Description
used to develop a new measure of contagion
using exchange rate data from the Asian Crisis
of 1997 and beyond. Connection weight
changes during retraining of networks used to
forecast exchange rates form the basis of this
measure. These weight changes are used in
obtaining a contribution factor for independent
variables used in a forecasting process.
Accounting Research Journal
Measuring Contagion in the South-East Asian Economic Crisis: An Exploration
Using Artificial Neural Networks
Callum Scott
Article information:
To cite this document:
Callum Scott, (2006),"Measuring Contagion in the South-East Asian Economic Crisis: An
Exploration Using Artificial Neural Networks", Accounting Research J ournal, Vol. 19 Iss 2 pp. 139 -
152
Permanent link to this document:http://dx.doi.org/10.1108/10309610680000684
Downloaded on: 24 January 2016, At: 21:00 (PT)
References: this document contains references to 0 other documents.
To copy this document: [email protected]
The fulltext of this document has been downloaded 401 times since 2006*
Users who downloaded this article also downloaded:
Freddie Choo, Kim Tan, (2006),"A Commentary on Sample Design Issues in Behavioral
Accounting Experiments", Accounting Research J ournal, Vol. 19 Iss 2 pp. 153-158 http://
dx.doi.org/10.1108/10309610680000685
Pak K. Auyeung, Ron Dagwell, Chew Ng, J ohn Sands, (2006),"Educators’ Epistemological Beliefs
of Accounting Ethics Teaching: A Cross-Cultural Study", Accounting Research J ournal, Vol. 19 Iss
2 pp. 122-138http://dx.doi.org/10.1108/10309610680000683
Yass A. Alkafaji, Nauzer Balsara, J udith N. Aburmishan, (2006),"FASB’s Statement No. 133 on
Derivatives and Barings Bank: The Case for Value at Risk (VAR)", Accounting Research J ournal,
Vol. 19 Iss 2 pp. 94-104http://dx.doi.org/10.1108/10309610680000681
Access to this document was granted through an Emerald subscription provided by emerald-
srm:115632 []
For Authors
If you would like to write for this, or any other Emerald publication, then please use our
Emerald for Authors service information about how to choose which publication to write for and
submission guidelines are available for all. Please visit www.emeraldinsight.com/authors for
more information.
About Emerald www.emeraldinsight.com
Emerald is a global publisher linking research and practice to the benefit of society. The
company manages a portfolio of more than 290 journals and over 2,350 books and book series
volumes, as well as providing an extensive range of online products and additional customer
resources and services.
Emerald is both COUNTER 4 and TRANSFER compliant. The organization is a partner of the
Committee on Publication Ethics (COPE) and also works with Portico and the LOCKSS initiative
for digital archive preservation.
D
o
w
n
l
o
a
d
e
d
b
y
P
O
N
D
I
C
H
E
R
R
Y
U
N
I
V
E
R
S
I
T
Y
A
t
2
1
:
0
0
2
4
J
a
n
u
a
r
y
2
0
1
6
(
P
T
)
*Related content and download information correct at time of
download.
D
o
w
n
l
o
a
d
e
d
b
y
P
O
N
D
I
C
H
E
R
R
Y
U
N
I
V
E
R
S
I
T
Y
A
t
2
1
:
0
0
2
4
J
a
n
u
a
r
y
2
0
1
6
(
P
T
)
Measuring Contagion in the South-East Asian Economic Crisis
139
Measuring Contagion in the South-East
Asian Economic Crisis: An Exploration
Using Artificial Neural Networks
Callum Scott
Department of Finance
University of Melbourne
Abstract
An artificial neural network methodology is
used to develop a new measure of contagion
using exchange rate data from the Asian Crisis
of 1997 and beyond. Connection weight
changes during retraining of networks used to
forecast exchange rates form the basis of this
measure. These weight changes are used in
obtaining a contribution factor for independent
variables used in a forecasting process.
Volatilities of contribution factors form the
basis of the measure of contagion obtained.
These volatilities are statistically validated
through a series of simulations where critical
values for them are derived. The measures of
contagion obtained are then matched to
concurrent economic and financial shocks that
occurred during the crisis. It is found that there
is good correlation between these events and the
contagion measures obtained.
1. Introduction
This paper proposes and assesses a novel
measure of contagion using an artificial neural
network (ANN) forecasting methodology. This
methodology is non-parametric. If ANNs are
used to forecast, contribution factors of
independent variables can be generated. A
contribution factor
1
can be thought of as a
measure of the efficacy, or “importance” of a
variable to the forecast. Thus if, for example, we
make a forecast of the exchange rate of
Thailand, using as one of the input variables a
lag of the exchange rate of Malaysia, the
contribution factor of the Malaysian exchange
rate can be regarded as a measure of the impact
of the Malaysian exchange rate on the exchange
rate of Thailand.
1 Contribution factors are fully defined in Section 2.
We propose that the size of these contribution
factors, their volatilities and their differences
2
are possible measures of contagion. Rather than
compare our results with other contagion
measures, we examine underlying economic
events that could impact on the relevant
economies to see if there is a correlation with the
values of contribution factors obtained for that
time. This is done in Section 5.
We use exchange rate data for four countries
that suffered massive economic downturns
during the Asian financial crisis. Obviously,
more economies suffered greatly then, but we
restrict our study to the chosen four to test our
proposal. This set could easily be expanded.
Several researchers have presented
explanations for the onset of the Asian crisis of
1997. A useful starting point in reviewing some
of these explanations is (Jackson 1999), where
several authors address causes and consequences
of the crisis with respect to several individual
economies of the region. Chowdhry and Goyal
(2000) highlight the effect of inadequate and lax
banking practices, suggesting recapitalization
of performing banks and the closure of
underperformers. Others include Roubini
(1998), Thomson (1998), Johnson, Boone,
Breach and Friedman (1998), Chang and
Velasco (1998), Mishkin (1999) and Radelet and
Sachs (1998). Proposed causes of the crisis
include short-term debt used to finance long-
term projects, an increasing emphasis on
projects of questionable value, moral hazard,
pegged exchange rates and political corruption.
However, it is likely that the cause of the event
2 For example, if three variables x, y, z affect a fourth, q,
then differences of contribution factors of x, y and z
through time, such as CF
x
- CF
y
where CF represents a
contribution factor, will chart the changing importance of
each variable relative to the other in its effect on q.
D
o
w
n
l
o
a
d
e
d
b
y
P
O
N
D
I
C
H
E
R
R
Y
U
N
I
V
E
R
S
I
T
Y
A
t
2
1
:
0
0
2
4
J
a
n
u
a
r
y
2
0
1
6
(
P
T
)
ACCOUNTING RESEARCH JOURNAL VOLUME 19 NO 2 (2006)
140
was a combination of such factors. Whatever
the cause, or causes, the accompanying
phenomenon of contagion had a devastating
effect on several of the economies in the region.
The process of contagion, consisting of
related devaluations in asset markets across
countries, reflects the possible transmission of
sentiment across those markets. One possible
definition of contagion in an economic or
financial context is a significant increase in
cross-market correlation during a period of crisis
and confusion. Pericoli and Sbracia (2003), in
their overview of financial contagion, offer five
representative definitions used in the literature
that are largely subsumed by the definition stated
above. They also present a useful summary of
the theory of contagion and a review of the
empirical literature. Dornbusch, Park and
Claessens (2000) define contagion as a
significant increase in cross-market linkages
resulting from some for of shock, pointing out
that during a crisis, the mode of transmission of
such shocks differs from stable times. Probably
the most frequently used test for contagion is
the threshold index model of Eichengreen,
Wyplosz and Rose (1997) and extended by Van
Rijckeghem and Weder (2001). ANNs and
contribution factors
3
are used in this study to
capture and quantify the contagion effect
resulting from shocks.
One possible measure of contagion could be
the comparing of covariance between stock
markets during a stable period with that
determined during a crisis. Other methods of
detecting and quantifying contagion have been
proposed, such as vector error correction
models, impulse response functions and variance
decomposition of VAR models, utilised by, for
example, Tan (1998). Baig and Goldfajn (1999)
used dummy variables and daily news to study
cross-border contagion in the trigger
4
economies
and Korea during the Asian crisis, while Drazen
(2005) examined political contagion during
the European Monetary System crisis of
1992-3. Glick and Rose (1998) addressed the
phenomenon of clustering in contagion in an
attempt to explain the regional nature of some
3 Contribution factors measure the importance of input
variables in an ANN forecasting system. They are fully
defined later in the next section.
4 The Philippines, Thailand, Malaysia and Indonesia.
contagious events using trade linkages. Forbes
and Rigobon (2002) tested for stock market co-
movement using data from the 1994 Mexican
peso crisis, the 1987 US stock market collapse
and the 1997 Asian crisis. They concluded that
in fact there was no real contagion, but a
continuation of the already existing cross-market
linkages. Kim, Kose and Plummer (2001) tested
two explanations of contagion during the Asia
crisis. These explanations were propagation of
adverse shocks and international investor
behaviour, along with symmetric problems in
domestic economic fundamentals.
Co-integration analysis was used by Chen,
Firth and Rui (2002) to investigate the
relationships between South American stock
markets, and Nieh (2002) examined the
relationships between exchange rate volatilities,
exports, imports and productivity for several
Asian economies. Finally, Daly (2003)
investigated the interdependence of several
South-East Asian stock markets before and after
the crisis. We however, adopt a new approach
using ANN contribution factors generated
through a forecasting process with data
consisting of exchange rates from the South-East
Asian crisis of 1997.
An ANN is a numerical algorithm that uses a
set of input data to find the best numerical
approximation to a set of output data. The ANN
consists of three main features. Firstly, it consists
of three layers, an input layer where input data is
received, a hidden layer where processing occurs
through a summation and activation function
5
,
and an output layer. Secondly, it consists of
connection weights placed on the links between
the input and hidden layers, and between the
hidden and output layers. All data passing across
these links are multiplied by the weight. Thirdly,
during the training process of an ANN, using
historic data, these weights change to optimize
the forecast. Training involves an iterative
process of forward and backward propagation
with an objective to minimize the error in
predicting the output variables. The literature
often describes this as a connectionist approach
5 Activation functions form part of the nodes that make up
the hidden or middle layer of an ANN. Usually, they are
a “squashing “ function that constrains values of the
node’s output to between 0 and 1 or -1 and +1. They add
non-linearity to the ANN. An example is the logistic
function that takes the form 1/(1+e
x
).
D
o
w
n
l
o
a
d
e
d
b
y
P
O
N
D
I
C
H
E
R
R
Y
U
N
I
V
E
R
S
I
T
Y
A
t
2
1
:
0
0
2
4
J
a
n
u
a
r
y
2
0
1
6
(
P
T
)
Measuring Contagion in the South-East Asian Economic Crisis
141
and has been widely applied to financial time
series that exhibit non-linearity that changes over
time.
In this study we use sigmoid activation
functions in the middle and output nodes of the
network. This introduces nonlinearity to the
process. Given that this nonlinearity is thus
introduced, this is reflected in the way
contribution factors change over time through
the retraining process used to generate time
series of contribution factors. Thus, although the
contribution factors, fully defined in the next
section, depend on connection weights that are
linear components, the back-propagation process
used in retraining ensures that these weights
change in a nonlinear way. This nonlinearity
therefore flows into the changes in the
contribution factors that are dependent on weight
changes.
Non-linearity and time variance has been
discussed and tested by researchers such as Lee,
White and Granger (1993) and Zhang (1999).
Hoptroff (1993) offers a practical discussion on
the merits of using ANNs in time series
forecasting covering several issues from a
practical trading viewpoint, including non-linear
modelling, robustness to noise, interpolation and
extrapolation.
Useful comparisons of the ANN approach
with more traditional methodologies have been
made by, for example, Marquez, Hill, Worthley
and Remus (1992). Lippman (1987) discussed
ANNs with respect to robustness and fault
tolerance, while Connor (1988) explored their
ability to discern correct variable
transformations and to handle outliers in data.
Many other studies have also been undertaken to
compare ANN performance with traditional
statistical methods, such as the comparison with
Box-Jenkins techniques made by Tang, de
Almeida and Fishwick (1991), Marquez et al
(1991) with regression methods and Baker and
Richard’s (1999) comparison with linear
regression methods. Generally, ANNs compared
well with these methods, indicating their
effectiveness in a nonlinear, time varying
application
6
. Lim and McNelis (1998) tested
predictability in daily stock price returns in
Australia, and used ANNs to examine the
6 Zhang, Timmis and Hu (1999) reviewed articles
comparing ANN forecasting with other techniques but
few business applications were examined.
response of these stocks to shocks in the
Japanese and US markets. Other models tested
were an autoregressive linear model and a non-
linear GARCH-M model. The ANN model
outperformed the GARCH-M model but not the
linear model.
The methodology used here differs from
those in the contagion literature discussed
earlier, and is non-parametric. No attempt has
been made to optimize the ANNs to improve
performance in forecasting as this is an
exploratory study. Further research could
however be carried out where different ANN
structure and design are tested. For example,
different activation functions could be assessed.
Data from the four economies that probably
suffered the most during the crisis are used
7
,
and a measure of the contagion of each on the
other is derived. Concurrent economic events
are examined to detect any correlation with
corresponding contribution factors.
The economies used here had close trade
links, financial links, and similar
macroeconomic policies. Rapid economic
growth driven mainly by export growth had
occurred in all of them. Most importantly, they
compete in the same export markets and rely on
the same sources of capital. Their currencies
were also all effectively pegged to the $US.
Devaluation of one currency would make that
country’s exports immediately more
competitive than the others, precipitating their
devaluation also. Thus, a cycle of competitive
de-valuations could be initiated.
The exchange rates of Indonesia, The
Philippines, Malaysia and Thailand exhibited a
sharp decline from July 1997 until February
1998 where they stabilize at a much lower
value. We therefore proceeded by making
forecasts of each country’s exchange rate (with
respect to the $US) in turn, using lagged values
of the other three. Forecasts were repeated
moving through time from 1995 – 99, and
adding to the data set used to retrain the ANNs.
Obviously, such a forecast is liable to be
improved with the inclusion of the lags of the
dependent variable as inputs, but this would
negate the collection of contribution factors that
related strictly to the exchange rates of the other
economies.
7 These are Indonesia, The Philippines, Malaysia and
Thailand.
D
o
w
n
l
o
a
d
e
d
b
y
P
O
N
D
I
C
H
E
R
R
Y
U
N
I
V
E
R
S
I
T
Y
A
t
2
1
:
0
0
2
4
J
a
n
u
a
r
y
2
0
1
6
(
P
T
)
ACCOUNTING RESEARCH JOURNAL VOLUME 19 NO 2 (2006)
142
Time series of contribution factors and their
volatilities for each exchange rate were
therefore generated during the forecasting
process. These series mirror the changing data
structure as we move from stability before the
crisis, into the disruption at its onset and
beyond. We now define contribution factors in
terms of ANN connection weights. This method
is that used by Ward Systems (1998) in their
NeuroShell 2 software that is used in this
application.
2. Artificial Neural Networks and
Contribution Factors
2.1 Defining a Contribution Factor
The schematic representation of an ANN is
given in Figure 1. We denote by W1(i,j) the
connection weight
8
connecting the i-th input
layer to the j-th hidden layer. We denote by
W2(j,k) the connection weight connecting the j-
th hidden layer and the k-th output layer. N
i
, N
j
and N
k
are the number of input nodes, hidden
layer nodes and output nodes respectively.
We denote by N1(j) the sum of the
connection weights for the j-th hidden node, that
is
N
i
N1(j) = ? W1(i, j)
i=1
and similarly for the k-th output node
N
j
N2(k) = ? W2(j, k)
j=1
N1(j) and N2(k) represent normalization
factors. That is, they enable the calculation of
the proportion of absolute total connection
weight attributable to any one input. For each
input, i, and each output, k:
N
j
CF
(i,k)
= ? W1(i, j)/N1(j) x W2(j, k)/N2(k) (2.1)
j=1
CF(i,k) in (2.1) represents the contribution of
the i-th input to the k-th output. We require the
total contribution of any input to all outputs.
Thus, for any input, i:
) , (
1
) ( 2
) ( k i CF
k
k N
i CF
=
? =
(2.2)
8 When we refer to a connection weight, we mean its
absolute value.
Thus, CF(i) represents the total contribution
factor of any input, i.
Figure 1
Simple three-layer back
propagation network
It was noted in the Introduction on page 141
that nonlinearity was introduced through the use
of sigmoid functions in the middle and output
nodes of the ANNs. Importantly, we should also
note that the contribution factors as defined
above are a function of the ANN connection
weights. Given the back-propagation
methodology used here to train the ANNs, and
the use of sigmoid functions, changes in the
connection weights during training will behave
in a nonlinear fashion. Our definition of
contribution factors aggregates the connection
weights of each independent variable, therefore
this nonlinearity will flow through to changes in
contribution factors.
2.2 Contribution Factors and Market
Shocks
Time series of contribution factor volatilities of
independent variables in a forecasting process,
and in particular their differences at any given
time, across the different variables used in the
forecasts form the basis of our measure of
contagion. Forecasts of each of the economies
exchange rates are made using input variables to
the ANN that consist of the other economies
lagged exchange rates. Thus, a forecast of the
exchange rate of Thailand would be made using
the lagged exchange rates of Indonesia, The
Philippines and Malaysia. This is repeated for
each economy in turn. We define contribution
factor volatility as
Volatility = ?
t
2
= [ln(CF(X)
t+1
/CF(X)
t
]
2
(2.3)
CF(X) = contribution factor of X
Input Layer
Hidden Layer
Output Layer
W1(i, j)
W2(j, k)
D
o
w
n
l
o
a
d
e
d
b
y
P
O
N
D
I
C
H
E
R
R
Y
U
N
I
V
E
R
S
I
T
Y
A
t
2
1
:
0
0
2
4
J
a
n
u
a
r
y
2
0
1
6
(
P
T
)
Measuring Contagion in the South-East Asian Economic Crisis
143
We now propose how contribution factors of
independent variables used in a forecasting
process would behave where the relevant
economies are subjected to unexpected shocks.
This builds on the introductory discussion of
contribution factor behaviour in Scott (2003).
Consider a market where learning, in the
EMH sense is complete. Assume a shock occurs
at time t = 0 and define pre-shock time as t < 0.
Thus post-shock time will be defined as t > 0.
The contribution factor of an independent
variable, CF(x), at t < 0 should be
approximately constant. Thus
CF(x)
(t 0)
> ?
2
CF(x) critical
(2.6)
When the effects of a shock have been
absorbed, contribution factors may revert to
their previous levels, or, if a permanent change
to market conditions has been established by the
shock, levels that differ from those exhibited
pre-shock:
CF(x)
pre-shock
= CF(x)
post-shock
or
CF(x)
pre-shock
? CF(x)
post-shock
(2.7)
We would therefore expect that contribution
factor volatility for each exchange rate should
be near zero before a divergent event such as
the Asian crisis, increasing at its onset and for
some time thereafter. Volatilities of these
contribution factors in independent variables are
liable to be correlated to economic events
during the crisis. Therefore a matching of
contribution factor behaviour to relevant
economic events was also carried out and is
explored in Section 5.
We now briefly discuss the simulations that
were performed to obtain critical values for
contribution factors and their volatilities. We
require these critical values to enable us to make
statements with some confidence about
contribution factor behaviour. To enable this,
we need to examine contribution factor
behaviour where there is no structure in the
relevant data. This will enable us to benchmark
contribution factor behaviour over time. The
Brownian process is stochastic with increments
that are independent, stationary and normally
distributed. That is, we need random time series
so that contribution factors from these series can
be compared with those generated from series
that probably contain data that is highly
nonlinear and structured. This is the purpose of
the critical values discussed below. These are
our benchmarks for structure indentification.
We generated sets of four such time series and
used lags of three to forecast the fourth.
Contribution factors were collected from this
process to obtain a distribution of contribution
factor values and their volatilities. This
forecasting process is the same as that used in
Section 3 on real data.
2.3 Critical Values for Contribution
Factors and Their Volatilities
The simulation that we adopt assumes that asset
prices follow Brownian motion with drift,
represented as
dS
t
= ?S
t
dt + ?S
t
dW
t
(2.8)
where W
t
is a Wiener process, ? is the drift and
? the volatility of the process. An
approximation to this is:
?S = µS?t + ?S???t (2.9)
?t represents the time increment used and ? is a
random drawing from a standardised normal
distribution. Using the above concept, four sets
of random artificial time series were generated.
Lags of three series were then used to predict
the fourth. These simulations were repeated
sufficient times to obtain critical values for
contribution factors and their volatilities.
The upper 95% confidence level for a
contribution factor is 0.191. Thus, for the
random case, there is a 5% probability that a
contribution factor will be greater than or equal
to 0.191, given the parameters of the simulation.
A similar procedure to that above was
conducted for each simulation using instead the
volatility of the contribution factors as
previously defined at equation 2.3. A 95%
confidence level for volatility, equal to 0.25,
was determined.
D
o
w
n
l
o
a
d
e
d
b
y
P
O
N
D
I
C
H
E
R
R
Y
U
N
I
V
E
R
S
I
T
Y
A
t
2
1
:
0
0
2
4
J
a
n
u
a
r
y
2
0
1
6
(
P
T
)
ACCOUNTING RESEARCH JOURNAL VOLUME 19 NO 2 (2006)
144
The ANN software used in these simulations
was NeuroShell 2, see Ward Systems (1998).
Given the number of runs required by this
simulation exercise, a macro, Macro Express,
was used to generate a program to automate
running NeuroShell 2.
3. Data and Method
Data used are the daily exchange rates of the
Thai baht, Malaysian ringgit, Philippines peso
and Indonesian rupiah for the period July 1996
to February 1999. The corresponding equity
market proxies are the Stock Exchange of
Thailand Index (SET), the Kuala Lumpur
Composite Index (KLC), the Philippines
Composite Index (PCOMP) and the Jakarta
Composite Index (JCI). Data were obtained
from Bloomberg.
A forecast of each country’s exchange rate
against the $US was made one day ahead using
the one-day lags of each of the other countries
in the group. Thus, a one-day ahead forecast of
Indonesia’s exchange rate was made using the
one-day lag exchange rates for Malaysia,
Thailand and The Philippines, etc. Initially, the
ANN was trained on 110 vectors, tested on the
next ten and contribution factors recorded.
Cessation of training criteria were an average
training set error less than 0.0002, number of
epochs since minimum average training error
greater than 20,000, and number of test set
vector presentations since minimum average
error greater than 20,000. If any one of these
criteria was met, training ceased. A further ten
vectors were then added chronologically to the
training and testing set, and the process was
repeated until all data had been used. Time
series of contribution factors and their
volatilities were thus generated.
4. Results and Discussion
Figures 2 to 5 plot the contribution factors of the
independent variables, obtained from the
forecasts for each country’s exchange rate,
against time. The contribution factors for all
inputs for all forecasts are very stable pre-July
1997, which is what would be expected given
the then current pegging policies of these
currencies. The fact that the ANNs produce
contribution factors with very low volatility at
this point, as we would expect, helps to confirm
their validity for the use we make of them in
this study.
Post-July 1997, the volatility of the
contribution factors shown in Figures 6 to 9
increased and generally took on higher values.
In each case the 95% confidence level of 0.25
obtained in Section 2 has been exceeded. F-tests
conducted on the contribution factors of the
independent variables for each forecasting run
showed clearly that the volatilities of the
respective contribution factors before and after
the onset of the crisis are from different
populations.
Recalling our propositions from Section 2,
we see that they are substantiated. Figures 2 to 4
show that contribution factors are tending
towards their initial values, but are still
exhibiting increased volatility. This volatility is
understandable given that the relevant
currencies at that point were no longer pegged
to the US dollar. The case of Figure 5 is
discussed in the next section. A much larger
data set would probably be needed to
substantiate our fourth proposition.
We will now discuss in general the
implications of the above results with respect to
contagion between the exchange rates of the
four economies, and develop a method of
measuring changes in contagion based on
contribution factor differences. This is followed
by an examination of each of the economies in
turn using the method developed.
5. Correlation of Contagion
Measures and Concurrent
Economic Events
Differences in contribution factors of the ANN
inputs over time when forecasting each
exchange rate are shown in Figures 10 – 13. For
example, when forecasting the peso using
lagged inputs of the rupiah, baht and ringgit,
plots over time are given of the following
contribution factor differences: baht – ringgit,
baht – rupiah and ringgit – rupiah. Variations of
these differences illustrate the relative changes
in importance of each input over time. In each
case, prior to July 1997, contribution factors
exhibited very low volatility. Contribution
factor differences, as shown in figures 10 – 13,
were also small. This is not surprising, given
that the exchanges rates were effectively pegged
at that time.
D
o
w
n
l
o
a
d
e
d
b
y
P
O
N
D
I
C
H
E
R
R
Y
U
N
I
V
E
R
S
I
T
Y
A
t
2
1
:
0
0
2
4
J
a
n
u
a
r
y
2
0
1
6
(
P
T
)
Measuring Contagion in the South-East Asian Economic Crisis
145
Figure 2
Contribution Factors Forecasting One Day Ahead - The Philippines Peso
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
2
8
-
O
c
t
-
9
6
1
7
-
D
e
c
-
9
6
5
-
F
e
b
-
9
7
2
7
-
M
a
r
-
9
7
2
5
-
J
u
l
-
9
7
1
3
-
S
e
p
-
9
7
2
-
N
o
v
-
9
7
2
2
-
D
e
c
-
9
7
1
0
-
F
e
b
-
9
8
1
-
A
p
r
-
9
8
2
1
-
M
a
y
-
9
8
1
0
-
J
u
l
-
9
8
2
9
-
A
u
g
-
9
8
1
8
-
O
c
t
-
9
8
7
-
D
e
c
-
9
8
2
6
-
J
a
n
-
9
9
Date
C
o
n
t
r
i
b
u
t
i
o
n
F
a
c
t
o
r
Thai Baht
Malay Ringgit
Indon Rupiah
Figure 3
Contribution Factors Forecasting One Day Ahead - The Thai Baht
0
0.05
0.1
0.15
0.2
0.25
2
8
-
O
c
t
-
1
9
9
6
6
-
J
a
n
-
1
9
9
7
1
7
-
M
a
r
-
1
9
9
7
4
-
A
u
g
-
1
9
9
7
1
3
-
O
c
t
-
1
9
9
7
2
2
-
D
e
c
-
1
9
9
7
2
-
M
a
r
-
1
9
9
8
1
1
-
M
a
y
-
1
9
9
8
2
0
-
J
u
l
-
1
9
9
8
2
8
-
S
e
p
-
1
9
9
8
7
-
D
e
c
-
1
9
9
8
1
5
-
F
e
b
-
1
9
9
9
Date
C
o
n
t
r
i
b
u
t
i
o
n
F
a
c
t
o
r
Malay Ringgit
Philippines Peso
Indon Rupiah
Figure 4
Contribution Factors Forecasting One Day Ahead - The Malaysian Ringgit
0
0.05
0.1
0.15
0.2
0.25
2
8
-
O
c
t
-
1
9
9
6
6
-
J
a
n
-
1
9
9
7
1
7
-
M
a
r
-
1
9
9
7
4
-
A
u
g
-
1
9
9
7
1
3
-
O
c
t
-
1
9
9
7
2
2
-
D
e
c
-
1
9
9
7
2
-
M
a
r
-
1
9
9
8
1
1
-
M
a
y
-
1
9
9
8
2
0
-
J
u
l
-
1
9
9
8
2
8
-
S
e
p
-
1
9
9
8
7
-
D
e
c
-
1
9
9
8
1
5
-
F
e
b
-
1
9
9
9
Date
C
o
n
t
r
i
b
u
t
i
o
n
F
a
c
t
o
r
s
Thai Baht
Philippines Peso
Indon Rupiah
D
o
w
n
l
o
a
d
e
d
b
y
P
O
N
D
I
C
H
E
R
R
Y
U
N
I
V
E
R
S
I
T
Y
A
t
2
1
:
0
0
2
4
J
a
n
u
a
r
y
2
0
1
6
(
P
T
)
ACCOUNTING RESEARCH JOURNAL VOLUME 19 NO 2 (2006)
146
Figure 5
Contribution Factors Forecasting One Day Ahead - Indonesian Rupiah
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.18
0.2
2
8
-
O
c
t
-
1
9
9
6
6
-
J
a
n
-
1
9
9
7
1
7
-
M
a
r
-
1
9
9
7
4
-
A
u
g
-
1
9
9
7
1
3
-
O
c
t
-
1
9
9
7
2
2
-
D
e
c
-
1
9
9
7
2
-
M
a
r
-
1
9
9
8
1
1
-
M
a
y
-
1
9
9
8
2
0
-
J
u
l
-
1
9
9
8
2
8
-
S
e
p
-
1
9
9
8
7
-
D
e
c
-
1
9
9
8
1
5
-
F
e
b
-
1
9
9
9
Date
C
o
n
t
r
i
b
u
t
i
o
n
F
a
c
t
o
r
Thai Baht
Malay Ringgit
Philippines Peso
Figure 6
Volatilites Forecasting Philippines Peso
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
7
-
N
o
v
-
9
6
1
6
-
J
a
n
-
9
7
2
7
-
M
a
r
-
9
7
1
4
-
A
u
g
-
9
7
2
3
-
O
c
t
-
9
7
1
-
J
a
n
-
9
8
1
2
-
M
a
r
-
9
8
2
1
-
M
a
y
-
9
8
3
0
-
J
u
l
-
9
8
8
-
O
c
t
-
9
8
1
7
-
D
e
c
-
9
8
2
5
-
F
e
b
-
9
9
Date
V
o
l
a
t
i
l
i
t
yThai Baht
Malay Ringgit
Indon Rupiah
During the crisis, contribution factors
exhibited increased volatility. This was mirrored
by contribution factor differences. Each
contribution factor difference, at any given time
in a given forecast, gives a measure of the
importance of one exchange rate relative to one
other with respect to the forecast. Thus, whilst
forecasting the peso, if we examine the
difference of the baht contribution factor with
the ringgit’s shown Figure 10 at the end of July
1997, we can see that the importance of the baht
with respect to the ringgit has markedly
increased. These contribution factors have been
extracted from a forecast of the peso and
therefore demonstrate that the baht took on a
much higher level of importance in determining
the forecast at that time. This is an alternative
way of stating that the contagion from the baht
to the peso has increased at that point, which
can be repeated for other currencies.
Each forecast will now be discussed to match
some economic events to the behaviour of
relevant contribution factors. These economic
events are documented not only in the financial
press of the relevant time but in, for example,
Henderson (1998), Daly and Logan (1998) and
Roubini (1998).
6. The Philippines and the Peso
From Figure 2 we see that the contribution
factors for the baht, ringgit and rupiah are
virtually constant until the end of July 1997.
The baht has the higher contribution factor,
while the ringgit and rupiah are almost equal.
D
o
w
n
l
o
a
d
e
d
b
y
P
O
N
D
I
C
H
E
R
R
Y
U
N
I
V
E
R
S
I
T
Y
A
t
2
1
:
0
0
2
4
J
a
n
u
a
r
y
2
0
1
6
(
P
T
)
Measuring Contagion in the South-East Asian Economic Crisis
147
Figure 7
Volatilites Forecasting the Indonesian Rupiah
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
7
-
N
o
v
-
1
9
9
6
1
6
-
J
a
n
-
1
9
9
7
2
7
-
M
a
r
-
1
9
9
7
1
4
-
A
u
g
-
1
9
9
7
2
3
-
O
c
t
-
1
9
9
7
1
-
J
a
n
-
1
9
9
8
1
2
-
M
a
r
-
1
9
9
8
2
1
-
M
a
y
-
1
9
9
8
3
0
-
J
u
l
-
1
9
9
8
8
-
O
c
t
-
1
9
9
8
1
7
-
D
e
c
-
1
9
9
8
2
5
-
F
e
b
-
1
9
9
9
Date
V
o
l
a
t
i
l
i
t
y
Thai Baht
Malay Ringgit
Philippines Peso
Figure 8
Volatilities Forecasting the Malaysian Ringgit
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
7
-
N
o
v
-
1
9
9
6
1
6
-
J
a
n
-
1
9
9
7
2
7
-
M
a
r
-
1
9
9
7
1
4
-
A
u
g
-
1
9
9
7
2
3
-
O
c
t
-
1
9
9
7
1
-
J
a
n
-
1
9
9
8
1
2
-
M
a
r
-
1
9
9
8
2
1
-
M
a
y
-
1
9
9
8
3
0
-
J
u
l
-
1
9
9
8
8
-
O
c
t
-
1
9
9
8
1
7
-
D
e
c
-
1
9
9
8
2
5
-
F
e
b
-
1
9
9
9
Date
V
o
l
a
t
i
l
i
t
y
Thai Baht
Philippines Peso
Indon Rupiah
Figure 9
Volatilities Forecasting the Thai Baht
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
7
-
N
o
v
-
1
9
9
6
1
6
-
J
a
n
-
1
9
9
7
2
7
-
M
a
r
-
1
9
9
7
1
4
-
A
u
g
-
1
9
9
7
2
3
-
O
c
t
-
1
9
9
7
1
-
J
a
n
-
1
9
9
8
1
2
-
M
a
r
-
1
9
9
8
2
1
-
M
a
y
-
1
9
9
8
3
0
-
J
u
l
-
1
9
9
8
8
-
O
c
t
-
1
9
9
8
1
7
-
D
e
c
-
1
9
9
8
2
5
-
F
e
b
-
1
9
9
9
Date
V
o
l
a
t
i
l
i
t
y
Malay Ringgit
Philippines Peso
Indon Rupiah
D
o
w
n
l
o
a
d
e
d
b
y
P
O
N
D
I
C
H
E
R
R
Y
U
N
I
V
E
R
S
I
T
Y
A
t
2
1
:
0
0
2
4
J
a
n
u
a
r
y
2
0
1
6
(
P
T
)
ACCOUNTING RESEARCH JOURNAL VOLUME 19 NO 2 (2006)
148
After July 1997 the contribution factor for the
baht with respect to the ringgit rises rapidly
whilst that of the ringgit with respect to the
rupiah increases slightly before steadily rising to
a much higher difference by mid-October. The
difference between the contribution factors of
the baht and rupiah rises rapidly over August
1997. All the differences exhibit greatly
increased variation but over time this volatility
dampens and the differences all decrease
towards February 1999. How do we explain this
behaviour in financial and economic event
terms?
The onset of the crisis was marked by the
devaluation of the Thai baht at the beginning of
July 1997 after the Thai government had spent
billions unsuccessfully defending the pegged
currency. A rise of initially 13% in overnight
rates in The Philippines was made, probably to
counter spill-over effects from Thailand. This
was then increased in two moves to 24%.
The Philippines differed substantially from
Thailand. It did not have the same level of debt,
nor excessive “bubble” property projects and its
banking system was in far better order.
Nevertheless, on 11 July, The Philippines gave
up the defence of the peso, which then devalued
by more than 10% to the $US. This is captured
by the magnitude of the baht contribution factor
differences in Figure 10 demonstrating a
marked increase in the baht contribution factor.
Previous currency crises elsewhere had been
relatively short-lived and brought under control
reasonable quickly, which appeared to be the
initial expectation in this crisis and would help
to explain the large drop in contribution factor
differences not long after the sudden rises at the
start of the crisis. The contribution factors again
rose rapidly probably once it became apparent
that the problems of the region would not be
rectified in the short term.
A downward trend in contribution factor
differences then ensued for the remainder of the
period, punctuated with volatility marking
economic and financial shocks that were
transmitted throughout this period. The
uncertainty generated by these shocks could
account for this volatility and reflects the
changing degree of contagion felt by the peso
from the three other currencies.
Thailand cooperated with the IMF and
accepted its guidelines and conditions. Malaysia
at this time had large external borrowings, a
large current account deficit and similar
property investment problems to those of
Thailand. 30% of bank loans were to property
companies. The ringgit was devalued on 14 July
and appeared to prompt a domino effect of
currency devaluation driven by the necessity of
countries to maintain their export
competitiveness. Each country’s economic
viability was threatened in turn by the most
recent devaluations of the others.
Figure 10 captures the immediate effect of
the devaluations of the baht and ringgit on the
peso. Mahathir, the Malaysian President,
initially blamed international speculators and
hedge funds for his country’s demise and
rejected IMF recommendations, which may
account for the steady rise of the ringgit’s
Figure 10
Contribution Factor Differences: Forecasting the Peso
-0.1
-0.05
0
0.05
0.1
0.15
0.2
0.25
0.3
2
8
-
O
c
t
-
1
9
9
6
6
-
J
a
n
-
1
9
9
7
1
7
-
M
a
r
-
1
9
9
7
4
-
A
u
g
-
1
9
9
7
1
3
-
O
c
t
-
1
9
9
7
2
2
-
D
e
c
-
1
9
9
7
2
-
M
a
r
-
1
9
9
8
1
1
-
M
a
y
-
1
9
9
8
2
0
-
J
u
l
-
1
9
9
8
2
8
-
S
e
p
-
1
9
9
8
7
-
D
e
c
-
1
9
9
8
1
5
-
F
e
b
-
1
9
9
9
Date
C
o
n
t
r
i
b
u
t
i
o
n
f
a
c
t
o
r
D
i
f
f
e
r
e
n
c
e
baht-ringgit
baht-rupiah
ringgit-rupiah
D
o
w
n
l
o
a
d
e
d
b
y
P
O
N
D
I
C
H
E
R
R
Y
U
N
I
V
E
R
S
I
T
Y
A
t
2
1
:
0
0
2
4
J
a
n
u
a
r
y
2
0
1
6
(
P
T
)
Measuring Contagion in the South-East Asian Economic Crisis
149
Figure 11
Contribution Factor Differences: Forecasting the Baht
-0.15
-0.1
-0.05
0
0.05
0.1
0.15
0.2
2
8
-
O
c
t
-
1
9
9
6
6
-
J
a
n
-
1
9
9
7
1
7
-
M
a
r
-
1
9
9
7
4
-
A
u
g
-
1
9
9
7
1
3
-
O
c
t
-
1
9
9
7
2
2
-
D
e
c
-
1
9
9
7
2
-
M
a
r
-
1
9
9
8
1
1
-
M
a
y
-
1
9
9
8
2
0
-
J
u
l
-
1
9
9
8
2
8
-
S
e
p
-
1
9
9
8
7
-
D
e
c
-
1
9
9
8
1
5
-
F
e
b
-
1
9
9
9
Date
C
o
n
t
r
i
b
u
t
i
o
n
F
a
c
t
o
r
D
i
f
f
e
r
e
n
c
e
ringgit-peso
ringgit-rupiah
peso-rupiah
Figure 12
Contribution Factor Differences: Forecasting the Ringgit
-0.1
-0.05
0
0.05
0.1
0.15
0.2
2
8
-
O
c
t
-
1
9
9
6
6
-
J
a
n
-
1
9
9
7
1
7
-
M
a
r
-
1
9
9
7
4
-
A
u
g
-
1
9
9
7
1
3
-
O
c
t
-
1
9
9
7
2
2
-
D
e
c
-
1
9
9
7
2
-
M
a
r
-
1
9
9
8
1
1
-
M
a
y
-
1
9
9
8
2
0
-
J
u
l
-
1
9
9
8
2
8
-
S
e
p
-
1
9
9
8
7
-
D
e
c
-
1
9
9
8
1
5
-
F
e
b
-
1
9
9
9
Date
C
o
n
t
r
i
b
u
t
i
o
n
F
a
c
t
o
r
D
i
f
f
e
r
e
n
c
e
baht-peso
baht-rupiah
peso-rupiah
Figure 13
Contribution Factor Differences: Forecasting the Rupiah
-0.1
-0.05
0
0.05
0.1
0.15
2
8
-
O
c
t
-
1
9
9
6
6
-
J
a
n
-
1
9
9
7
1
7
-
M
a
r
-
1
9
9
7
4
-
A
u
g
-
1
9
9
7
1
3
-
O
c
t
-
1
9
9
7
2
2
-
D
e
c
-
1
9
9
7
2
-
M
a
r
-
1
9
9
8
1
1
-
M
a
y
-
1
9
9
8
2
0
-
J
u
l
-
1
9
9
8
2
8
-
S
e
p
-
1
9
9
8
7
-
D
e
c
-
1
9
9
8
1
5
-
F
e
b
-
1
9
9
9
Date
C
o
n
t
r
i
b
u
t
i
o
n
F
a
c
t
o
r
D
i
f
f
e
r
e
n
c
e
baht-ringgit
baht-peso
peso-ringgit
D
o
w
n
l
o
a
d
e
d
b
y
P
O
N
D
I
C
H
E
R
R
Y
U
N
I
V
E
R
S
I
T
Y
A
t
2
1
:
0
0
2
4
J
a
n
u
a
r
y
2
0
1
6
(
P
T
)
ACCOUNTING RESEARCH JOURNAL VOLUME 19 NO 2 (2006)
150
contribution factor in this forecast, from the
onset of the crisis until January 1998.
Meanwhile, Indonesia widened the trading band
of the rupiah on 11 July, effectively signalling
increased risk associated with the rupiah.
Indonesia devalued the rupiah on 14 August, by
which time its reserves had fallen by $US2.2
billion.
The contribution factor for the rupiah does
not exhibit the same level of volatility as the
other currencies in this forecast or in any of the
others. Generally, Indonesia suffered more and
longer from the effects of the crisis, and still
continues to do so. It depended more heavily on
foreign loans, and government agencies were,
unlike the other trigger economies, heavily
committed to borrowing. Banks were weak and
directed by government to whom they should
lend. Foreign debt was around 50% of GDP.
The economy, compared to the other countries,
was a closed one, with international trade
running at around 25% of GDP compared with
a regional norm of around 50%. Export growth
had also fallen significantly prior to the crisis
and could be regarded as the poor cousin of the
export growth economies of the region. Given
this, it is not surprising that its contribution
factor behaviour is the least relevant and
exhibits the least volatility in forecasting the
Philippines peso.
7. Malaysia and the Ringgit
Figures 4 and 12 show that the baht and peso
contribution factors in the ringgit forecast
appear correlated, with the peso being the more
dominant. Again, the rupiah’s contribution
factor indicates that the effect of the rupiah is
largely minimal. The initial changes in the
relevant contribution factors occur in mid-
August, 1997, roughly six weeks after the
devaluation of the baht.
The baht’s contribution factor increases with
respect to the rupiah at the onset of the crisis
and then drops back. One possible explanation,
already offered in the previous section, is that
there was an expectation that the crisis would be
dealt with swiftly. The baht’s contribution
factor rose again shortly thereafter, possibly due
to the realization that the crisis was more
profound and complex than initially anticipated,
and would take longer with more radical
remedial action to resolve. The peso’s
contribution factors largely replicate this
behaviour. Thereafter, they trend downward,
exhibiting volatility representing the financial
and economic shocks experienced throughout
this period.
Of the four countries, Malaysia recovered
more swiftly and more fully. Interestingly, it did
not accept or adopt any IMF guidelines in
dealing with the crisis. Capital controls were
also imposed at the end of August, 1998.
Stability appears to be returning around the start
of 1998. The contribution factors at that time
have returned to their pre-crisis levels.
However, further research could be conducted
to observe if these levels are sustained.
8. Thailand and the Baht
Contribution factors for the ringgit, rupiah and
peso shown in Figures 3 and 11 vary minimally
until the first week in September 1997 when the
peso and ringgit’s contribution factors increase
with respect to the rupiah’s. The rupiah appears
largely irrelevant. In November 1997 Chavalit,
the prime minister of Thailand, announced his
resignation, and the stock exchange briefly
made gains not seen for months. James
Wolfensohn, President of the World Bank,
9
announced at this time that the worst of the
currency crisis in South-East Asia was over. All
contribution factor differences reduce markedly
at this point as shown in Figure 11, indicating
that the contagion effects of these currencies
are, at least temporarily, greatly reduced. This
drop was followed by a subsequent rise for the
peso to nearly its former level, and for the
ringgit to a level much higher. Once again the
contribution factors from this point trend
downward, but continue to exhibit volatility and
reflect economic shocks. Towards the end of the
period this volatility reduces as shown in Figure
9, which may indicate a lessening of the impact
of the crisis, uncertainty diminishing as the
market learns to deal with the crisis, and some
remedial measures beginning to take effect.
The Indonesian rupiah’s contribution factor
shows an increase in volatility over this period
but is nowhere near as relevant as the other two
currencies. The reasons for this are the same as
those discussed in the case of the Philippines
peso.
9 See the Economist, 8 November 1997.
D
o
w
n
l
o
a
d
e
d
b
y
P
O
N
D
I
C
H
E
R
R
Y
U
N
I
V
E
R
S
I
T
Y
A
t
2
1
:
0
0
2
4
J
a
n
u
a
r
y
2
0
1
6
(
P
T
)
Measuring Contagion in the South-East Asian Economic Crisis
151
9. Indonesia and the Rupiah
Until mid-September, 1997, Figures 5 and 13
show almost constant contribution factors for
the baht, ringgit and peso in forecasting the
Indonesian rupiah. Given the exchange rate
policy until then, this was to be expected. Figure
13 shows that contribution factor differences
vary slightly, but much less than the other
examples above. In late December 1997 and
January 1998, the contribution factor of the peso
relative to the ringgit and baht shows a very
large increase. In December 1997 the ringgit,
baht and rupiah had fallen to their lowest level
ever against the $US. In particular, the ringgit
had depreciated by 33%. Malaysia’s stock
market had also fallen by more than 50%,
which may account for the ringgit’s increased
contribution factor at this time. Thereafter,
Figure 13 shows a reduction in all contribution
factor differences.
Other authors, such as Radelet and Sachs
(1998) and Goldstein, Kaminsky and Reinhart
(2000) concluded that Indonesia was the country
that suffered most from contagion. Our results do
not show this. Dungey and Martin (2004) found
that the contagion was largely de to shocks arising
from Thailand. Figures 2, 4 and 5, showing
contribution factor values for the Thai baht, and
Figures 6, 7, and 8 showing corresponding
contribution factor volatilities, support this, at
least in the early stages of the crisis.
10. Conclusion
A measure of the contagion between the
currencies of The Philippines, Malaysia,
Thailand and Indonesia during the 1997 crisis
has been developed using ANNs. Forecasts
were made of each country’s exchange rate
using one-day lags of the others as ANN inputs.
The contribution factors of the inputs were
collected whilst forecasts were made at ten-day
intervals from July 1996, to February 1999. The
volatilities of these contribution factors and
their differences with respect to each other give
a measure of the level of contagion between the
input currencies and the one being forecast.
Thus, a method of measuring changes in
contagion over time was developed. These
changes were successfully matched to economic
and financial shocks experienced during the
crisis.
Independent variables used here are liable to
be correlated. Examination of predictive errors
using restrictions on the independent variables
would be useful. For example, predictions could
be carried out using restricted sets of
independent variables.
References
B. Baker and C. Richards, A Comparison of Conventional
Linear Regression Methods and Neural Networks for
Forecasting Educational Spending, Economics of
Education Review, 18, 4, pp. 405–415, 1999.
T. Baig and I. Goldfajn, Financial Market Contagion in the
Asian Crisis, IMF Staff Papers, Vol. 46, No. 2, June,
1999.
R. Chang and A. Velasco, The Asian Liquidity Crisis,
Working Paper 6796, National bureau of Economic
Research, Cambridge, Mass., 1998.
G. M. Chen, M. Firth and O. M. Rui, Stock Market
Linkages: Evidence from Latin America, Journal of
Banking and Finance, Vol. 26, No. 6, 2002.
B. Chowdhry and A. Goyal, Understanding the financial
crisis in Asia, Pacific-Basin Finance Journal, vol. 8,
pp. 135–152, 2000.
D. Connor, Data Transformation Explains the Basics of
Neural Networks, EDN, pp. 138–144, May, 1988.
M.T. Daly and M.I. Logan, Reconstructing Asia, RMIT
University Press, Melbourne, 1998.
R. Dornbusch, Y. C. Park and S. Claessens, Contagion:
Understanding How it Spreads, The World bank
Research observer, Vol. 15, No. 2, pp. 177–197,
August, 2000.
A. Drazen, Political Contagion in Currency Crises, in M.
Dungey and D. Tambakis (eds), Identifying
International Financial Contagion: Progress and
Challenges, Cambridge University Press, 2005.
M. Dungey and V. Martin, A Multifactor Model of
Exchange Rates with Unanticipated Shocks: Measuring
Contagion in the East Asian Currency Crisis, Journal of
Emerging Market Finance, Vol. 3, No. 3, pp. 305–330,
December, 2004.
B. Eichengreen, A. K. Rose and C. Wyplosz, Contagious
Currency Crises, NBER Working Paper, 5681, 1997.
K. Forbes and R. Rigobon, No Contagion, Only
Interdependence: Measuring Stock Market Co-
Movements, Journal of Finance, Vol. 57, No. 5,
pp. 2223–2261, Oct. 2002.
R. Glick and A. Rose, Contagion and Trade: Why are
Currency Crises Regional?, NBER Working Paper
6806, National Bureau of Economic Research,
Cambridge, Mass., 1998.
M. Goldstein, G. Kaminsky and C. Reinhart, Assessing
Financial Vulnerability: An Early warning System for
Emerging Markets, Institute for International
Economics, Washington, 2000.
C. Henderson, Asia Falling, McGraw-Hill, New York, 1998.
R. G. Hoptroff, The Principles and Practice of Time Series
Forecasting and Business Modelling Using Neural Nets,
Neural Computing & Application, 1, pp. 59–66, 1993.
K. D. Jackson (Ed.), Asian Contagion, The Causes and
Consequences of a Financial Crisis, Westview Press,
Oxford, U.K., 1999.
D
o
w
n
l
o
a
d
e
d
b
y
P
O
N
D
I
C
H
E
R
R
Y
U
N
I
V
E
R
S
I
T
Y
A
t
2
1
:
0
0
2
4
J
a
n
u
a
r
y
2
0
1
6
(
P
T
)
ACCOUNTING RESEARCH JOURNAL VOLUME 19 NO 2 (2006)
152
S. Johnson, P. Boone, A. Breach and E. Friedman, Working
Paper, No. 137, Stockholm Institute of Transition
Economics and East European Economies, Stockholm,
September, 1998.
S. H. Kim, M. A. Kose and M. G. Plummer, Understanding
the Asian Contagion, Asian Economic Journal, Vol 15,
No 2, June, 2001.
T. Lee, H. White and C. W. J. Granger, Testing for
Neglected Nonlinearity in Time Series Models, Journal
of Economics, 56, pp. 269–290, 1993.
G.C. Lim and P.D. McNelis, The Effect of the Nikkei and
the S&P on the All-Ordinaries: A Comparison of Three
Models, International Journal of Finance and
Economics, 3, pp. 217–228, 1998.
R. P. Lippman, An Introduction to Computing with Neural
Nets, IEEE ASSP Magazine, pp. 4–22, April, 1987.
L. Marquez, T. Hill, R. Worthley and W. Remus, Neural
network Models as an Alternative to Regression, in R.R.
Trippi and E. Turban (eds), Neural Networks in Finance
and Investing, Probus Publishing, Chicago, IL, 1992.
F. S. Mishkin, Lessons from the Asian Crisis, Journal of
International Money and Finance, 18, 4, pp. 709–723,
August, 1999.
C. Nieh, The Effect of the Asian Financial Crisis on the
Relationships among Open Macroeconomic Factors for
Asian Countries, Applied Economics, Vol. 34, No. 4,
2003.
M. Pericoli and M. Sbracia, A Primer On Financial
Contagion, Journal of Economic Surveys, Vol. 17,
No. 4, 2003.
S. Radelet and J. Sachs, The East Asian Financial Crisis:
Diagnosis, Remedies, Prospects, Brookings Paper, Vol.
28, No. 1, pp. 1–74, 1998.
N. Roubini, Chronology of the Asian Crisis and its
Global Contagion, in Roubini’s homepage:
TUhttp://www.stern.nyu.edu/nroubiniUT, September 6
1998.
C. Scott, The South-East Asia Crisis, Neural networks and
Market Behaviour: An Exploratory Study, Review of
Pacific Basin Financial markets and Policies, Vol. 6,
No. 3, pp. 349–379, September, 2003.
J.A.R. Tan, Contagion Effects During the Asian Financial
Crisis: Some Evidence from Stock Price Data, Working
Paper PB98-06, Center for Pacific Basin Monetary and
Economic Studies, Economic Research Department,
Federal Bank of San Francisco, 1998.
Z. Tang, C. de Almeida and P.A. Fishwick, Time Series
Forecasting Using Neural Networks vs. Box-Jenkins
Methodology. Simulation, pp. 303–310, November
1991.
W. R. Thomson, The Asian Economic Crisis: Causes and
Current Status, The Asian Crisis: Economic Analysis
and Market Intelligence Conference, University of
Melbourne, Melbourne, May, 1998.
Ward Systems Group Inc., Fredrick M.D., 1998. (See web-
sitehttp://www.wardsystems.com/)
C. V. Van Rijckeghem and B. Weder, Sources of Contagion:
Is it Finance or Trade, Journal of international
Economics, 54, pp. 293–300, 2001.
B. Zhang, R. Timmis and W. Hu, A Neural Network Based
Pattern Recognition system for Somatic Embryos of
Douglas Fir, Plant Cell, tissue and Organ Culture, 56,
pp. 25–35, 1999.
J. Zhang, Developing Robust Non-Linear Models through
Bootstrap Aggregated Neural Networks,
Neurocomputing, 25, pp. 93–113, 1999.
D
o
w
n
l
o
a
d
e
d
b
y
P
O
N
D
I
C
H
E
R
R
Y
U
N
I
V
E
R
S
I
T
Y
A
t
2
1
:
0
0
2
4
J
a
n
u
a
r
y
2
0
1
6
(
P
T
)
doc_572277814.pdf
used to develop a new measure of contagion
using exchange rate data from the Asian Crisis
of 1997 and beyond. Connection weight
changes during retraining of networks used to
forecast exchange rates form the basis of this
measure. These weight changes are used in
obtaining a contribution factor for independent
variables used in a forecasting process.
Accounting Research Journal
Measuring Contagion in the South-East Asian Economic Crisis: An Exploration
Using Artificial Neural Networks
Callum Scott
Article information:
To cite this document:
Callum Scott, (2006),"Measuring Contagion in the South-East Asian Economic Crisis: An
Exploration Using Artificial Neural Networks", Accounting Research J ournal, Vol. 19 Iss 2 pp. 139 -
152
Permanent link to this document:http://dx.doi.org/10.1108/10309610680000684
Downloaded on: 24 January 2016, At: 21:00 (PT)
References: this document contains references to 0 other documents.
To copy this document: [email protected]
The fulltext of this document has been downloaded 401 times since 2006*
Users who downloaded this article also downloaded:
Freddie Choo, Kim Tan, (2006),"A Commentary on Sample Design Issues in Behavioral
Accounting Experiments", Accounting Research J ournal, Vol. 19 Iss 2 pp. 153-158 http://
dx.doi.org/10.1108/10309610680000685
Pak K. Auyeung, Ron Dagwell, Chew Ng, J ohn Sands, (2006),"Educators’ Epistemological Beliefs
of Accounting Ethics Teaching: A Cross-Cultural Study", Accounting Research J ournal, Vol. 19 Iss
2 pp. 122-138http://dx.doi.org/10.1108/10309610680000683
Yass A. Alkafaji, Nauzer Balsara, J udith N. Aburmishan, (2006),"FASB’s Statement No. 133 on
Derivatives and Barings Bank: The Case for Value at Risk (VAR)", Accounting Research J ournal,
Vol. 19 Iss 2 pp. 94-104http://dx.doi.org/10.1108/10309610680000681
Access to this document was granted through an Emerald subscription provided by emerald-
srm:115632 []
For Authors
If you would like to write for this, or any other Emerald publication, then please use our
Emerald for Authors service information about how to choose which publication to write for and
submission guidelines are available for all. Please visit www.emeraldinsight.com/authors for
more information.
About Emerald www.emeraldinsight.com
Emerald is a global publisher linking research and practice to the benefit of society. The
company manages a portfolio of more than 290 journals and over 2,350 books and book series
volumes, as well as providing an extensive range of online products and additional customer
resources and services.
Emerald is both COUNTER 4 and TRANSFER compliant. The organization is a partner of the
Committee on Publication Ethics (COPE) and also works with Portico and the LOCKSS initiative
for digital archive preservation.
D
o
w
n
l
o
a
d
e
d
b
y
P
O
N
D
I
C
H
E
R
R
Y
U
N
I
V
E
R
S
I
T
Y
A
t
2
1
:
0
0
2
4
J
a
n
u
a
r
y
2
0
1
6
(
P
T
)
*Related content and download information correct at time of
download.
D
o
w
n
l
o
a
d
e
d
b
y
P
O
N
D
I
C
H
E
R
R
Y
U
N
I
V
E
R
S
I
T
Y
A
t
2
1
:
0
0
2
4
J
a
n
u
a
r
y
2
0
1
6
(
P
T
)
Measuring Contagion in the South-East Asian Economic Crisis
139
Measuring Contagion in the South-East
Asian Economic Crisis: An Exploration
Using Artificial Neural Networks
Callum Scott
Department of Finance
University of Melbourne
Abstract
An artificial neural network methodology is
used to develop a new measure of contagion
using exchange rate data from the Asian Crisis
of 1997 and beyond. Connection weight
changes during retraining of networks used to
forecast exchange rates form the basis of this
measure. These weight changes are used in
obtaining a contribution factor for independent
variables used in a forecasting process.
Volatilities of contribution factors form the
basis of the measure of contagion obtained.
These volatilities are statistically validated
through a series of simulations where critical
values for them are derived. The measures of
contagion obtained are then matched to
concurrent economic and financial shocks that
occurred during the crisis. It is found that there
is good correlation between these events and the
contagion measures obtained.
1. Introduction
This paper proposes and assesses a novel
measure of contagion using an artificial neural
network (ANN) forecasting methodology. This
methodology is non-parametric. If ANNs are
used to forecast, contribution factors of
independent variables can be generated. A
contribution factor
1
can be thought of as a
measure of the efficacy, or “importance” of a
variable to the forecast. Thus if, for example, we
make a forecast of the exchange rate of
Thailand, using as one of the input variables a
lag of the exchange rate of Malaysia, the
contribution factor of the Malaysian exchange
rate can be regarded as a measure of the impact
of the Malaysian exchange rate on the exchange
rate of Thailand.
1 Contribution factors are fully defined in Section 2.
We propose that the size of these contribution
factors, their volatilities and their differences
2
are possible measures of contagion. Rather than
compare our results with other contagion
measures, we examine underlying economic
events that could impact on the relevant
economies to see if there is a correlation with the
values of contribution factors obtained for that
time. This is done in Section 5.
We use exchange rate data for four countries
that suffered massive economic downturns
during the Asian financial crisis. Obviously,
more economies suffered greatly then, but we
restrict our study to the chosen four to test our
proposal. This set could easily be expanded.
Several researchers have presented
explanations for the onset of the Asian crisis of
1997. A useful starting point in reviewing some
of these explanations is (Jackson 1999), where
several authors address causes and consequences
of the crisis with respect to several individual
economies of the region. Chowdhry and Goyal
(2000) highlight the effect of inadequate and lax
banking practices, suggesting recapitalization
of performing banks and the closure of
underperformers. Others include Roubini
(1998), Thomson (1998), Johnson, Boone,
Breach and Friedman (1998), Chang and
Velasco (1998), Mishkin (1999) and Radelet and
Sachs (1998). Proposed causes of the crisis
include short-term debt used to finance long-
term projects, an increasing emphasis on
projects of questionable value, moral hazard,
pegged exchange rates and political corruption.
However, it is likely that the cause of the event
2 For example, if three variables x, y, z affect a fourth, q,
then differences of contribution factors of x, y and z
through time, such as CF
x
- CF
y
where CF represents a
contribution factor, will chart the changing importance of
each variable relative to the other in its effect on q.
D
o
w
n
l
o
a
d
e
d
b
y
P
O
N
D
I
C
H
E
R
R
Y
U
N
I
V
E
R
S
I
T
Y
A
t
2
1
:
0
0
2
4
J
a
n
u
a
r
y
2
0
1
6
(
P
T
)
ACCOUNTING RESEARCH JOURNAL VOLUME 19 NO 2 (2006)
140
was a combination of such factors. Whatever
the cause, or causes, the accompanying
phenomenon of contagion had a devastating
effect on several of the economies in the region.
The process of contagion, consisting of
related devaluations in asset markets across
countries, reflects the possible transmission of
sentiment across those markets. One possible
definition of contagion in an economic or
financial context is a significant increase in
cross-market correlation during a period of crisis
and confusion. Pericoli and Sbracia (2003), in
their overview of financial contagion, offer five
representative definitions used in the literature
that are largely subsumed by the definition stated
above. They also present a useful summary of
the theory of contagion and a review of the
empirical literature. Dornbusch, Park and
Claessens (2000) define contagion as a
significant increase in cross-market linkages
resulting from some for of shock, pointing out
that during a crisis, the mode of transmission of
such shocks differs from stable times. Probably
the most frequently used test for contagion is
the threshold index model of Eichengreen,
Wyplosz and Rose (1997) and extended by Van
Rijckeghem and Weder (2001). ANNs and
contribution factors
3
are used in this study to
capture and quantify the contagion effect
resulting from shocks.
One possible measure of contagion could be
the comparing of covariance between stock
markets during a stable period with that
determined during a crisis. Other methods of
detecting and quantifying contagion have been
proposed, such as vector error correction
models, impulse response functions and variance
decomposition of VAR models, utilised by, for
example, Tan (1998). Baig and Goldfajn (1999)
used dummy variables and daily news to study
cross-border contagion in the trigger
4
economies
and Korea during the Asian crisis, while Drazen
(2005) examined political contagion during
the European Monetary System crisis of
1992-3. Glick and Rose (1998) addressed the
phenomenon of clustering in contagion in an
attempt to explain the regional nature of some
3 Contribution factors measure the importance of input
variables in an ANN forecasting system. They are fully
defined later in the next section.
4 The Philippines, Thailand, Malaysia and Indonesia.
contagious events using trade linkages. Forbes
and Rigobon (2002) tested for stock market co-
movement using data from the 1994 Mexican
peso crisis, the 1987 US stock market collapse
and the 1997 Asian crisis. They concluded that
in fact there was no real contagion, but a
continuation of the already existing cross-market
linkages. Kim, Kose and Plummer (2001) tested
two explanations of contagion during the Asia
crisis. These explanations were propagation of
adverse shocks and international investor
behaviour, along with symmetric problems in
domestic economic fundamentals.
Co-integration analysis was used by Chen,
Firth and Rui (2002) to investigate the
relationships between South American stock
markets, and Nieh (2002) examined the
relationships between exchange rate volatilities,
exports, imports and productivity for several
Asian economies. Finally, Daly (2003)
investigated the interdependence of several
South-East Asian stock markets before and after
the crisis. We however, adopt a new approach
using ANN contribution factors generated
through a forecasting process with data
consisting of exchange rates from the South-East
Asian crisis of 1997.
An ANN is a numerical algorithm that uses a
set of input data to find the best numerical
approximation to a set of output data. The ANN
consists of three main features. Firstly, it consists
of three layers, an input layer where input data is
received, a hidden layer where processing occurs
through a summation and activation function
5
,
and an output layer. Secondly, it consists of
connection weights placed on the links between
the input and hidden layers, and between the
hidden and output layers. All data passing across
these links are multiplied by the weight. Thirdly,
during the training process of an ANN, using
historic data, these weights change to optimize
the forecast. Training involves an iterative
process of forward and backward propagation
with an objective to minimize the error in
predicting the output variables. The literature
often describes this as a connectionist approach
5 Activation functions form part of the nodes that make up
the hidden or middle layer of an ANN. Usually, they are
a “squashing “ function that constrains values of the
node’s output to between 0 and 1 or -1 and +1. They add
non-linearity to the ANN. An example is the logistic
function that takes the form 1/(1+e
x
).
D
o
w
n
l
o
a
d
e
d
b
y
P
O
N
D
I
C
H
E
R
R
Y
U
N
I
V
E
R
S
I
T
Y
A
t
2
1
:
0
0
2
4
J
a
n
u
a
r
y
2
0
1
6
(
P
T
)
Measuring Contagion in the South-East Asian Economic Crisis
141
and has been widely applied to financial time
series that exhibit non-linearity that changes over
time.
In this study we use sigmoid activation
functions in the middle and output nodes of the
network. This introduces nonlinearity to the
process. Given that this nonlinearity is thus
introduced, this is reflected in the way
contribution factors change over time through
the retraining process used to generate time
series of contribution factors. Thus, although the
contribution factors, fully defined in the next
section, depend on connection weights that are
linear components, the back-propagation process
used in retraining ensures that these weights
change in a nonlinear way. This nonlinearity
therefore flows into the changes in the
contribution factors that are dependent on weight
changes.
Non-linearity and time variance has been
discussed and tested by researchers such as Lee,
White and Granger (1993) and Zhang (1999).
Hoptroff (1993) offers a practical discussion on
the merits of using ANNs in time series
forecasting covering several issues from a
practical trading viewpoint, including non-linear
modelling, robustness to noise, interpolation and
extrapolation.
Useful comparisons of the ANN approach
with more traditional methodologies have been
made by, for example, Marquez, Hill, Worthley
and Remus (1992). Lippman (1987) discussed
ANNs with respect to robustness and fault
tolerance, while Connor (1988) explored their
ability to discern correct variable
transformations and to handle outliers in data.
Many other studies have also been undertaken to
compare ANN performance with traditional
statistical methods, such as the comparison with
Box-Jenkins techniques made by Tang, de
Almeida and Fishwick (1991), Marquez et al
(1991) with regression methods and Baker and
Richard’s (1999) comparison with linear
regression methods. Generally, ANNs compared
well with these methods, indicating their
effectiveness in a nonlinear, time varying
application
6
. Lim and McNelis (1998) tested
predictability in daily stock price returns in
Australia, and used ANNs to examine the
6 Zhang, Timmis and Hu (1999) reviewed articles
comparing ANN forecasting with other techniques but
few business applications were examined.
response of these stocks to shocks in the
Japanese and US markets. Other models tested
were an autoregressive linear model and a non-
linear GARCH-M model. The ANN model
outperformed the GARCH-M model but not the
linear model.
The methodology used here differs from
those in the contagion literature discussed
earlier, and is non-parametric. No attempt has
been made to optimize the ANNs to improve
performance in forecasting as this is an
exploratory study. Further research could
however be carried out where different ANN
structure and design are tested. For example,
different activation functions could be assessed.
Data from the four economies that probably
suffered the most during the crisis are used
7
,
and a measure of the contagion of each on the
other is derived. Concurrent economic events
are examined to detect any correlation with
corresponding contribution factors.
The economies used here had close trade
links, financial links, and similar
macroeconomic policies. Rapid economic
growth driven mainly by export growth had
occurred in all of them. Most importantly, they
compete in the same export markets and rely on
the same sources of capital. Their currencies
were also all effectively pegged to the $US.
Devaluation of one currency would make that
country’s exports immediately more
competitive than the others, precipitating their
devaluation also. Thus, a cycle of competitive
de-valuations could be initiated.
The exchange rates of Indonesia, The
Philippines, Malaysia and Thailand exhibited a
sharp decline from July 1997 until February
1998 where they stabilize at a much lower
value. We therefore proceeded by making
forecasts of each country’s exchange rate (with
respect to the $US) in turn, using lagged values
of the other three. Forecasts were repeated
moving through time from 1995 – 99, and
adding to the data set used to retrain the ANNs.
Obviously, such a forecast is liable to be
improved with the inclusion of the lags of the
dependent variable as inputs, but this would
negate the collection of contribution factors that
related strictly to the exchange rates of the other
economies.
7 These are Indonesia, The Philippines, Malaysia and
Thailand.
D
o
w
n
l
o
a
d
e
d
b
y
P
O
N
D
I
C
H
E
R
R
Y
U
N
I
V
E
R
S
I
T
Y
A
t
2
1
:
0
0
2
4
J
a
n
u
a
r
y
2
0
1
6
(
P
T
)
ACCOUNTING RESEARCH JOURNAL VOLUME 19 NO 2 (2006)
142
Time series of contribution factors and their
volatilities for each exchange rate were
therefore generated during the forecasting
process. These series mirror the changing data
structure as we move from stability before the
crisis, into the disruption at its onset and
beyond. We now define contribution factors in
terms of ANN connection weights. This method
is that used by Ward Systems (1998) in their
NeuroShell 2 software that is used in this
application.
2. Artificial Neural Networks and
Contribution Factors
2.1 Defining a Contribution Factor
The schematic representation of an ANN is
given in Figure 1. We denote by W1(i,j) the
connection weight
8
connecting the i-th input
layer to the j-th hidden layer. We denote by
W2(j,k) the connection weight connecting the j-
th hidden layer and the k-th output layer. N
i
, N
j
and N
k
are the number of input nodes, hidden
layer nodes and output nodes respectively.
We denote by N1(j) the sum of the
connection weights for the j-th hidden node, that
is
N
i
N1(j) = ? W1(i, j)
i=1
and similarly for the k-th output node
N
j
N2(k) = ? W2(j, k)
j=1
N1(j) and N2(k) represent normalization
factors. That is, they enable the calculation of
the proportion of absolute total connection
weight attributable to any one input. For each
input, i, and each output, k:
N
j
CF
(i,k)
= ? W1(i, j)/N1(j) x W2(j, k)/N2(k) (2.1)
j=1
CF(i,k) in (2.1) represents the contribution of
the i-th input to the k-th output. We require the
total contribution of any input to all outputs.
Thus, for any input, i:
) , (
1
) ( 2
) ( k i CF
k
k N
i CF
=
? =
(2.2)
8 When we refer to a connection weight, we mean its
absolute value.
Thus, CF(i) represents the total contribution
factor of any input, i.
Figure 1
Simple three-layer back
propagation network
It was noted in the Introduction on page 141
that nonlinearity was introduced through the use
of sigmoid functions in the middle and output
nodes of the ANNs. Importantly, we should also
note that the contribution factors as defined
above are a function of the ANN connection
weights. Given the back-propagation
methodology used here to train the ANNs, and
the use of sigmoid functions, changes in the
connection weights during training will behave
in a nonlinear fashion. Our definition of
contribution factors aggregates the connection
weights of each independent variable, therefore
this nonlinearity will flow through to changes in
contribution factors.
2.2 Contribution Factors and Market
Shocks
Time series of contribution factor volatilities of
independent variables in a forecasting process,
and in particular their differences at any given
time, across the different variables used in the
forecasts form the basis of our measure of
contagion. Forecasts of each of the economies
exchange rates are made using input variables to
the ANN that consist of the other economies
lagged exchange rates. Thus, a forecast of the
exchange rate of Thailand would be made using
the lagged exchange rates of Indonesia, The
Philippines and Malaysia. This is repeated for
each economy in turn. We define contribution
factor volatility as
Volatility = ?
t
2
= [ln(CF(X)
t+1
/CF(X)
t
]
2
(2.3)
CF(X) = contribution factor of X
Input Layer
Hidden Layer
Output Layer
W1(i, j)
W2(j, k)
D
o
w
n
l
o
a
d
e
d
b
y
P
O
N
D
I
C
H
E
R
R
Y
U
N
I
V
E
R
S
I
T
Y
A
t
2
1
:
0
0
2
4
J
a
n
u
a
r
y
2
0
1
6
(
P
T
)
Measuring Contagion in the South-East Asian Economic Crisis
143
We now propose how contribution factors of
independent variables used in a forecasting
process would behave where the relevant
economies are subjected to unexpected shocks.
This builds on the introductory discussion of
contribution factor behaviour in Scott (2003).
Consider a market where learning, in the
EMH sense is complete. Assume a shock occurs
at time t = 0 and define pre-shock time as t < 0.
Thus post-shock time will be defined as t > 0.
The contribution factor of an independent
variable, CF(x), at t < 0 should be
approximately constant. Thus
CF(x)
(t 0)
> ?
2
CF(x) critical
(2.6)
When the effects of a shock have been
absorbed, contribution factors may revert to
their previous levels, or, if a permanent change
to market conditions has been established by the
shock, levels that differ from those exhibited
pre-shock:
CF(x)
pre-shock
= CF(x)
post-shock
or
CF(x)
pre-shock
? CF(x)
post-shock
(2.7)
We would therefore expect that contribution
factor volatility for each exchange rate should
be near zero before a divergent event such as
the Asian crisis, increasing at its onset and for
some time thereafter. Volatilities of these
contribution factors in independent variables are
liable to be correlated to economic events
during the crisis. Therefore a matching of
contribution factor behaviour to relevant
economic events was also carried out and is
explored in Section 5.
We now briefly discuss the simulations that
were performed to obtain critical values for
contribution factors and their volatilities. We
require these critical values to enable us to make
statements with some confidence about
contribution factor behaviour. To enable this,
we need to examine contribution factor
behaviour where there is no structure in the
relevant data. This will enable us to benchmark
contribution factor behaviour over time. The
Brownian process is stochastic with increments
that are independent, stationary and normally
distributed. That is, we need random time series
so that contribution factors from these series can
be compared with those generated from series
that probably contain data that is highly
nonlinear and structured. This is the purpose of
the critical values discussed below. These are
our benchmarks for structure indentification.
We generated sets of four such time series and
used lags of three to forecast the fourth.
Contribution factors were collected from this
process to obtain a distribution of contribution
factor values and their volatilities. This
forecasting process is the same as that used in
Section 3 on real data.
2.3 Critical Values for Contribution
Factors and Their Volatilities
The simulation that we adopt assumes that asset
prices follow Brownian motion with drift,
represented as
dS
t
= ?S
t
dt + ?S
t
dW
t
(2.8)
where W
t
is a Wiener process, ? is the drift and
? the volatility of the process. An
approximation to this is:
?S = µS?t + ?S???t (2.9)
?t represents the time increment used and ? is a
random drawing from a standardised normal
distribution. Using the above concept, four sets
of random artificial time series were generated.
Lags of three series were then used to predict
the fourth. These simulations were repeated
sufficient times to obtain critical values for
contribution factors and their volatilities.
The upper 95% confidence level for a
contribution factor is 0.191. Thus, for the
random case, there is a 5% probability that a
contribution factor will be greater than or equal
to 0.191, given the parameters of the simulation.
A similar procedure to that above was
conducted for each simulation using instead the
volatility of the contribution factors as
previously defined at equation 2.3. A 95%
confidence level for volatility, equal to 0.25,
was determined.
D
o
w
n
l
o
a
d
e
d
b
y
P
O
N
D
I
C
H
E
R
R
Y
U
N
I
V
E
R
S
I
T
Y
A
t
2
1
:
0
0
2
4
J
a
n
u
a
r
y
2
0
1
6
(
P
T
)
ACCOUNTING RESEARCH JOURNAL VOLUME 19 NO 2 (2006)
144
The ANN software used in these simulations
was NeuroShell 2, see Ward Systems (1998).
Given the number of runs required by this
simulation exercise, a macro, Macro Express,
was used to generate a program to automate
running NeuroShell 2.
3. Data and Method
Data used are the daily exchange rates of the
Thai baht, Malaysian ringgit, Philippines peso
and Indonesian rupiah for the period July 1996
to February 1999. The corresponding equity
market proxies are the Stock Exchange of
Thailand Index (SET), the Kuala Lumpur
Composite Index (KLC), the Philippines
Composite Index (PCOMP) and the Jakarta
Composite Index (JCI). Data were obtained
from Bloomberg.
A forecast of each country’s exchange rate
against the $US was made one day ahead using
the one-day lags of each of the other countries
in the group. Thus, a one-day ahead forecast of
Indonesia’s exchange rate was made using the
one-day lag exchange rates for Malaysia,
Thailand and The Philippines, etc. Initially, the
ANN was trained on 110 vectors, tested on the
next ten and contribution factors recorded.
Cessation of training criteria were an average
training set error less than 0.0002, number of
epochs since minimum average training error
greater than 20,000, and number of test set
vector presentations since minimum average
error greater than 20,000. If any one of these
criteria was met, training ceased. A further ten
vectors were then added chronologically to the
training and testing set, and the process was
repeated until all data had been used. Time
series of contribution factors and their
volatilities were thus generated.
4. Results and Discussion
Figures 2 to 5 plot the contribution factors of the
independent variables, obtained from the
forecasts for each country’s exchange rate,
against time. The contribution factors for all
inputs for all forecasts are very stable pre-July
1997, which is what would be expected given
the then current pegging policies of these
currencies. The fact that the ANNs produce
contribution factors with very low volatility at
this point, as we would expect, helps to confirm
their validity for the use we make of them in
this study.
Post-July 1997, the volatility of the
contribution factors shown in Figures 6 to 9
increased and generally took on higher values.
In each case the 95% confidence level of 0.25
obtained in Section 2 has been exceeded. F-tests
conducted on the contribution factors of the
independent variables for each forecasting run
showed clearly that the volatilities of the
respective contribution factors before and after
the onset of the crisis are from different
populations.
Recalling our propositions from Section 2,
we see that they are substantiated. Figures 2 to 4
show that contribution factors are tending
towards their initial values, but are still
exhibiting increased volatility. This volatility is
understandable given that the relevant
currencies at that point were no longer pegged
to the US dollar. The case of Figure 5 is
discussed in the next section. A much larger
data set would probably be needed to
substantiate our fourth proposition.
We will now discuss in general the
implications of the above results with respect to
contagion between the exchange rates of the
four economies, and develop a method of
measuring changes in contagion based on
contribution factor differences. This is followed
by an examination of each of the economies in
turn using the method developed.
5. Correlation of Contagion
Measures and Concurrent
Economic Events
Differences in contribution factors of the ANN
inputs over time when forecasting each
exchange rate are shown in Figures 10 – 13. For
example, when forecasting the peso using
lagged inputs of the rupiah, baht and ringgit,
plots over time are given of the following
contribution factor differences: baht – ringgit,
baht – rupiah and ringgit – rupiah. Variations of
these differences illustrate the relative changes
in importance of each input over time. In each
case, prior to July 1997, contribution factors
exhibited very low volatility. Contribution
factor differences, as shown in figures 10 – 13,
were also small. This is not surprising, given
that the exchanges rates were effectively pegged
at that time.
D
o
w
n
l
o
a
d
e
d
b
y
P
O
N
D
I
C
H
E
R
R
Y
U
N
I
V
E
R
S
I
T
Y
A
t
2
1
:
0
0
2
4
J
a
n
u
a
r
y
2
0
1
6
(
P
T
)
Measuring Contagion in the South-East Asian Economic Crisis
145
Figure 2
Contribution Factors Forecasting One Day Ahead - The Philippines Peso
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
2
8
-
O
c
t
-
9
6
1
7
-
D
e
c
-
9
6
5
-
F
e
b
-
9
7
2
7
-
M
a
r
-
9
7
2
5
-
J
u
l
-
9
7
1
3
-
S
e
p
-
9
7
2
-
N
o
v
-
9
7
2
2
-
D
e
c
-
9
7
1
0
-
F
e
b
-
9
8
1
-
A
p
r
-
9
8
2
1
-
M
a
y
-
9
8
1
0
-
J
u
l
-
9
8
2
9
-
A
u
g
-
9
8
1
8
-
O
c
t
-
9
8
7
-
D
e
c
-
9
8
2
6
-
J
a
n
-
9
9
Date
C
o
n
t
r
i
b
u
t
i
o
n
F
a
c
t
o
r
Thai Baht
Malay Ringgit
Indon Rupiah
Figure 3
Contribution Factors Forecasting One Day Ahead - The Thai Baht
0
0.05
0.1
0.15
0.2
0.25
2
8
-
O
c
t
-
1
9
9
6
6
-
J
a
n
-
1
9
9
7
1
7
-
M
a
r
-
1
9
9
7
4
-
A
u
g
-
1
9
9
7
1
3
-
O
c
t
-
1
9
9
7
2
2
-
D
e
c
-
1
9
9
7
2
-
M
a
r
-
1
9
9
8
1
1
-
M
a
y
-
1
9
9
8
2
0
-
J
u
l
-
1
9
9
8
2
8
-
S
e
p
-
1
9
9
8
7
-
D
e
c
-
1
9
9
8
1
5
-
F
e
b
-
1
9
9
9
Date
C
o
n
t
r
i
b
u
t
i
o
n
F
a
c
t
o
r
Malay Ringgit
Philippines Peso
Indon Rupiah
Figure 4
Contribution Factors Forecasting One Day Ahead - The Malaysian Ringgit
0
0.05
0.1
0.15
0.2
0.25
2
8
-
O
c
t
-
1
9
9
6
6
-
J
a
n
-
1
9
9
7
1
7
-
M
a
r
-
1
9
9
7
4
-
A
u
g
-
1
9
9
7
1
3
-
O
c
t
-
1
9
9
7
2
2
-
D
e
c
-
1
9
9
7
2
-
M
a
r
-
1
9
9
8
1
1
-
M
a
y
-
1
9
9
8
2
0
-
J
u
l
-
1
9
9
8
2
8
-
S
e
p
-
1
9
9
8
7
-
D
e
c
-
1
9
9
8
1
5
-
F
e
b
-
1
9
9
9
Date
C
o
n
t
r
i
b
u
t
i
o
n
F
a
c
t
o
r
s
Thai Baht
Philippines Peso
Indon Rupiah
D
o
w
n
l
o
a
d
e
d
b
y
P
O
N
D
I
C
H
E
R
R
Y
U
N
I
V
E
R
S
I
T
Y
A
t
2
1
:
0
0
2
4
J
a
n
u
a
r
y
2
0
1
6
(
P
T
)
ACCOUNTING RESEARCH JOURNAL VOLUME 19 NO 2 (2006)
146
Figure 5
Contribution Factors Forecasting One Day Ahead - Indonesian Rupiah
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.18
0.2
2
8
-
O
c
t
-
1
9
9
6
6
-
J
a
n
-
1
9
9
7
1
7
-
M
a
r
-
1
9
9
7
4
-
A
u
g
-
1
9
9
7
1
3
-
O
c
t
-
1
9
9
7
2
2
-
D
e
c
-
1
9
9
7
2
-
M
a
r
-
1
9
9
8
1
1
-
M
a
y
-
1
9
9
8
2
0
-
J
u
l
-
1
9
9
8
2
8
-
S
e
p
-
1
9
9
8
7
-
D
e
c
-
1
9
9
8
1
5
-
F
e
b
-
1
9
9
9
Date
C
o
n
t
r
i
b
u
t
i
o
n
F
a
c
t
o
r
Thai Baht
Malay Ringgit
Philippines Peso
Figure 6
Volatilites Forecasting Philippines Peso
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
7
-
N
o
v
-
9
6
1
6
-
J
a
n
-
9
7
2
7
-
M
a
r
-
9
7
1
4
-
A
u
g
-
9
7
2
3
-
O
c
t
-
9
7
1
-
J
a
n
-
9
8
1
2
-
M
a
r
-
9
8
2
1
-
M
a
y
-
9
8
3
0
-
J
u
l
-
9
8
8
-
O
c
t
-
9
8
1
7
-
D
e
c
-
9
8
2
5
-
F
e
b
-
9
9
Date
V
o
l
a
t
i
l
i
t
yThai Baht
Malay Ringgit
Indon Rupiah
During the crisis, contribution factors
exhibited increased volatility. This was mirrored
by contribution factor differences. Each
contribution factor difference, at any given time
in a given forecast, gives a measure of the
importance of one exchange rate relative to one
other with respect to the forecast. Thus, whilst
forecasting the peso, if we examine the
difference of the baht contribution factor with
the ringgit’s shown Figure 10 at the end of July
1997, we can see that the importance of the baht
with respect to the ringgit has markedly
increased. These contribution factors have been
extracted from a forecast of the peso and
therefore demonstrate that the baht took on a
much higher level of importance in determining
the forecast at that time. This is an alternative
way of stating that the contagion from the baht
to the peso has increased at that point, which
can be repeated for other currencies.
Each forecast will now be discussed to match
some economic events to the behaviour of
relevant contribution factors. These economic
events are documented not only in the financial
press of the relevant time but in, for example,
Henderson (1998), Daly and Logan (1998) and
Roubini (1998).
6. The Philippines and the Peso
From Figure 2 we see that the contribution
factors for the baht, ringgit and rupiah are
virtually constant until the end of July 1997.
The baht has the higher contribution factor,
while the ringgit and rupiah are almost equal.
D
o
w
n
l
o
a
d
e
d
b
y
P
O
N
D
I
C
H
E
R
R
Y
U
N
I
V
E
R
S
I
T
Y
A
t
2
1
:
0
0
2
4
J
a
n
u
a
r
y
2
0
1
6
(
P
T
)
Measuring Contagion in the South-East Asian Economic Crisis
147
Figure 7
Volatilites Forecasting the Indonesian Rupiah
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
7
-
N
o
v
-
1
9
9
6
1
6
-
J
a
n
-
1
9
9
7
2
7
-
M
a
r
-
1
9
9
7
1
4
-
A
u
g
-
1
9
9
7
2
3
-
O
c
t
-
1
9
9
7
1
-
J
a
n
-
1
9
9
8
1
2
-
M
a
r
-
1
9
9
8
2
1
-
M
a
y
-
1
9
9
8
3
0
-
J
u
l
-
1
9
9
8
8
-
O
c
t
-
1
9
9
8
1
7
-
D
e
c
-
1
9
9
8
2
5
-
F
e
b
-
1
9
9
9
Date
V
o
l
a
t
i
l
i
t
y
Thai Baht
Malay Ringgit
Philippines Peso
Figure 8
Volatilities Forecasting the Malaysian Ringgit
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
7
-
N
o
v
-
1
9
9
6
1
6
-
J
a
n
-
1
9
9
7
2
7
-
M
a
r
-
1
9
9
7
1
4
-
A
u
g
-
1
9
9
7
2
3
-
O
c
t
-
1
9
9
7
1
-
J
a
n
-
1
9
9
8
1
2
-
M
a
r
-
1
9
9
8
2
1
-
M
a
y
-
1
9
9
8
3
0
-
J
u
l
-
1
9
9
8
8
-
O
c
t
-
1
9
9
8
1
7
-
D
e
c
-
1
9
9
8
2
5
-
F
e
b
-
1
9
9
9
Date
V
o
l
a
t
i
l
i
t
y
Thai Baht
Philippines Peso
Indon Rupiah
Figure 9
Volatilities Forecasting the Thai Baht
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
7
-
N
o
v
-
1
9
9
6
1
6
-
J
a
n
-
1
9
9
7
2
7
-
M
a
r
-
1
9
9
7
1
4
-
A
u
g
-
1
9
9
7
2
3
-
O
c
t
-
1
9
9
7
1
-
J
a
n
-
1
9
9
8
1
2
-
M
a
r
-
1
9
9
8
2
1
-
M
a
y
-
1
9
9
8
3
0
-
J
u
l
-
1
9
9
8
8
-
O
c
t
-
1
9
9
8
1
7
-
D
e
c
-
1
9
9
8
2
5
-
F
e
b
-
1
9
9
9
Date
V
o
l
a
t
i
l
i
t
y
Malay Ringgit
Philippines Peso
Indon Rupiah
D
o
w
n
l
o
a
d
e
d
b
y
P
O
N
D
I
C
H
E
R
R
Y
U
N
I
V
E
R
S
I
T
Y
A
t
2
1
:
0
0
2
4
J
a
n
u
a
r
y
2
0
1
6
(
P
T
)
ACCOUNTING RESEARCH JOURNAL VOLUME 19 NO 2 (2006)
148
After July 1997 the contribution factor for the
baht with respect to the ringgit rises rapidly
whilst that of the ringgit with respect to the
rupiah increases slightly before steadily rising to
a much higher difference by mid-October. The
difference between the contribution factors of
the baht and rupiah rises rapidly over August
1997. All the differences exhibit greatly
increased variation but over time this volatility
dampens and the differences all decrease
towards February 1999. How do we explain this
behaviour in financial and economic event
terms?
The onset of the crisis was marked by the
devaluation of the Thai baht at the beginning of
July 1997 after the Thai government had spent
billions unsuccessfully defending the pegged
currency. A rise of initially 13% in overnight
rates in The Philippines was made, probably to
counter spill-over effects from Thailand. This
was then increased in two moves to 24%.
The Philippines differed substantially from
Thailand. It did not have the same level of debt,
nor excessive “bubble” property projects and its
banking system was in far better order.
Nevertheless, on 11 July, The Philippines gave
up the defence of the peso, which then devalued
by more than 10% to the $US. This is captured
by the magnitude of the baht contribution factor
differences in Figure 10 demonstrating a
marked increase in the baht contribution factor.
Previous currency crises elsewhere had been
relatively short-lived and brought under control
reasonable quickly, which appeared to be the
initial expectation in this crisis and would help
to explain the large drop in contribution factor
differences not long after the sudden rises at the
start of the crisis. The contribution factors again
rose rapidly probably once it became apparent
that the problems of the region would not be
rectified in the short term.
A downward trend in contribution factor
differences then ensued for the remainder of the
period, punctuated with volatility marking
economic and financial shocks that were
transmitted throughout this period. The
uncertainty generated by these shocks could
account for this volatility and reflects the
changing degree of contagion felt by the peso
from the three other currencies.
Thailand cooperated with the IMF and
accepted its guidelines and conditions. Malaysia
at this time had large external borrowings, a
large current account deficit and similar
property investment problems to those of
Thailand. 30% of bank loans were to property
companies. The ringgit was devalued on 14 July
and appeared to prompt a domino effect of
currency devaluation driven by the necessity of
countries to maintain their export
competitiveness. Each country’s economic
viability was threatened in turn by the most
recent devaluations of the others.
Figure 10 captures the immediate effect of
the devaluations of the baht and ringgit on the
peso. Mahathir, the Malaysian President,
initially blamed international speculators and
hedge funds for his country’s demise and
rejected IMF recommendations, which may
account for the steady rise of the ringgit’s
Figure 10
Contribution Factor Differences: Forecasting the Peso
-0.1
-0.05
0
0.05
0.1
0.15
0.2
0.25
0.3
2
8
-
O
c
t
-
1
9
9
6
6
-
J
a
n
-
1
9
9
7
1
7
-
M
a
r
-
1
9
9
7
4
-
A
u
g
-
1
9
9
7
1
3
-
O
c
t
-
1
9
9
7
2
2
-
D
e
c
-
1
9
9
7
2
-
M
a
r
-
1
9
9
8
1
1
-
M
a
y
-
1
9
9
8
2
0
-
J
u
l
-
1
9
9
8
2
8
-
S
e
p
-
1
9
9
8
7
-
D
e
c
-
1
9
9
8
1
5
-
F
e
b
-
1
9
9
9
Date
C
o
n
t
r
i
b
u
t
i
o
n
f
a
c
t
o
r
D
i
f
f
e
r
e
n
c
e
baht-ringgit
baht-rupiah
ringgit-rupiah
D
o
w
n
l
o
a
d
e
d
b
y
P
O
N
D
I
C
H
E
R
R
Y
U
N
I
V
E
R
S
I
T
Y
A
t
2
1
:
0
0
2
4
J
a
n
u
a
r
y
2
0
1
6
(
P
T
)
Measuring Contagion in the South-East Asian Economic Crisis
149
Figure 11
Contribution Factor Differences: Forecasting the Baht
-0.15
-0.1
-0.05
0
0.05
0.1
0.15
0.2
2
8
-
O
c
t
-
1
9
9
6
6
-
J
a
n
-
1
9
9
7
1
7
-
M
a
r
-
1
9
9
7
4
-
A
u
g
-
1
9
9
7
1
3
-
O
c
t
-
1
9
9
7
2
2
-
D
e
c
-
1
9
9
7
2
-
M
a
r
-
1
9
9
8
1
1
-
M
a
y
-
1
9
9
8
2
0
-
J
u
l
-
1
9
9
8
2
8
-
S
e
p
-
1
9
9
8
7
-
D
e
c
-
1
9
9
8
1
5
-
F
e
b
-
1
9
9
9
Date
C
o
n
t
r
i
b
u
t
i
o
n
F
a
c
t
o
r
D
i
f
f
e
r
e
n
c
e
ringgit-peso
ringgit-rupiah
peso-rupiah
Figure 12
Contribution Factor Differences: Forecasting the Ringgit
-0.1
-0.05
0
0.05
0.1
0.15
0.2
2
8
-
O
c
t
-
1
9
9
6
6
-
J
a
n
-
1
9
9
7
1
7
-
M
a
r
-
1
9
9
7
4
-
A
u
g
-
1
9
9
7
1
3
-
O
c
t
-
1
9
9
7
2
2
-
D
e
c
-
1
9
9
7
2
-
M
a
r
-
1
9
9
8
1
1
-
M
a
y
-
1
9
9
8
2
0
-
J
u
l
-
1
9
9
8
2
8
-
S
e
p
-
1
9
9
8
7
-
D
e
c
-
1
9
9
8
1
5
-
F
e
b
-
1
9
9
9
Date
C
o
n
t
r
i
b
u
t
i
o
n
F
a
c
t
o
r
D
i
f
f
e
r
e
n
c
e
baht-peso
baht-rupiah
peso-rupiah
Figure 13
Contribution Factor Differences: Forecasting the Rupiah
-0.1
-0.05
0
0.05
0.1
0.15
2
8
-
O
c
t
-
1
9
9
6
6
-
J
a
n
-
1
9
9
7
1
7
-
M
a
r
-
1
9
9
7
4
-
A
u
g
-
1
9
9
7
1
3
-
O
c
t
-
1
9
9
7
2
2
-
D
e
c
-
1
9
9
7
2
-
M
a
r
-
1
9
9
8
1
1
-
M
a
y
-
1
9
9
8
2
0
-
J
u
l
-
1
9
9
8
2
8
-
S
e
p
-
1
9
9
8
7
-
D
e
c
-
1
9
9
8
1
5
-
F
e
b
-
1
9
9
9
Date
C
o
n
t
r
i
b
u
t
i
o
n
F
a
c
t
o
r
D
i
f
f
e
r
e
n
c
e
baht-ringgit
baht-peso
peso-ringgit
D
o
w
n
l
o
a
d
e
d
b
y
P
O
N
D
I
C
H
E
R
R
Y
U
N
I
V
E
R
S
I
T
Y
A
t
2
1
:
0
0
2
4
J
a
n
u
a
r
y
2
0
1
6
(
P
T
)
ACCOUNTING RESEARCH JOURNAL VOLUME 19 NO 2 (2006)
150
contribution factor in this forecast, from the
onset of the crisis until January 1998.
Meanwhile, Indonesia widened the trading band
of the rupiah on 11 July, effectively signalling
increased risk associated with the rupiah.
Indonesia devalued the rupiah on 14 August, by
which time its reserves had fallen by $US2.2
billion.
The contribution factor for the rupiah does
not exhibit the same level of volatility as the
other currencies in this forecast or in any of the
others. Generally, Indonesia suffered more and
longer from the effects of the crisis, and still
continues to do so. It depended more heavily on
foreign loans, and government agencies were,
unlike the other trigger economies, heavily
committed to borrowing. Banks were weak and
directed by government to whom they should
lend. Foreign debt was around 50% of GDP.
The economy, compared to the other countries,
was a closed one, with international trade
running at around 25% of GDP compared with
a regional norm of around 50%. Export growth
had also fallen significantly prior to the crisis
and could be regarded as the poor cousin of the
export growth economies of the region. Given
this, it is not surprising that its contribution
factor behaviour is the least relevant and
exhibits the least volatility in forecasting the
Philippines peso.
7. Malaysia and the Ringgit
Figures 4 and 12 show that the baht and peso
contribution factors in the ringgit forecast
appear correlated, with the peso being the more
dominant. Again, the rupiah’s contribution
factor indicates that the effect of the rupiah is
largely minimal. The initial changes in the
relevant contribution factors occur in mid-
August, 1997, roughly six weeks after the
devaluation of the baht.
The baht’s contribution factor increases with
respect to the rupiah at the onset of the crisis
and then drops back. One possible explanation,
already offered in the previous section, is that
there was an expectation that the crisis would be
dealt with swiftly. The baht’s contribution
factor rose again shortly thereafter, possibly due
to the realization that the crisis was more
profound and complex than initially anticipated,
and would take longer with more radical
remedial action to resolve. The peso’s
contribution factors largely replicate this
behaviour. Thereafter, they trend downward,
exhibiting volatility representing the financial
and economic shocks experienced throughout
this period.
Of the four countries, Malaysia recovered
more swiftly and more fully. Interestingly, it did
not accept or adopt any IMF guidelines in
dealing with the crisis. Capital controls were
also imposed at the end of August, 1998.
Stability appears to be returning around the start
of 1998. The contribution factors at that time
have returned to their pre-crisis levels.
However, further research could be conducted
to observe if these levels are sustained.
8. Thailand and the Baht
Contribution factors for the ringgit, rupiah and
peso shown in Figures 3 and 11 vary minimally
until the first week in September 1997 when the
peso and ringgit’s contribution factors increase
with respect to the rupiah’s. The rupiah appears
largely irrelevant. In November 1997 Chavalit,
the prime minister of Thailand, announced his
resignation, and the stock exchange briefly
made gains not seen for months. James
Wolfensohn, President of the World Bank,
9
announced at this time that the worst of the
currency crisis in South-East Asia was over. All
contribution factor differences reduce markedly
at this point as shown in Figure 11, indicating
that the contagion effects of these currencies
are, at least temporarily, greatly reduced. This
drop was followed by a subsequent rise for the
peso to nearly its former level, and for the
ringgit to a level much higher. Once again the
contribution factors from this point trend
downward, but continue to exhibit volatility and
reflect economic shocks. Towards the end of the
period this volatility reduces as shown in Figure
9, which may indicate a lessening of the impact
of the crisis, uncertainty diminishing as the
market learns to deal with the crisis, and some
remedial measures beginning to take effect.
The Indonesian rupiah’s contribution factor
shows an increase in volatility over this period
but is nowhere near as relevant as the other two
currencies. The reasons for this are the same as
those discussed in the case of the Philippines
peso.
9 See the Economist, 8 November 1997.
D
o
w
n
l
o
a
d
e
d
b
y
P
O
N
D
I
C
H
E
R
R
Y
U
N
I
V
E
R
S
I
T
Y
A
t
2
1
:
0
0
2
4
J
a
n
u
a
r
y
2
0
1
6
(
P
T
)
Measuring Contagion in the South-East Asian Economic Crisis
151
9. Indonesia and the Rupiah
Until mid-September, 1997, Figures 5 and 13
show almost constant contribution factors for
the baht, ringgit and peso in forecasting the
Indonesian rupiah. Given the exchange rate
policy until then, this was to be expected. Figure
13 shows that contribution factor differences
vary slightly, but much less than the other
examples above. In late December 1997 and
January 1998, the contribution factor of the peso
relative to the ringgit and baht shows a very
large increase. In December 1997 the ringgit,
baht and rupiah had fallen to their lowest level
ever against the $US. In particular, the ringgit
had depreciated by 33%. Malaysia’s stock
market had also fallen by more than 50%,
which may account for the ringgit’s increased
contribution factor at this time. Thereafter,
Figure 13 shows a reduction in all contribution
factor differences.
Other authors, such as Radelet and Sachs
(1998) and Goldstein, Kaminsky and Reinhart
(2000) concluded that Indonesia was the country
that suffered most from contagion. Our results do
not show this. Dungey and Martin (2004) found
that the contagion was largely de to shocks arising
from Thailand. Figures 2, 4 and 5, showing
contribution factor values for the Thai baht, and
Figures 6, 7, and 8 showing corresponding
contribution factor volatilities, support this, at
least in the early stages of the crisis.
10. Conclusion
A measure of the contagion between the
currencies of The Philippines, Malaysia,
Thailand and Indonesia during the 1997 crisis
has been developed using ANNs. Forecasts
were made of each country’s exchange rate
using one-day lags of the others as ANN inputs.
The contribution factors of the inputs were
collected whilst forecasts were made at ten-day
intervals from July 1996, to February 1999. The
volatilities of these contribution factors and
their differences with respect to each other give
a measure of the level of contagion between the
input currencies and the one being forecast.
Thus, a method of measuring changes in
contagion over time was developed. These
changes were successfully matched to economic
and financial shocks experienced during the
crisis.
Independent variables used here are liable to
be correlated. Examination of predictive errors
using restrictions on the independent variables
would be useful. For example, predictions could
be carried out using restricted sets of
independent variables.
References
B. Baker and C. Richards, A Comparison of Conventional
Linear Regression Methods and Neural Networks for
Forecasting Educational Spending, Economics of
Education Review, 18, 4, pp. 405–415, 1999.
T. Baig and I. Goldfajn, Financial Market Contagion in the
Asian Crisis, IMF Staff Papers, Vol. 46, No. 2, June,
1999.
R. Chang and A. Velasco, The Asian Liquidity Crisis,
Working Paper 6796, National bureau of Economic
Research, Cambridge, Mass., 1998.
G. M. Chen, M. Firth and O. M. Rui, Stock Market
Linkages: Evidence from Latin America, Journal of
Banking and Finance, Vol. 26, No. 6, 2002.
B. Chowdhry and A. Goyal, Understanding the financial
crisis in Asia, Pacific-Basin Finance Journal, vol. 8,
pp. 135–152, 2000.
D. Connor, Data Transformation Explains the Basics of
Neural Networks, EDN, pp. 138–144, May, 1988.
M.T. Daly and M.I. Logan, Reconstructing Asia, RMIT
University Press, Melbourne, 1998.
R. Dornbusch, Y. C. Park and S. Claessens, Contagion:
Understanding How it Spreads, The World bank
Research observer, Vol. 15, No. 2, pp. 177–197,
August, 2000.
A. Drazen, Political Contagion in Currency Crises, in M.
Dungey and D. Tambakis (eds), Identifying
International Financial Contagion: Progress and
Challenges, Cambridge University Press, 2005.
M. Dungey and V. Martin, A Multifactor Model of
Exchange Rates with Unanticipated Shocks: Measuring
Contagion in the East Asian Currency Crisis, Journal of
Emerging Market Finance, Vol. 3, No. 3, pp. 305–330,
December, 2004.
B. Eichengreen, A. K. Rose and C. Wyplosz, Contagious
Currency Crises, NBER Working Paper, 5681, 1997.
K. Forbes and R. Rigobon, No Contagion, Only
Interdependence: Measuring Stock Market Co-
Movements, Journal of Finance, Vol. 57, No. 5,
pp. 2223–2261, Oct. 2002.
R. Glick and A. Rose, Contagion and Trade: Why are
Currency Crises Regional?, NBER Working Paper
6806, National Bureau of Economic Research,
Cambridge, Mass., 1998.
M. Goldstein, G. Kaminsky and C. Reinhart, Assessing
Financial Vulnerability: An Early warning System for
Emerging Markets, Institute for International
Economics, Washington, 2000.
C. Henderson, Asia Falling, McGraw-Hill, New York, 1998.
R. G. Hoptroff, The Principles and Practice of Time Series
Forecasting and Business Modelling Using Neural Nets,
Neural Computing & Application, 1, pp. 59–66, 1993.
K. D. Jackson (Ed.), Asian Contagion, The Causes and
Consequences of a Financial Crisis, Westview Press,
Oxford, U.K., 1999.
D
o
w
n
l
o
a
d
e
d
b
y
P
O
N
D
I
C
H
E
R
R
Y
U
N
I
V
E
R
S
I
T
Y
A
t
2
1
:
0
0
2
4
J
a
n
u
a
r
y
2
0
1
6
(
P
T
)
ACCOUNTING RESEARCH JOURNAL VOLUME 19 NO 2 (2006)
152
S. Johnson, P. Boone, A. Breach and E. Friedman, Working
Paper, No. 137, Stockholm Institute of Transition
Economics and East European Economies, Stockholm,
September, 1998.
S. H. Kim, M. A. Kose and M. G. Plummer, Understanding
the Asian Contagion, Asian Economic Journal, Vol 15,
No 2, June, 2001.
T. Lee, H. White and C. W. J. Granger, Testing for
Neglected Nonlinearity in Time Series Models, Journal
of Economics, 56, pp. 269–290, 1993.
G.C. Lim and P.D. McNelis, The Effect of the Nikkei and
the S&P on the All-Ordinaries: A Comparison of Three
Models, International Journal of Finance and
Economics, 3, pp. 217–228, 1998.
R. P. Lippman, An Introduction to Computing with Neural
Nets, IEEE ASSP Magazine, pp. 4–22, April, 1987.
L. Marquez, T. Hill, R. Worthley and W. Remus, Neural
network Models as an Alternative to Regression, in R.R.
Trippi and E. Turban (eds), Neural Networks in Finance
and Investing, Probus Publishing, Chicago, IL, 1992.
F. S. Mishkin, Lessons from the Asian Crisis, Journal of
International Money and Finance, 18, 4, pp. 709–723,
August, 1999.
C. Nieh, The Effect of the Asian Financial Crisis on the
Relationships among Open Macroeconomic Factors for
Asian Countries, Applied Economics, Vol. 34, No. 4,
2003.
M. Pericoli and M. Sbracia, A Primer On Financial
Contagion, Journal of Economic Surveys, Vol. 17,
No. 4, 2003.
S. Radelet and J. Sachs, The East Asian Financial Crisis:
Diagnosis, Remedies, Prospects, Brookings Paper, Vol.
28, No. 1, pp. 1–74, 1998.
N. Roubini, Chronology of the Asian Crisis and its
Global Contagion, in Roubini’s homepage:
TUhttp://www.stern.nyu.edu/nroubiniUT, September 6
1998.
C. Scott, The South-East Asia Crisis, Neural networks and
Market Behaviour: An Exploratory Study, Review of
Pacific Basin Financial markets and Policies, Vol. 6,
No. 3, pp. 349–379, September, 2003.
J.A.R. Tan, Contagion Effects During the Asian Financial
Crisis: Some Evidence from Stock Price Data, Working
Paper PB98-06, Center for Pacific Basin Monetary and
Economic Studies, Economic Research Department,
Federal Bank of San Francisco, 1998.
Z. Tang, C. de Almeida and P.A. Fishwick, Time Series
Forecasting Using Neural Networks vs. Box-Jenkins
Methodology. Simulation, pp. 303–310, November
1991.
W. R. Thomson, The Asian Economic Crisis: Causes and
Current Status, The Asian Crisis: Economic Analysis
and Market Intelligence Conference, University of
Melbourne, Melbourne, May, 1998.
Ward Systems Group Inc., Fredrick M.D., 1998. (See web-
sitehttp://www.wardsystems.com/)
C. V. Van Rijckeghem and B. Weder, Sources of Contagion:
Is it Finance or Trade, Journal of international
Economics, 54, pp. 293–300, 2001.
B. Zhang, R. Timmis and W. Hu, A Neural Network Based
Pattern Recognition system for Somatic Embryos of
Douglas Fir, Plant Cell, tissue and Organ Culture, 56,
pp. 25–35, 1999.
J. Zhang, Developing Robust Non-Linear Models through
Bootstrap Aggregated Neural Networks,
Neurocomputing, 25, pp. 93–113, 1999.
D
o
w
n
l
o
a
d
e
d
b
y
P
O
N
D
I
C
H
E
R
R
Y
U
N
I
V
E
R
S
I
T
Y
A
t
2
1
:
0
0
2
4
J
a
n
u
a
r
y
2
0
1
6
(
P
T
)
doc_572277814.pdf