Physica A 390 (2011) 707–718

Contents lists available at ScienceDirect

Physica A
journal homepage: www.elsevier.com/locate/physa

Currency crises and the evolution of foreign exchange market: Evidence

from minimum spanning tree
Wooseok Jang a , Junghoon Lee a , Woojin Chang b,∗
a
Technology Management Economics and Policy Program, Seoul National University, Republic of Korea
b
Department of Industrial Engineering, Seoul National University, 599 Gwanak-ro, Gwanak-gu, Seoul 151-742, Republic of Korea

article info abstract

Article history: We examined the time series properties of the foreign exchange market for 1990–2008
Received 8 December 2009 in relation to the history of the currency crises using the minimum spanning tree (MST)
Received in revised form 17 September approach and made several meaningful observations about the MST of currencies. First,
2010
around currency crises, the mean correlation coefficient between currencies decreased
Available online 5 November 2010
whereas the normalized tree length increased. The mean correlation coefficient dropped
dramatically passing through the Asian crisis and remained at the lowered level after
Keywords:
Foreign exchange rate
that. Second, the Euro and the US dollar showed a strong negative correlation after 1997,
Currency crisis implying that the prices of the two currencies moved in opposite directions. Third, we
Minimum spanning tree observed that Asian countries and Latin American countries moved away from the cluster
center (USA) passing through the Asian crisis and Argentine crisis, respectively.
© 2010 Elsevier B.V. All rights reserved.

1. Introduction

During the past decade, the application of statistical physics and complexity to financial market data has attracted
much interest. Network analysis, in particular, has been playing a leading role among the related techniques because
the representation of a financial market as a network topology provides efficient ways of understanding its structural
properties [1–4]. In the stock exchange market, stock networks generated by trading activity represent the similarities
between stocks and have significant implications for portfolio optimization [5,6].
Several attempts have also been made to apply network theory to the analysis of the foreign exchange market. Mizuno
et al. [7] analyzed foreign exchange market data and derived a hierarchical taxonomy of currencies constructing a minimum
spanning tree (MST). The identified currency clusters matched nicely with the corresponding countries’ geographical regions
around the world. McDonald et al. [8] developed a network analysis of currency correlations in the foreign exchange market
using the MST approach, and showed that global foreign exchange dynamics such as dominant and dependent currency
structures can be found in MSTs.
However, previous research into the foreign exchange market has shed little light on the time series properties of
currency networks for currency crises. It is important to examine the time series properties of currency networks because
network shapes vary as time goes on. Whether the time series have typical properties in response to currency crises is also
a serious question.
In this paper, our particular interest lies in the application of network theory to the analysis of the foreign exchange
market. We examine the network properties of the market and interpret the network topology in financial terms. Analysis
of the time series properties of currency networks in relation to the history of currency crises is the goal of this paper.

∗ Corresponding author. Tel.: +82 28808335.

E-mail address: changw@snu.ac.kr (W. Chang).

0378-4371/$ – see front matter © 2010 Elsevier B.V. All rights reserved.
doi:10.1016/j.physa.2010.10.028
708 W. Jang et al. / Physica A 390 (2011) 707–718

Table 1
Countries and respective symbols.
Continent Country Symbol Continent Country Symbol
Egypt EGY Poland POL
Africa Nigeria NGR Portugal POR
South Africa RSA Russia RUS
China CHN Europe Spain ESP
Chinese Taipei TPE Sweden SWE
Hong Kong HKG Switzerland SUI
India IND Turkey TUR
Indonesia INA United Kingdom GBR
Asia Japan JPN Bahrain BRN
Korea (Rep. of) KOR Israel ISR
Malaysia MAS Jordan JOR
Philippines PHI Kuwait KUW
Middle East
Singapore SIN Lebanon LIB
Sri Lanka LKA Qatar QAT
Thailand THA Saudi Arabia KSA
Austria AUT United Arab Emirates UAE
Belgium BEL Canada CAN
Czech Rep. CZE North America Mexico MEX
Denmark DEN United States USA
Euro EUR Australia AUS
Pacific Ocean
Finland FIN New Zealand NZL
France FRA Argentina ARG
Europe Germany GER Bolivia BOL
Greece GRE Brazil BRA
Hungary HUN Chile CHI
Iceland ISL South America Colombia COL
Ireland IRL Ecuador ECU
Italy ITA Peru PER
Luxembourg LUX Uruguay URU
Netherlands NED Venezuela VEN
Norway NOR

The remainder of the paper is organized as follows. In Section 2, we present our data set and discuss some of its specific
aspects. In Section 3 we describe the methods that we use to construct the MST derived from the entire sample of data. In
Section 4 we present the results obtained from analysis of the data set, and in Section 5 we draw our conclusions.

2. Data

We analyze the foreign exchange rate of 61 countries from January 1990 to December 2008. The data consist of daily
closing prices provided by Thomson DataStream. The 61 countries in our study and the respective currency symbols are
presented in Table 1.
The foreign exchange rate between two currencies specifies how much one currency is worth in terms of the other and
is generally stated in relation to the US dollar. As this feature excludes the US dollar from the currency network, Mizuno
et al. [7] suggested that a precious metal such as gold, silver, or platinum should be used as the basis of exchange rate.
However, the result could be seriously affected by the variances of the prices of these metals.
To include the US dollar and prevent exogenous factors from affecting the currency network, we base it on Special
Drawing Right (SDR), which is a potential claim on the freely usable currencies of International Monetary Fund (IMF)
members. The value of the SDR is the combined value of a basket of major currencies used in international trade and finance.
At present, the basket consists of the US dollar, the Euro, the Japanese yen, and the pound sterling. For 2006–2010, one SDR
was the sum of 0.6320 US Dollars, 0.4100 Euro, 18.4000 Japanese yen and 0.0903 pound sterling. The currency proportions
of 1 SDR are 0.44(USD), 0.34(EUR), 0.11(JPY) and 0.11(GBP) approximately. As the relative value of each currency varies,
the SDR value continuously fluctuates. The latest value of the SDR in terms of the US dollar is available from the IMF and is
updated daily.
The exchange rate of a currency for dollars can be converted into the exchange rate for SDRs by multiplying the exchange
rate for SDRs by dollars. The daily exchange rate of each country’s currency for SDRs is used as basic data in this study.
We examine changes in the properties of the currency network caused by currency crises. To compile a list of currency
crises, we referenced and updated Kaminsky and Reinhart [9] and Didier et al. [10]. These studies used a weighted average of
exchange-rate changes as an index of currency market turbulence and reserve changes in the spirit of Eichengreen et al. [11].
The chronology of currency crises after 1990 is summarized in Table 2.
Currency crises became common in the 1990s when the neoliberal ideology shaped the globalized financial capitalism
environment. We experienced the European exchange rate mechanism (ERM) crisis in 1992 through 1993, the Mexican peso
crisis in 1994, the Asian currency crisis in 1997, and the Russian ruble crisis in 1998, and the Argentine crisis in 2001, to name
 W. Jang et al. / Physica A 390 (2011) 707–718 709

Table 2
Chronology of currency crises after 1990.
Currency crisis Start of crisis

European exchange rate mechanism (ERM) crisis September 1992

Mexican crisis December 1994
Asian crisis July 1997
Russian default August 1998
Brazilian devaluation January 1999
Argentine crisis December 2001
US sub-prime crisis June 2007
Sources: Kaminsky and Reinhart [9]; Didier et al. [10]; Wall Street Journal, various issues.

a few. The global financial crisis that began in 2007 caused severe currency crises all over the world. Table 2 provides the
chronology of currency crises after 1990.

3. Methodology

In this section, we describe technical methodologies used for the treatment of the data. To construct currency networks,
we employ a method of hierarchical taxonomy introduced by Mizuno et al. [7]. First, we calculate the correlation coefficients
between all pairs of currencies. The logarithmic return of exchange rate Ri (t ) is defined as
Ri (t ) = ln Pi (t ) − ln Pi (t − 1), (1)
where Pi (t ) is the daily exchange rate of currency i at the time t. The correlation coefficient between currencies i and j at
time t , ρit,j , is given by

T (Rti · Rtj ) − (Rti · 1T )(Rtj · 1T )

ρijt =  , (2)
(T (Rti · Rti ) − (Rti · 1T )2 )(T (Rtj · Rtj ) − (Rtj · 1T )2 )
where
Rti = (Ri (t − T + 1), . . . , Ri (t )) i = 1, . . . , N ,
1T = (1, . . . , 1) T number of 1’s,
x · y is the inner product of two vectors x and y, and T is the length of the time window, T = 250 days (1 year). If two
currencies, i and j, are completely correlated (anti-correlated) then ρit,j = +1(−1), while if the two currencies are completely
uncorrelated then ρijt = 0.
Correlation coefficients between all pairs of currencies form a N × N symmetric matrix, the correlation matrix C , with
diagonal elements equal to unity. The mean of correlation coefficients at time t,
N −1 −
N
2 −
ρ̄ t = ρijt . (3)
N (N − 1) i=1 j=i+1

MST is constructed by applying the correlation matrix C . The MST approach in this paper is based on the work of
Mantegna [12]. The correlation coefficient between currencies i and j is transformed to the distance between them at
time t,

dtij = 2(1 − ρijt ). (4)

Based on N × N matrix of dtij , the MST for N currencies is constructed using Kruskal’s algorithm [13,14].
MST is a spanning tree having N − 1 edges with weight less than or equal to the weight of every other spanning tree, and
the edge distances satisfy the three axioms of a Euclidean distance.
The property of the foreign exchange market can be observed by examining the normalized tree length at time t , L(t ),
which is the average distance of edges in MST defined below.
1 −
L( t ) = dtij , (5)
N −1
dtij ∈Θ

where Θ is the set of edges. Here dtij in Eq. (4) becomes the edge distance between nodes (currencies) i and j at time
t [6]. Changes in the density of the MST can be examined with the mean occupation layer, as defined by Onnela et al. [6].
The mean occupation layer for the central node vc at time t is defined as
N
1 −
ℓ(vc ) = η(vi ), (6)
N i=1
710 W. Jang et al. / Physica A 390 (2011) 707–718

0.35
ERM crisis
Mexican crisis
Asian crisis
Russian default
0.3 Brazilian devaluation
Argentine crisis
US subprime crisis

0.25

mean correlation
0.2

0.15

0.1

90 92 94 96 98 00 02 04 06 08
time (year)

Fig. 1. Mean of correlation coefficients (Eq. (3)) as a function of T . We used time windows of T = 250 days and window step length parameter
δ T = 1 day.

where η(vi ) denotes the level of a node vi in relation to the central node vc , whose level is taken to be zero. The mean
occupation layer stands for the layer on which the mass of the tree, on average, is conceived to be located.
The mean occupation layer is calculated with a continuously updated central node. Even though there are various
measures of the node centrality that determine the relative importance of a node within the network, the results are similar
regardless of measure. We employed the betweenness centrality of a node (currency) v [15],
− − σij (v)
CB (v) = , (7)
i̸=v j̸=v,j̸=i
σij
where σij is the number of shortest geodesic paths from i to j, and σij (v) is the number of shortest geodesic paths from node
i to node j that pass through a node v .
The maximum number of links of MST, kmax , is defined as the number of edges of the most connected node in the MST.

4. The evolution of foreign exchange network

In this section, we examine the time series properties of the currency network and analyze those properties in relation
to the currency crises.

4.1. Overall characteristics

Fig. 1 presents the mean correlation coefficient between all currencies in the foreign exchange market from 1990 to 2008.
The time window for the mean correlation is T = 250 days (1 year).
A very interesting observation is that the mean values of the correlation coefficients decreased around all currency
crises except for the Russian default and the Brazilian devaluation. In particular, the mean correlation coefficient dropped
dramatically around the Asian crisis. In the aftermath, the mean correlation coefficient remained at the lowered value. During
this period of the 1990s, the financial market experienced rapid globalization and liberalization.
In the case of the Mexican crisis, the mean correlation coefficient jumped at the beginning, then rapidly decreased for
a short time and then increased sharply. We do not find any clear mean correlation decrement around the Russian default
and the Brazilian devaluation. The mean correlation coefficient fell slightly at the beginning of the Russian default and then
increased. During the Brazilian devaluation, the mean correlation coefficient slowly decreased but the trend then switched
to an upward direction.
The Mexican crisis, the Russian default, and the Brazilian devaluation do not have any significant mean correlation
decrease. These crises were rather regional and limited to a few nations, and the crises of Russia and Brazil occurred just
after the destructive Asian crisis period so that their impacts on the international currency network were relatively small.
The MST analyses confirm that the Asian crisis significantly changed the currency network and some recovery followed even
in the periods of Russian default and Brazilian devaluation.
Fig. 2 plots the correlation coefficients between the three major currencies (yen, pound, and Euro) and the US dollar. The
situation is similar to that of the whole currency network. In particular, the correlation between the Euro and the US dollar
was strongly negative after 1997, which implies that the prices of two currencies moved in opposite directions. If the value
of the Euro went up, the value of the US dollar went down.
 W. Jang et al. / Physica A 390 (2011) 707–718 711

1
ERM crisis
euro
Mexican crisis
0.8 yen
Asian crisis
Russian default pound
0.6
Brazilian devaluation
Argentine crisis
0.4
US subprime crisis

0.2

correlation
0

-0.2

-0.4

-0.6

-0.8

-1
92 94 96 98 00 02 04 06 08
time (year)

Fig. 2. Correlation with US dollar. We use time windows of T = 250 days and window step length parameter δ T = 1 day.

Correlation Coefficient Normalized Tree Length

0.26 0.65

0.24
0.6
0.22

0.2
0.55

0.18

1992 1993 1994 1992 1993 1994

Mean Occupation Layer Maximum Number of Links

6
9

5 8

7
4
6

3 5

1992 1993 1994 1992 1993 1994

Fig. 3. Correlation coefficient, normalized tree length, mean occupation layer, and maximum number of links of MST during 1992–1993. The start of the
ERM crisis is marked on September 1992.

We now present four case studies that show typical changes in the currency network properties: (1) ERM crisis, (2) Asian
crisis, (3) Argentine crisis, and (4) US subprime crisis.

4.2. ERM crisis

On Black Wednesday, September 16 1992, the pound sterling was withdrawn from the European ERM after the British
government failed to keep it above its agreed lower limit. The British government could not thwart short-selling currency
speculators.
Fig. 3 shows what happened to the international exchange market around the crisis. The mean correlation coefficient
decreased in 1992 whereas the normalized tree length increased. This implies that the correlations among currencies tend
to decrease as the foreign exchange market experiences financial crises. The mean occupation layer, in contrast, did not
show special movement around the ERM crisis. The maximum number of links in Fig. 3 was at its lower level around the
Black Wednesday in comparison with the levels before and after the crisis.
712 W. Jang et al. / Physica A 390 (2011) 707–718

Fig. 4. MST before and after the ERM crisis which started on September 1992. Panel (a) shows the MST during September 1991–August 1992. Panel (b)
shows the MST during October 1992–September 1993.

The numerical comparisons of network features between MSTs before and after the ERM crisis are organized in Table 3.
We average the values of four network property measures (correlation coefficient, normalized tree length, mean occupation
layer, and maximum number of links), respectively, during the period of 1 month, 3 months, 6 months, and 1 year before
and after September 1992 when the ERM crisis started. The first column of the table shows the average values of one
month before (August) and after (October) the crisis. The last column compares the average values of one year before
(1991.9.–1992.8.) and after (1992.10.–1993.9.) the crisis. We can see that the mean correlation coefficient decreased and
the normalized tree length increased through the crisis. For mean occupation layer and maximum number of links, we
cannot find any consistent characteristic.
Fig. 4 shows the minimum spanning trees based on one year period before (1991.9.–1992.8.) and after (1992.10.–1993.9.)
the ERM crisis. Grouped features of European countries have been changed very little. The position of US dollar is noticeable.
Before the ERM crisis, the US Dollar was a dominant center in non Europe area. However, passing through the ERM crisis US
Dollar lost the central position in the MST.

4.3. Asian crisis

The Asian financial crisis started in July 1997 with the collapse of the Thai baht. This crisis raised fears of a worldwide
economic meltdown as other Asian currencies, such as those of Indonesia, South Korea, Hong Kong, Malaysia and Philippines,
began to collapse.
The decrease of correlation coefficient and the increase of normalized tree length in Fig. 5 appear to replicate the pattern
of the ERM crisis. It showed consistency regardless of the time span varying from one month to one year. The fall of the
mean correlation coefficient around the Asian crisis was steeper and larger than that around the ERM crisis. The maximum
number of links seems to drop just after the Asian crisis.
 W. Jang et al. / Physica A 390 (2011) 707–718 713

Correlation Coefficient Normalized Tree Length

0.8
0.25

0.2 0.7

0.15
0.6
0.1

0.5
1996 1997 1998 1999 1996 1997 1998 1999

Mean Occupation Layer Maximum Number of Links

6
20

5
15
4

10
3

5
2
1996 1997 1998 1999 1996 1997 1998 1999

Fig. 5. Correlation coefficient, normalized tree length, mean occupation layer, and maximum number of links of MST during 1996–1998. The start of Asian
crisis is marked in July 1997.

Table 3
Correlation coefficient, normalized tree length, mean occupation layer, and maximum number of links of MST before and after the ERM Crisis which started
on September 1992.
1 month 3 months 6 months 12 months

Before 0.209 0.215 0.224 0.226

Correlation coefficient After 0.200 0.200 0.190 0.190
−0.009 −0.015 −0.034 −0.036
Before 0.578 0.568 0.560 0.558
Normalized tree length After 0.583 0.582 0.600 0.598
0.005 0.014 0.040 0.040
Before 3.852 3.787 3.833 4.032
Mean occupation layer After 3.929 3.698 3.607 3.716
0.077 −0.089 −0.226 −0.316
Before 5.000 5.242 5.916 6.682
Maximum number of links After 5.773 6.348 6.477 6.540
0.773 1.106 0.561 −0.142

Through the Asian crisis, the Russian default, and the Brazilian devaluation, the international financial market
experienced significant change. The lowered mean correlation coefficient never recovered to its previous level.
Table 4, which is the Asian crisis version of the comparison table (Table 3) in Section 4.2, shows the changes of four
network property measures around the Asian crisis. Each entry of the table is computed in the same way of Table 3. It is
observed that the mean of correlation coefficient decreased and the normalized tree length increased through the crisis.
Fig. 6 shows the minimum spanning tree before (1996.7.–1997.6.) and after (1997.8.–1998.7.) the Asian crisis. We can
notice that many Asian countries moved away from the cluster center after the crisis. The average layer of the Asian countries
to cluster center (USA) increased to 3.41 from 2.17.

4.4. Argentine crisis

The Argentine economic crisis, in macro-economic terms, started with the decrease of real GDP in 1999 and ended in
2002. As the flight of money away from the country increased, people began withdrawing money from banks in 2001,
turning pesos into US dollars and sending them abroad. This caused bank runs and riots in Argentine during December
2001.
In Fig. 7 and Table 5, the pattern that the correlation coefficient decreases and the normalized tree length increases
repeated during the Argentine crisis while the changes of mean occupation layer and the maximum number of links are not
clear.
714 W. Jang et al. / Physica A 390 (2011) 707–718

Table 4
Correlation coefficient, normalized tree length, mean occupation layer, and maximum number of links of MST before and after the Asian Crisis which
started on July 1997.

1 month 3 months 6 months 12 months

Before 0.090 0.097 0.114 0.153

Correlation coefficient After 0.086 0.085 0.087 0.086
−0.004 −0.012 −0.027 −0.068
Before 0.727 0.727 0.728 0.701
Normalized tree length After 0.746 0.733 0.729 0.726
0.018 0.007 0.001 0.025
Before 3.501 3.342 3.429 3.399
Mean occupation layer After 4.167 3.986 3.799 4.031
0.667 0.643 0.370 0.632
Before 13.571 15.769 15.620 14.165
Maximum number of links After 10.619 11.136 11.069 13.234
−2.952 −4.633 −4.551 −0.931

Fig. 6. MST before and after the Asian crisis which started on July 1997. Panel (a) shows the MST during July 1996–June 1997. Panel (b) shows the MST
during August 1997–July 1998.

Fig. 8 illustrates the changes of MST before (2000.12.–2001.11.) and after (2002.1.–2002.12.) the Argentine crisis. Many
Latin American countries including Argentina moved away from the USA through the currency crisis. The average layer of
the Latin America countries to cluster center (USA) increased to 3.33 from 2.78.
 W. Jang et al. / Physica A 390 (2011) 707–718 715

Correlation Coefficient Normalized Tree Length

0.15 0.82

0.8
0.14

0.78
0.13
0.76
0.12
0.74

0.11 0.72
2000 2001 2002 2003 2000 2001 2002 2003

Mean Occupation Layer Maximum Number of Links

4 15

3.5

3 10

2.5

2 5
2000 2001 2002 2003 2000 2001 2002 2003

Fig. 7. Correlation coefficient, normalized tree length, mean occupation layer, and maximum number of links of MST during 2000–2002. The start of
Argentine crisis is marked in December 2001.

Table 5
Correlation coefficient, normalized tree length, mean occupation layer, and maximum number of links of MST before and after the Argentine Crisis which
started in December 2001.
1 month 3 months 6 months 12 months

Before 0.130 0.128 0.126 0.129

Correlation coefficient After 0.123 0.120 0.117 0.121
−0.008 −0.009 −0.009 −0.008
Before 0.743 0.744 0.743 0.740
Normalized tree length After 0.764 0.777 0.788 0.788
0.021 0.033 0.044 0.048
Before 3.388 3.315 3.311 3.332
Mean occupation layer After 3.428 3.172 3.057 3.108
0.040 −0.144 −0.254 −0.224
Before 7.500 7.569 7.160 7.084
Maximum number of links After 7.000 7.406 8.000 7.923
−0.500 −0.163 0.840 0.839

4.5. US subprime crisis

The global financial crisis of 2007–2009 is known as the worst financial crisis since the Great Depression of the 1930s.
Fig. 9 and Table 6 summarize the characteristics of the foreign exchange market around the crisis. As noted repeatedly
above, the correlation coefficient dropped during the currency crisis and resulted in an increased normalized tree length.
The maximum number of links jumped at the initial period of the crisis, which is different from the pattern in the Asian
crisis (Fig. 5).
Fig. 10 shows the minimum spanning tree before (2006.6.–2007.5.) and after (2007.7.–2008.6.) the US sub-prime crisis.
One visible change is the distance between North American countries. Before the crisis, the layers to USA from Canada and
Mexico were 2 and 3, respectively. After the crisis, they increased 6 and 8 for Canada and Mexico, respectively.

5. Conclusions

Currency crises became common during the 1990s when globalized financial capitalism flourished. We have observed
that the currencies of nations with heavy foreign debts in emerging markets severely depreciated against major currencies
such as the US dollar when the global credit market tightened up.
716 W. Jang et al. / Physica A 390 (2011) 707–718

Fig. 8. MST before and after the Argentine crisis which started in December 2001. Panel (a) shows the MST during December 2000–November 2001. Panel
(b) shows the MST during January 2002–December 2002.

Table 6
Correlation coefficient, normalized tree length, mean occupation layer, and maximum number of links of MST before and after the US sub-prime Crisis
which started in June 2007.
1 month 3 months 6 months 12 months

Before 0.127 0.127 0.125 0.125

Correlation coefficient After 0.126 0.122 0.118 0.115
−0.001 −0.005 −0.008 −0.010
Before 0.713 0.706 0.701 0.695
Normalized tree length After 0.748 0.749 0.753 0.747
0.035 0.042 0.052 0.053
Before 3.886 3.963 3.872 3.730
Mean occupation layer After 3.681 3.865 3.933 3.985
−0.205 −0.098 0.061 0.254
Before 9.214 9.121 9.285 10.337
Maximum number of links After 12.963 13.200 12.817 12.502
3.749 4.079 3.532 2.165

This paper has focused on a network-based analysis of currency crises in foreign exchange market. We have examined
the time series properties of foreign exchange market and analyzed them in relation to currency crises, extracting some
meaningful observations.
 W. Jang et al. / Physica A 390 (2011) 707–718 717

Correlation Coefficient Normalized Tree Length

0.14

0.13 0.75

0.12
0.7
0.11

0.1 0.65
2006 2007 2008 2009 2006 2007 2008 2009

Mean Occupation Layer Maximum Number of Links

6

15
5

10
4

3
5
2006 2007 2008 2009 2006 2007 2008 2009

Fig. 9. Correlation coefficient, normalized tree length, mean occupation layer, and maximum number of links of MST during 2006–2008. The start of US
sub-prime crisis is marked on June 2007.

First, the mean values of the correlation coefficients decreased around most currency crises whereas the normalized
tree length increased. These patterns appear repeatedly in the ERM crisis (Table 3), the Asian crisis (Table 4), the Argentine
crisis (Table 5), and the US sub-prime crisis (Table 6). In particular, the mean correlation coefficient dropped dramatically
passing through the Asian crisis, and remained at the lowered value. The Asian crisis occurred as the financial market
experienced rapid globalization and liberalization in the 1990s. However, the mean correlation drop pattern for currency
crisis is ambiguous. The steep drop of correlation coefficient is found before the date Asian crisis begins (Fig. 5), which is
distinct from the ERM crisis where the drop is found after the crisis (Fig. 3) starts. As most currency crises are the outcomes
of various continuing economic events, it is difficult to pinpoint the beginning and ending dates of crises. In this paper, we
use the starting dates of crises in Refs. [9,10], and the multiple issues of Wall Street Journal.
Second, the correlation between the Euro and the US dollar dramatically turned negative from 1997, which implies that
the prices of the two currencies moved in opposite directions.
Finally, we observed that Asian countries and Latin American countries moved away from the cluster center (USA) of
MSTs going through the Asian crisis and Argentine crisis, respectively. This implies that the Asian (Latin American) countries,
belong to a cluster of MST before the Asian (Argentine) crisis, tend to move farther away from the cluster center of MST after
the crisis. North American countries such as Canada and Mexico also moved away from USA through the US subprime crisis.
However, these patterns are not shown in the ERM crisis.
These tentative results await refinement and correction in the light of further research. It would be dangerous to
generalize that the above-mentioned phenomena will be observable during all currency crises. As we have seen, the Mexican
crisis and the Russian crisis, which were well known for their contagious nature on the global financial market, did not show
obvious trends.
The methods used in this paper have the following limitations. For the validity of correlation analysis between currencies,
we need the assumption that the data from currency markets are mutually independent normal random variables. The
Pearson coefficient of linear correlation in Eq. (2) may not work well for non-random, strongly correlated data. The financial
data in practice, including ours, are likely to be correlated and show non-Gaussian behavior. In addition, although MST
analysis is beneficial in capturing the general views and changes of international currency market, the relatively simple
tree diagram structure in MST is definitely limited in not being able to represent the detailed characteristics of international
currency markets, which are complex and dynamic. For more accurate analysis, we should pay attention to these limitations
and improve the standard techniques.
Our work can suggest elementary ideas about the relationship between currency crises and the properties of the currency
network. The results of this study could be developed into the research on an early warning system (EWS) for financial crises.
With the growing frequency of financial crises, many researchers and practitioners have tried to develop a better EWS. It
remains for future work to investigate whether the performance of EWS can be improved by considering currency network
properties in addition to standard macroeconomic factors.
718 W. Jang et al. / Physica A 390 (2011) 707–718

Fig. 10. MST before and after the US sub-prime crisis which started on June 2007. Panel (a) shows the MST during June 2006–May 2007. Panel (b) shows
the MST during July 2007–June 2008.

References

[1] R.N. Mantegna, H.E. Stanley, An Introduction to Econophysics: Correlations and Complexity in Finance, Cambridge University Press, Cambridge, 2000.
[2] G. Bonanno, G. Caldarelli, F. Lillo, S. Miccichè, N. Vandewalle, R.N. Mantegna, Networks of equities in financial markets, Eur. Phys. J. B 38 (2004) 363.
[3] R. Coelho, C.G. Gilmore, B. Lucey, P. Richmond, S. Hutzler, The evolution of interdependence in world equity markets-evidence from minimum spanning
trees, Physica A 376 (2007) 455.
[4] C. Eom, G. Oh, S. Kim, Deterministic factors of stock networks based on cross-correlation in financial market, Physica A 383 (2007) 139.
[5] V. Tola, F. Lillo, M. Gallegati, R.N. Mantegna, Cluster analysis for portfolio optimization, J. Econom. Dynam. Control 32 (2008) 235.
[6] J.P. Onnela, A. Chakraborti, K. Kaski, J. Kertész, A. Kanto, Dynamics of market correlations: taxonomy and portfolio analysis, Phys. Rev. E 68 (2003)
561101.
[7] T. Mizuno, H. Takayasu, M. Takayasu, Correlation networks among currencies, Physica A 364 (2006) 336.
[8] M. McDonald, O. Suleman, S. Williams, S. Howison, N.F. Johnson, Detecting a currency’s dominance or dependence using foreign exchange network
trees, Phys. Rev. E 72 (2005) 046106.
[9] G.L. Kaminsky, C.M. Reinhart, The twin crises: the causes of banking and balance-of-payments problems, Amer. Econ. Rev. 89 (1999) 473.
[10] T. Didier, P. Mauro, S.L. Schmukler, Vanishing financial contagion? J. Pol. Model. 30 (2008) 775.
[11] B. Eichengreen, A.K. Rose, C. Wyplosz, Exchange market mayhem: the antecedents and aftermath of speculative attacks, Econ. Policy 10 (1995) 249.
[12] R.N. Mantegna, Hierarchical structure in financial markets, Eur. Phys. J. B 11 (1999) 193.
[13] D.B. West, Introduction to Graph Theory, Prentice-Hall, Englewood Cliffs, NJ, 1996, r15.
[14] C.H. Papadimitriou, K. Steiglitz, Combinatorial Optimization, Prentice-Hall, Englewood Cliffs, NJ, 1982.
[15] L.C. Freeman, A set of measures of centrality based on betweenness, Sociometry 40 (1977) 35.