be

expressed in matrix form' as

1
LM = -•f'Z(Z'Z)" Z'f
2

which is 1/2 the regression sums of squares from the regression of f on Z.

p
Furthermore, since under the null hypothesis f'f/T = 2 and the

arithmetic mean of the elements of f is , an asymptotically equivalent form

of the test statistic is

IK-8(**) . T fz(z;z)
,
rt _ T R2.

where R is the squared multiple correlation coefficient from the regression

of f on Z. The elements of z are easily shown to be

ah
= 1
8o
 13

and

*
ah (e /a ) •
1
t -2 t_ i
= a (i - 1.....P)
d4>. *
i >

Since a linear transformation of the variables in a regression does not

2 -2
affect the R , the test can be computed by regressing e on an intercept and

S
»
^-i> "
(i = 1.....P)

The independent variables of the auxiliary regression are seen to be Box-Cox

_2 * *
transformations of e where 6 is the Box-Cox parameter.
v When 6 =1, the
t

statistic is equivalent to Engle's ARCH test. As 6 — > 0, the test would be

-2 -2
based on the regression of e on an intercept and log(e .) (i = l,...,p).

This limiting case corresponds to the test for ARCH when the alternative is

the logarithmic model (1.2). If a regression package is available for

estimating a Box-Cox model with transformation only of the independent

variables, S could be obtained very easily.

4. SIMULATION EXPERIMENTS

In this section we conduct a simulation study to determine the accuracy

of the approximation (2.2). We also consider the power of Davies' test and

Engle's ARCH test when the alternative model is in the class of NARCH models
 .

14

Lastly, we compare the power of these two tests under a bilinear alternative,

to the LM test for that specific alternative.

To determine the accuracy of the approximate level of significance

(2.6), for various sample sizes ranging from 25 to 200, 500 random normal

samples were generated and Davies' test for NARCH (D-N) was computed.

Therefore, the maximum standard error of the estimates of type 1 error

probabilities and power in the following tables would be V 5x 5/500 ~ .022. .

The D-N test is based on the alternative hypothesis that the series is

generated with a conditional mean zero and NARCH(l) heteroskedasticity , that

is

y J $ t-l ~ N(0 h t } ' (4 - 1}

where

1/5
2 8
h
t
=
[*
(o ) + ^(y^) (4.2)

Computing the approximate level of significance requires finding the supremum

1/2
of S(<5) and the total variation of S (5) over the permissible range of 5.

Since the parameter space only imposes 5 > 0, an upper bound for 6 must be

chosen. We only present results for < 5 < 2. Below we discuss the

consequences of varying the upper bound. The supremum of S(<5) and the
1/2
turning points of S (5) were found using a grid search with step length

.01. Davies' test provides only an approximate p.- value for the statistic.

To examine the quality of this approximation, we choose a nominal level of

significance, then compute the p-value for each sample. When a computed
 15

p-value is less than the nominal significance level, a rejection is recorded.

Estimates of the type 1 error probabilities are obtained by counting the

number of times the null hypothesis is rejected and dividing by 500. We then

compare the those estimates to the chosen nominal significance level.

In Table 1, we present the estimates of the type 1 error probabilities

of D-N for different sample sizes and different nominal levels of

significance. For comparison, we also report the corresponding values for

Engle's LM test for ARCH (LM-A) against an ARCH(l) alternative. A rejection

2
for LM-A is recorded when the computed value of the statistic exceeds the x-,

critical value determined by the nominal level of significance. The results

for D-N indicate that the approximation (2.6) works well. All estimated

probabilities are quite close to the specified nominal significance levels.

The quality of the approximation for D-N is certainly no worse than the

approximation provided by the asymptotic distribution theory for LM-A.

We also give in Table 1 estimates of probabilities of type 1 error

2
obtained by comparing the supremum of S(5) to the x critical value for the

specified nominal significance level. This is the test which results from

omitting the second term in (2.6). The simulations indicate, as expected,

2
that the null hypothesis is rejected too frequently when the simple x->

2
critical value is used. Figure 1 is a plot of the the x-, density function

and a nonparametric estimate of the density function of D-N based on the 500

samples of size 100 from Table 1. A kernel estimator was used with a window

width of 1 and a biweight kernel function [see Silverman (1986)]. As seen in

2
Figure 1, the density of D-N is skewed further right than the x-, density.
 .

16

2
Again, this indicates that a critical value based on the x-, density will lead

to too frequent rejection of the null hypothesis.

The supremum search of S(5) was conducted over the interval < 5 < 2.

All quantities in Table 1 were also computed using 5 and 10 as an upper bound

for 6 . The results indicated that the choice of the upper bound does not

affect the quality of the approximation of the significance level of D-N.

2
The performance of the test using the x-i critical value became worse. Since

the supremum of S(5) cannot decrease when the upper bound of 6 increases, the

likelihood of rejecting the null can get larger as the upper bound for 6

increases

To determine the power of D-N, samples were generated from the model

(4.1) and (4.2). Experiments were conducted with points in the parameter

space at 4> G {.3, .5, .8} x 5 G (. 01, .1,. 3, .5, .8, 1.0,1. 5). To reduce the

computational burden, the step length for both the supremum search of S(6)
1/2
and the computation of the total variation of S (5) was increased to .025.

Again, all experiments were based on 500 replications. We also compute the

empirical power of LM-A for comparison. Results for samples of size 50, 100

and 150 are presented in Table 2.

The results indicate that D-N will significantly increase the ability to

detect conditional heteroskedasticity when the data are generated under

NARCH . For a given sample size T and given value of <f>

as 5 declines to 0,

that is as the alternative moves away from the linear ARCH model, the power

of D-N systematically increases relative to LM-A. In some instances, there

is a gain in power of more than 15%. Equally important, when 6=1, that is
 17

when the true model is precisely Engle's linear ARCH model, there is little

if any loss in power from using the D-N test relative to LM-A.

Regarding the computation of the D-N test statistic S = sup S(5), as we

mentioned earlier, it would be undesirable if S(5) possesses spurious peaks.

In Figures 2 and 3, we present plots from two random replications under the

null and alternative hypotheses, respectively. In Figure 2, although S(5)

has a clear maximum, the graph is somewhat flat. This may be due to the fact

that S is not identified under the null hypothesis. In Figure 3, S(5) has a

maximum very close to the true value .5. The value of 5 for which S(5) is a

maximum, provides an empirical estimate of S. For maximum likelihood

estimation after the testing this could provide a starting value.

It would also be desirable if the D-N test has good power against

other types of nonlinear models. Recently there has been interest in the

ability of different tests to detect a variety of nonlinear models [see

Keenan (1985), and Luukkonen, Saikkonen and Terasvirta (1988)]. In a

preliminary attempt to investigate this possibility, we generated samples

from the single term bilinear model

where

e - N(0,1)
t

As discussed in Higgins and Bera (1989a), this process has second moments

which are very similar to the ARCH model. In Table 3, we present the
 ,

18

estimated power, based on 500 replications, of the D-M

test for

P e {.1,.3,.5,.8} and samples of size 50, 100 and 150. For comparison, we

report the estimated power of LM-A and the LM test

(LM-B) for the specific

alternative given by (4.3). •

The results in Table 5 are self explanatory.

When p is only of moderate size, < .5, the power of D-N and LM-A are very

similar. At - .8, D-N does appear to be slightly more powerful

than LM-A.
Unfortunately, when compared to LM-B, neither test
does well. This, however,
is not unexpected since LM-B is based upon the specific
alternative model
given in (4.3). These preliminary results do indicate that D-N
can detect
types of nonlinearity other than those encompassed
in the NARCH

specification.

5. AN APPLICATION

We motivate our test by suggesting that

the LM test for the linear ARCH

model (LM-A) may not readily detect different

kinds of nonlinearity and

conditional heteroskedasticity . We suggest a LM test (D-N) based on a

broader alternative, the NARCH model, which may

be able detect a wider range

of nonlinearity. To illustrate this possibility, in this section we examine

the spot exchange rates between the U.S.

dollar and the French franc (Ff )
German mark (Cm), Italian lire (II), Japanese
yen (Jy) , Swiss franc (Sf) and
British pound (Bp) The data are monthly from January 1973 to
.
December 1986.
The series actually analyzed are the first
differences of the logarithms
center about their means. These particular series were chosen because
their
conditional means can be represented by a simple
autoregressive (AR) model.
 19

The sample autocorrelation and partial autocorrelation functions of the

series indicate that an AR(1) process is a adequate model for the conditional

mean of each of the series. Since the presence of conditional

heteroskedasticity is anticipated, the significance of the autocorrelations

were tested using standard errors and a portmanteau test robust to the

presence of linear ARCH [see Milhaj (1985) and Diebold (1986)]. The AR(1)

models were estimated by least squares and the least squares residuals were

used to compute Engle ' s LM test for linear ARCH and our test for NARCH for

orders 1 through 12. The p-values for each test statistic are reported in

Table 4. Examining the p-values reveal that the two tests can give very

different impressions about the presence of ARCH and nonlinearity . LM-A does

not indicate any ARCH, at conventional levels of significance, for Ff, II, Jy

and Sf; however, D-N finds ARCH significant at the 10% level for at least one

order for each of these series. For Gm and Bp , the two tests are in close

agreement through all orders of the test.

To further illustrate that LM-A may fail to detect nonlinearity when the

conditional heteroskedasticity is not linear, both the linear ARCH and NARCH

models were estimated for the II series. The II series was chosen because

the discrepancy between the two test seems greatest. The smallest p-value of

LM-A is .43, while the p-values for D-N are less than .10 at orders 3, 4, 8,

11 and 12. Table 5 shows maximum likelihood estimates for the linear ARCH(3)

model (L-ARCH) and the NARCH(3) models. Other order models were also

estimated, but these gave the best fit. In spite of LM-A being

insignificant at all orders, it is evident from Table 5 that some form of

nonlinearity is present in II. The parameter a is significant in L-ARCH and

 20

the parameters a~ and a~ are significant in NARCH. The salient result in

Table 5 is that the nonlinearity parameter 5 is estimated as 4.00 in the

NARCH specification. The high degree of nonlinearity in the conditional

variance function may explain why LM-A does not detect conditional

heteroskedasticity in this series.

6. CONCLUSIONS

Our Monte Carlo results present evidence that the approximation given by

Davies (1987) is sufficiently accurate in small samples so as to be able to

confidently use D-N. The power studies indicate that D-N can be

significantly more powerful than the LM test for linear ARCH when the

alternative is NARCH. As illustrated by our simulation study, this is

particularly true when the nonlinearity parameter S is quite small. From our

experience in estimating NARCH models with exchange rate data, small values

of 8 are frequently encountered. Also in the empirical example presented

here, LM-A could not detect heteroskedasticity when the estimated value of 6

was quite high. However, the D-N test was able to pick up this nonlinear

conditional heteroskedasticity. Hence, D-N should be a useful test when some

type of conditional heteroskedasticity is suspected.

 ,

21

REFERENCES

BOLLERSLEV, T. (1986) Generalized autoregressive conditional

heteroskedasticity. J. Econometrics 37, 307-27.

BOLLERSLEV, T. (1988) On the correlation structure for the generalized

autoregressive conditional heteroskedastic process. J. Time Series Anal.
9, 121-31.

BREUSCH, T. S. and PAGAN, A. R. (1980) The Lagrange multiplier test and its
applications to model specification in econometrics. Rev. of Economic
Studies 47, 239-53.

DAVIES, R. B. (1977) Hypothesis testing when a nuisance parameter is present

only under the alternative. Biometrika 64, 247-54.

DAVIES, R. B. (1987) Hypothesis testing when a nuisance parameter is present

only under the alternative. Biometrika 74, 33-43.

DIEBOLD, F. X. (1985) Testing for serial correlation in the presence of

heteroskedasticity. Proceedings of the American Statistical Association
Business and Economic Statistics Section, 323-328.

ENGLE, R. (1982) Autoregressive conditional heteroscedasticity with

F.
estimates of the variance of United Kingdom inflation. Econometrica 50,
987-1007.

ENGLE, R. F. and BOLLERSLEV, T. (9186) Modelling the persistence of

conditional variances. Econometric Rev. 5, 1-50.

GEWEKE, J. (1986) Comment. Econometric Rev. 5, 57-61.

HIGGINS, M. L. and BERA, A. K. (1989a) A joint test for ARCH and bilinearity
in the regression model. Econometric Rev. 7, 171-81.

HIGGINS, M. L. and BERA, A. K. (1989b) A class of nonlinear ARCH models:

properties, testing and applications. BEBR Faculty Working Paper
89-1554.

KEENAN, D. M. (1985) A Tukey nonadditivity- type test for time series

nonlinearity . Biometrika 72, 39-44.

LUUKKONEN, R. SAIKKONEN P. and TERASVIRTA, T. (1988) Testing linearity in

, ,

univariate time series models. Scand J. Statist. 15, 161-75.

.

MILH0J, A. (1985) The moment structure of ARCH models. Scand. J. Statist. 12,
281-292.

NELSON, D. B. (1989) ARCH models as diffusion approximations. University of

Chicago, Graduate School of Business Working Paper 89-77.
 22

PAGAN, A. R. and WICKENS M. R. (1989) A survey of some recent econometric

,

methods. Economic J. 99, 962-1025.

PANTULA, S. G. (1986) Comment. Econometric Rev. 5, 71-73.

ROY, S. N. (1953) On a heuristic method of test construction and its use in

multivariate analysis. Ann. Math. Statist. 24, 220-38.

Silverman, B. W. (1986) Densitity Estimation for Statistics and Data

Analysis London: Chapman and Hall.
.

WATSON, M. W. and ENGLE, R. F. (1985) Testing for regression coefficient

stability with a stationary AR(1) alternative. Rev. Economics and
Statistics 67, 341-46.

WEISS, A. A. (1984) ARMA models with ARCH errors. J. Time Series Anal. 4,
269-43.
 TABLE 1

ESTIMATES OF THE TYPE 1 ERROR PROBABILITIES

10% 5% 1%

SAMPLE D-N LM-A *1 D-N LM-A x\ D-N LM-A X\

SIZE

25 .082 .056 .152 .038 .022 .092 .018 .008 .030

50 .066 .050 .178 .032 .016 .086 .006 .004 .024

75 .076 .082 .158 .032 .024 .090 .014 .008 .022

100 .072 .088 .150 .030 .034 .080 .010 .004 .016

150 .094 .078 .166 .056 .036 .114 .010 .012 .036

200 .102 .104 .196 .058 .044 .126 .020 .018 .038
 TABLE 2

ESTIMATED POWER OF D-N AND LM-A AGAINST NARCH(l)

i
3 5 8

SAMPLE
SIZE 6 D-N LM-A D-N LM-A D-N LM-A

.01 .466 .374 .876 .690 .998 .946

.10 .404 .318 .806 .642 .994 .902
.30 .334 .314 .640 .536 .936 .826
50 .50 .298 .302 .542 .482 .890 .774
.80 .276 .272 .524 .490 .790 .748
1.00 .278 .292 .444 .448 .712 .660
1.50 .350 .364 .494 .498 .694 .674

.01 .872 .668 1.000 .938 1.000 .994

.10 .646 .538 .946 .796 1.000 .982
.30 .488 .458 .858 .758 .998 .958
100 .50 .402 .446 .792 .720 .962 .896
.80 .376 .396 .720 .704 .932 .888
1.00 .404 .424 .678 .680 .910 .868
1.50 .448 .490 .662 .670 .840 .812

.01 .970 .838 1.000 .992 1.000 1.000

.10 .746 .612 .990 .904 1.000 .996
.30 .624 .566 .954 .858 1.000 .978
150 .50 .544 .526 .888 .818 .998 .974
.80 .482 .498 .814 .790 .980 .974
1.00 .540 .546 .808 .808 .954 .936
1.50 .606 .638 .818 .816 .938 .914
 TABLE 3

ESTIMATED POWER OF D-N, LM-A AND LM-B AGAINST BILINEARITY

P
SAMPLE
SIZE .1 .3 .5 .8

D-N .100 .133 .167 .267

50 LM-A .033 .167 .133 .200
LM-B .167 .367 .567 .733

D-N .038 .126 .304 .456

100 LM-A .034 .124 .332 .404
LM-B .154 .692 .938 .928

D-N .050 .172 .404 .602

150 LM-A .056 .184 .436 .478
LM-B .224 .886 .994 .964
 TABLE 4

P-VALUES OF LM-A AND D-N FOR EXCHANGE RATE DATA

ORDER
Ff Gm 11 & Sf Bp

OF
ARCH LM-A D-N LM-A D-N LM-A D-N LM-A D-N LM-A D-N LM-A D-N

1 .47 .10 .01 .00 .85 .24 .37 .46 .41 .32 .01 .02
2 .50 .16 .03 .01 .52 .33 .42 .14 .62 .48 .02 .05
3 .46 .14 .02 .01 .43 .06 .61 .15 .23 .05 .05 .12
4 .63 .30 .03 .01 .50 .08 .64 .06 .30 .08 .09 .21
5 .40 .13 .07 .07 .59 .13 .70 .11 .24 .03 .16 .33
6 .59 .21 .17 .13 .67 .11 .69 .19 .18 .04 .12 .26
7 .38 .11 .06 .05 .53 .11 .78 .17 .21 .05 .18 .36
8 .40 .12 .05 .05 .63 .09 .86 .23 .21 .05 .12 .26
9 .52 .18 .32 .35 .70 .11 .87 .32 .33 .11 .11 .22
10 .59 .24 .35 .34 .69 .18 .91 .44 .43 .17 .13 .14
11 .63 .66 .58 .45 .66 .09 .80 .22 .55 .16 .16 .20
12 .79 .67 .85 .61 .70 .09 .83 .17 .67 .22 .12 .29
 TABLE 5

ESTIMATED MODELS FOR DOLLAR/LIRE EXCHANGE RATE'

L-ARCH: <j> a~ a.. a. a

.396 3.722 .020 .091 .048

(.095) (.686) (.054) (.035) (.073)
2
NARCH: 4> a a a a 8

.397 3.723 .0001 .006 .0008 4.00

(.0006) (.0006) (.0004) (.0006) (.00006) (.0005)

Standard errors of the estimates are shown in parenthesis.

55
i

O
u
o

<u
p
ni
e
•H
•P
W
0)

p
•H
CO
C
<u
T3

P.
P
<U
6
P
P<
C
o
55

D
O
o
d

o CO
•

CO
w
<u
&+J
o•
o
0*
r^ >>
&
f^
rA
o 5J
£
to
>-i

<u
T3
c
o 3
in •"N
<-o
v-/
CO

o
** ,

P
i-l
P-.

o .

to CM

W
od
O
o o
M
CM Pu

o
d
oo r^
*~ ^m
£ 2 o o o o
en oo
° t-
o
*™ •-»
£
^ w
1ECKMAN LSJ
INDERY INC. |M|
JUN95
»d To-fW- N. MANCHESTER,
INDIANA 46962