measuring the strangeness of gold and silver rates of return.pdf
TRANSCRIPT
-
8/10/2019 Measuring the Strangeness of Gold and Silver Rates of Return.pdf
1/16
The Review of Economic Studies, Ltd.
Measuring the Strangeness of Gold and Silver Rates of ReturnAuthor(s): Murray Frank and Thanasis StengosReviewed work(s):Source: The Review of Economic Studies, Vol. 56, No. 4 (Oct., 1989), pp. 553-567
Published by: Oxford University PressStable URL: http://www.jstor.org/stable/2297500.
Accessed: 12/05/2012 14:17
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at.http://www.jstor.org/page/info/about/policies/terms.jsp
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact [email protected].
Oxford University Pressand The Review of Economic Studies, Ltd.are collaborating with JSTOR to digitize,
preserve and extend access to The Review of Economic Studies.
http://www.jstor.org
http://www.jstor.org/action/showPublisher?publisherCode=ouphttp://www.jstor.org/stable/2297500?origin=JSTOR-pdfhttp://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/stable/2297500?origin=JSTOR-pdfhttp://www.jstor.org/action/showPublisher?publisherCode=oup -
8/10/2019 Measuring the Strangeness of Gold and Silver Rates of Return.pdf
2/16
Review of Economic Studies (1989) 56,
553-567
0034-6527/89/00360553$02.00
?
1989 The
Review of
Economic
Studies
Limited
easuring
t h
Strangeness
o
o l d
n d
S i l v e r R a t e s
o e t u r n
MURRAY FRANK
University of British Columbia
and
THANASIS
STENGOS
University of Guelph
First
version received January
1988;
final
version
accepted
March
1989 (Eds.)
The
predictability
of rates of return on
gold
and
silver
are
examined.
Econometric tests do
not reject the
martingale hypothesis for either asset. This failure to
reject
is
shown to
be
misleading.
Correlation dimension estimates
indicate
a
structure not
captured
by
ARCH.
The
correlation
dimension is
between 6 and
7
while the
Kolmogorov entropy is about
0-2
for both assets.
The
evidence is
consistent with
a
nonlinear deterministic data
generating process underlying the rates
of
return. The evidence is
certainly not sufficient to rule out
the possibility of some
degree
of
randomness
being present.
1. INTRODUCTION
The purpose of
this paper is to examine the
predictability of asset price changes. In
well-functioning
markets it
is often
held
that asset
price changes should
be
unpredictable.
Otherwise
speculators
will take
advantage
of the
predictable price change
thereby forcing
the
change
to
happen
immediately. Any subsequent
changes
are then left
unpredictable.
This
intuition was formalized
by
Samuelson
(1965)
and
has been
incorporated
in
standard
textbooks
such as
Brealey
and
Myers
(1984).
This intuition
is often
taken
to be
synony-
mous
with
the efficient markethypothesis.
Analysis
of
intertemporal general
equilibrium
models
indicates
that
this intuition is
only rigorously
justified
under
fairly stringent
conditions, see
Lucas
(1978)
and Brock
(1982). Despite being
a rather
special
case
theoretically,
considerable
empirical support
has, been
reported
for the
martingale
hypothesis ,
see
Fama
(1970).
Sims
(1984)
has
provided
an
interesting
rationalization of the
empirical
success
of
the
martingale hypothesis.
Our focus is
on
the extent to which the views of
Sims,
or
chaos
in
the
sense
of Brock
(1986)
might
be
responsible
for the
empirical
success
of
the
martingale hypothesis.
These
two
approaches
are
quite
different.
Scheinkman and
LeBaron
(1986) tested
for
chaos
using weekly
returns on
stocks
taken from the Center
for Research
in
Security
Prices at the
University
of
Chicago (CRSP). They
devoted
most
attention to
an index of
stocks but also considered
a
number of
individual stocks. Their
results are seemingly consistent with chaos.'
1.
The
behavior ...
seems to
leave
no
doubt that
past
weekly returns
help
predict
future
ones ...
Further
it
seems
that most
of
the
variation on
weekly
returns is
coming from
nonlinearities as
opposed
to
randomness.
Or more
moderately,
the data
is not
incompatible
with
a
theory
where
most of
the
variation
would come
from
nonlinearities
as
opposed
to
randomness
and
is not
compatible
with a
theory that
predicts
that the
returns are
generated
by
independent
random
variables.
Scheinkman and
LeBaron
(1986).
553
-
8/10/2019 Measuring the Strangeness of Gold and Silver Rates of Return.pdf
3/16
554
REVIEW
OF
ECONOMIC
STUDIES
The paper
is
organized
as follows.
Section 2
is
concerned
with some
theoretical
background.
First
we give an
indication of
the
restrictions
needed
to
generate the
martingale
hypothesis
in
a standard
economic
setup.
Then we
discuss
the approach
of
Sims
(1984). That
approach
was
developed
in
reaction to
the
restrictiveness of the
assumptions required to generate the martingale hypothesis in a standard setup. Next a
definition of
chaos is offered
in
order
to
illustrate the
sense
in
which
chaos provides an
alternative
to Sims
(19984).
Section
3 is concerned with the
testing
methods. Our work
is
based
on the
correlation
dimension and the
Kolmogorov
entropy.
Since
these are
not
yet
standard
tools
in
empirical
economic
research we
attempt
to
motivate their use.
As
the correlation
dimension
has
already
been
analyzed by
Brock
(1986)
and Brock
and
Dechert
(1988)
we
focus
more
attention
on the
Kolmogorov
entropy.
The
empirical results are set out
in
Section
4,
and
a brief
conclusion is offered
in
Section 5.
2.
THEORETICAL
UNDERPINNINGS
A.
Traditional
efficient
markets
hypothesis
A
standard
context
in
which to
consider asset
pricing
issues are the
intertemporal
models
of
Lucas
(1978)
and
Brock
(1982).
In
the Lucas
model one
derives a Euler
equation
for
the
representative
agent
relating
the
marginal
utility
of
consumption
at
date
t
to
expected
marginal
utility
of
consumption
at date
t
+
1.
In
order to
obtain the
martingalehypothesis
in
asset prices added
structure
is
required. One
needs to correct
for
dividends
and
discounting
and then either
assume
risk-neutrality
or
else
assume that there is no
aggregate
risk.
Let
Pt
be
the
price
of an
asset
at date
t
and
let
Et be the
expectations operator
conditioned on
the
information
available at
date t. A
process
{Pt}
is said
to
be a
martingale
if
E,p,+?
=
Pt
for
all s
>
0.
The claim
that asset
prices constitute
a
martingale
is
frequently
identified
in
textbooks as the essence of
the efficient
markets
hypothesis .
For
theoretical
analysis of such
economies see the
general
equilibrium models
of
Lucas
(1978)
and
Brock
(1982).
Hansen and
Singleton
(1983)
have
an
important
special
case of the
Lucas-Brock
approach
for which
there is a
closed-form
solution.2
Hansen
and
Singleton
(1983)
empirically
test their solution.
B.
Sims'
approach
To
obtain the
martingale
hypothesis
in
a
Lucas-Brock
setup requires
very
restrictive
assumptions. This led
Sims to
question whether
the
apparent
empirical
success of the
martingale
hypothesis
is little more
than a
robust
fluke. In
place of
the
Lucas-Brock
framework
Sims
(1984) provides
an
analysis
in
which
the
martingale
hypothesis
is
obtained
for
very
short
time
intervals.
Lengthy
time
intervals
need not
satisfy
the
martingale
property.
Definition 2.1. A process
{Pt}
is said to be instantaneously unpredictable (I.U.) if
li,
Et[pt+v
-
Et
(pt+v)2]
as
V-0Et[pt+v
-
(Pt)2]
2.
The
consumer
has constant relative
risk-aversion
and the
joint
distribution of
consumption
and
returns
is
lognormal.
-
8/10/2019 Measuring the Strangeness of Gold and Silver Rates of Return.pdf
4/16
FRANK & STENGOS GOLD AND SILVER 555
If asset prices satisfy Definition 2.1
and there are stationary increments, then a
regression of
(p,?
-p,)
on any variable known
at date t will have an R2 that approaches
zero as
s
approaches zero. Changes over
long time periods may be forecastable, but
short-term price changes should be
unpredictable. In modern financial theory diffusion
processes are often assumed. Diffusion processes satisfy the I.U. property. Any process
with
well-behaved derivatives will not satisfy
the I.U. property.
Sims
(1984) provides
a
detailed
discussion of the I.U. property and a very strong
argument
for
it
as
the basis
of
the empirical success of the martingale hypothesis. He
concludes that Except for the easily identified exceptional cases, neither the real world
nor an analytically manageable economic
model is likely to generate security prices which
fail to
be
instantaneously unpredictable .
C.
Chaos
In some respects chaos or strangeness is the polar opposite to a process that is
instantaneously unpredictable. There are a
number of different
definitionsg-of
chaos in
current use. Not all of the definitions are equally practical for empirical research.
Definition
2.2. Let fl be. a
space
with
metric d and
let
f
:
fl
-
fl be a continuous
mapping
defined
on
ft.
A
discrete
dynamical
system
(fQ,f) is
said to be chaotic
(or
strange)
if there exists a
8 >
0 such
that for
all c E: l and
all
?
>
0
there is
co
fE
and
k
such
that
d
(o, co')
< E
but
d
(fkc,fkw')
_8
In
this definition fkco denotes the k-fold iteration
of point Ctby the map
f
Definition
2.2
contains less
than
is frequently included
in
definitions of chaos. Devaney (1986) for
example includes the requirement that there exist dense orbits and that periodic points
are dense. We do
not take such
a
definition since the
additional conditions do not seem
to
be
verifiable nor refutable for empirical
systems. There seems little point to including
conditions that one cannot check. In practice
the existence of dense orbits must be
assumed
in
any case
in
order
to characterize
the system. Our definition follows the usage
of
Eckmann and
Ruelle
(1985)
who
take
chaos to be
synonymous
with sensitive
dependence
on
initial conditions .
Brock
(1986)
defines
chaos
in
terms
of
the
largest
Lyapunov exponent being positive.
The
Lyapunov exponent definition is
related
to the
Kolmogorov entropy.
The
Kolmogorov
entropy
is a lower
bound
on the sum of the
positive Lyapunov exponents,
see Eckmann and Ruelle
(1985).
The
Kolmogorov entropy
is discussed in Section 3B.
A chaotic
system
will be
quite predictable
over
very
short time horizons.
If
however
the
initial
conditions
are
only
known
with finite
precision,
then over
long
intervals
the
ability to predict
the
time path
will be lost. This is
despite
the
process being
deterministi-
cally generated. Typically
for chaotic
systems nearby trajectories locally separate exponen-
tially
fast.
The
mathematical
theory
of chaos is
currently
an active research
area,
see
Lasota
and
Mackey (1985), Devaney (1986)
and
Guckenheimer and Holmes
(1986).
There are
a
great many ways
in
which
chaos
might
enter an
economic
system.
Our work is not
tied
to any particular entry mechanism. If evidence of chaos is found then it becomes a
natural
topic
for
further
research
to
attempt
to
identify
its
source
or
sources.3
3. A
particularly simple
ad
hoc
example is as follows. Let
X,+,
4X,(1
-
X,) and
let
P,,1
3 P,
+ (X, -0.5).
Simulate this two
equation system
starting
with
Xl
E
(0, 1)
and P1 =
100. On the
simulated data
test
(P,,,
-
P,) =
ao+
a1(P,
-
P,1)
+
E,. One will
be unable to
reject
a0
=
a,
=
0
and the R2
will be very
close to zero.
However
this
example
is
anything
but
unpredictable.
This
example
is
discussed more
fully
in
Frank
and
Stengos (1988).
Also
consider footnote
12.
-
8/10/2019 Measuring the Strangeness of Gold and Silver Rates of Return.pdf
5/16
556
REVIEW OF ECONOMIC STUDIES
We have
three alternative
possible
interpretations
of the observed
irregularity
of asset
prices. The Lucas
(1978) framework considered
in
section
A does not tie the martingale
hypothesis to the
size
of the time intervals being employed empirically. Sims (1984)
generates unpredictability explicitly
for short time intervals. For lengthy time intervals
Sims' theory permits the asset price changes to be predictable. The chaos interpretation
leaves unspecified
the economic mechanism which generated
the data. It is consistent
with many possible theories including
versions
of
the
Lucas-Brock
setup. Chaos is not
consistent with Sims'
approach. If chaos is present then asset
price changes will (at least
in
principle) be predictable
over short
time intervals, but not over long time intervals.
Over long
time intervals,
due
to
the sensitive
dependence
on initial conditions,
the asset
prices will not be predictable.
It is worth emphasising that these three
approaches by no means exhaust the
set of
conceivable theories of
the observed irregularity
of
asset prices. These three approaches
are considered together
since each
has been suggested previously as
a
possible
interpreta-
tion. In rejecting one or more of these possible views we do not establish that a particular
alternative is true.
This familiar methodological
point is particularlypertinent with respect
to
any suggestion
of deterministic chaos.
3. TESTING METHODOLOGY
In
this section
we
describe
the two measures
on which
our empirical
work
is based.
The
theoretical
underpinning
is an
assumption
of
ergodicity.
Such an
assumption
is
required
if
we are to use time averages as representative
of
the
system's
behaviour. The first
invariant
of
the
system
is the correlation
dimension.
The
second invariant
of interest
is
the Kolmogorov entropy.
We take these
in
turn.
A. Correlation
dimension
The correlation
dimension is
originally
due to
Grassberger
and Procaccia
(1983)
and
Takens (1983).
For
more
detail than
we
provide
see
Brock and
Dechert
(1988)
and
Eckmann
and Ruelle
(1985).
Start
by
assuming
that the
system
is on an attractor .
An
attractor is
a
closed
compact
set
S with
a
neighbourhood
such that almost
all initial conditions
in the
neighbourhood have S itself as their forward-limit set. In other words, these initial
conditions
are
attracted
to S as time progresses. The neighbourhood
is termed the
basis of attraction
for
the attractor.
An attractor
satisfying
Definition 2.2 is then called
a
strange attractor
or
else
a chaotic attractor.
Consider
a
time-series
of rates of return
r,,
t
=
1, 2, 3,
.
.
.,
T.
We
suppose
that
these
were
generated by
an orbit or
trajectory
that is dense on the
attractor.
Use the time-series
to create
an
embedding.
In
other words create M-histories
as
rm
=
(r,,
rt+?,
*,
rt+M-1).
This converts
the series of scalars into a series of vectors with overlapping
entries.
If
the
true
system
which
generated
the
time-series is
n-dimensional,4
then
provided
M
?
2n
+
1
generically
the M-histories recreate
the
dynamics
of the
underlying system (there
is a
diffeomorphism between the M-histories and the underlying data generating system).
This extremely useful
mapping
between the underlying system
and
the M-histories
was
4. For an intuitive discussion of the meaning of dimension see Frank and Stengos (1988). Familiar,
smooth examples include:
a
point is zero-dimensional,
a
line is one-dimensional,
a
plane is two-dimensional.
These objects retain their dimensionality even when embedded in less restricted spaces, say R5.
-
8/10/2019 Measuring the Strangeness of Gold and Silver Rates of Return.pdf
6/16
FRANK & STENGOS GOLD
AND SILVER 557
established by
Takens (1980). It
is this result which permits the empirical
work. Broom-
head and King (1986) discuss certain
practical limitations on the use of
Takens' theorem.
Next one measures the spatial correlations
amongst the points (M-histories) on the
attractor by calculating the correlation
integral, CM(8). For a particular
embedding
dimension M, the correlation integral is defined to be
CM(E)={the
number of
pairs (i,j) whose distance
IIrM_rjMl
C??}/ T2. (3.1)
Here 1. denotes
the distance induced by the selected
norm. We
use the Euclidean
distance. The
other distance function
that is sometimes employed is
the sup-norm. By
Theorem
2.4
of
Brock (1986) the
correlation dimension is independent
of the choice of
norm.
In
principle
T
should
go to infinity,
but
in
practice
T
is
limited
by the length of
the available time
series.
This will
in
turn
place
limitations on the choice
of 8.
To obtain
the correlation dimension,
DM
take
D
=
lim6?O
{ln
Cm
()/ln
E}.
(3.2)
As
a
practical
matter
one
searches to see if the values of
DM
stabilize at some value D
as
M
increases.
If
so,
then D is the correlation dimension estimate. If
however,
as M
increases the
DM
continues to
increase at the same rate then the system
is taken to be
high
dimensional or
in
other words stochastic.
If
a low value for
DM
is found then
the
system
is
substantially
deterministic even
if
complicated.
In principle
an independently
and
identically
distributed stochastic
system is infinite dimensional.
Each time one
increases the available degrees of freedom, the system
utilizes that extra
freedom. With
finite data sets, high dimensionality
will be indistinguishable from infinite
dimensionality
empirically,
see
Ramsey
and Yuan
(1987) concerning
small data sets.
Two practical problems concerning
8
should be noted. If
8
is too large, then
CM(,)
=
1
and no information about the system is
obtained. It- s also possible for
8
to
be
too small.
With
finite
data sets there is a limit to the degree
of
detail that one may
discern. This limitation on the
ability
to
get
a detailed focus means that even
in
principle
one
can never exclude the
possibility
of the
system
containing
some
degree
of additive
noise. However,
the test still can
find
out
if
there are substantial nonlinearities moving
the
system.
Empirically finding
an
appropriate range
of
values
for
?
is
not
difficult for
these series.
There are
several papers
in
economics that use
the correlation integral. Barnett and
Chen (1988) used it to examine monetary aggregates, a low correlation dimension estimate
was
obtained. Brock
and
Sayers (1988) investigated
American
macroeconomic
time-series.
They reject
chaos
but
find
some evidence
of nonlinear
structures.
As
previously
indicated
Scheinkman and
LeBaron
(1986)
examined
American
stock
market data
and
obtained
results
strikingly
similar
to
those
that
we
obtain. Since the number
of related
papers
is
large
and
rapidly
growing
we
do not
carry
out a full
survey
here. For an overview of the
literature
see
Frank and
Stengos
(1988).
B.
Kolmogorov
entropy
Dimension measures the degree of complexity of a system. Entropy is a measure of time
dependence.
The
Kolmogorov5
entropy,
K,
quantifies
the
concept
of
sensitive depen-
dence on initial
conditions .
It
is
frequently
described
as
measuring
the
rate at which
5. It
is
also
termed
Kolmogorov-Sinai invariant ,
measure theoretic
entropy
or sometimes
simply
entropy .
-
8/10/2019 Measuring the Strangeness of Gold and Silver Rates of Return.pdf
7/16
558
REVIEW OF
ECONOMIC
STUDIES
information is
created. This is due to the
following argument. Consider two
trajectories
that are so close
initially
as
to
be
indistinguishable
to an
observer. As time passes, the
trajectories
may separate
and become
distinguishable.
The
entropy measures how
rapidly
this happens.
For an ordered system, that is to say quasi-periodic or less erratic still, K = 0, while
for
an independent and
identically distributed stochastic system K
=
+00. For a
deter-
ministic
chaotic system 0