Journal of Financial Markets 14 (2011) 82–108

www.elsevier.com/locate/finmar

Housewives of Tokyo versus the gnomes of Zurich:

Measuring price discovery in sequential markets$
Jianxin Wangn, Minxian Yang
Australian School of Business, University of New South Wales, Sydney 2052, Australia
Available online 7 August 2010

Abstract

This paper presents two methods to measure market-specific contributions to price discovery in
non-overlapping sequential markets: one is a non-parametric approach using high-frequency data
and the other is a structural VAR model based on open-to-close returns. The methods complement
the existing methodologies for comparing price discovery in parallel markets. Using these methods,
we estimate the information shares of four sequential markets for the trading of AUD, JPY, EUR,
and GBP against USD over an eight-year period. We find that price discovery in the foreign
exchange markets are still dominated by Europe and the United States, particularly the London–
New York overlapping trading hours. Asia is losing information shares to Europe in the trading of
AUD and JPY. The significance of the ‘‘housewives of Tokyo’’ in currency trading may have been
overstated.
Crown Copyright & 2010 Published by Elsevier B.V. All rights reserved.

JEL classification: G14; G15; C32

Keywords: Price discovery; Information share; Sequential markets; Realized variance; Beverage-Nelson
decomposition; Efficient price; Foreign exchange rate

‘‘Years ago, it was the gnomes of Zurich who shook the foreign exchange markets.
They have now been replaced by the housewives of Tokyo, who speculate in various

$
We appreciate helpful comments from Douglas Foster, Frederick Harris, Joel Hasbrouck, Bruce Lehmann,
Ron Masulis, S. Ghon Rhee, Maxwell Stevenson, Shuji Watanabe, and participants at seminars at the University
of New South Wales, Monash University, Xiamen University, the microstructure conference at the University of
Sydney, and the 2009 FMA Asian Conference. We thank the Securities Industry Research Centre of Asia-Pacific
(SIRCA) for providing the foreign exchange data.
n
Corresponding author. Tel.: þ61 2 9385 5863; fax: þ61 2 9385 6347.
E-mail addresses: jx.wang@unsw.edu.au (J. Wang), m.yang@unsw.edu.au (M. Yang).

1386-4181/$ - see front matter Crown Copyright & 2010 Published by Elsevier B.V. All rights reserved.
doi:10.1016/j.finmar.2010.08.002
 J. Wang, M. Yang / Journal of Financial Markets 14 (2011) 82–108 83

currencies. However, whereas the gnomes of Zurich were accused in their day of
destabilizing markets, the housewives of Tokyo are apparently acting to stabilize
them. Their presence seems to lie behind the marked decline in (perceived) volatility
in yen-dollar exchange rates. y The housewives are betting against professional
investors in the IMM,1 and seem to be profiting from their trading so far.’’
Dr. Kiyohiko Nishimura, Bank of Japan.2

1. Introduction

The above speech, given at the Brookings Institution in Washington on July 2, 2007,
created the metaphorical ‘‘Mrs. Watanabe,’’ whose currency trading and market impact
captured the imagination of financial press. ‘‘Now, a new beast is roiling markets — Mrs.
Watanabe, or Japan’s archetypal housewife,’’ says the Financial Times on the next day.
‘‘On her shoulders may lie responsibility for some of the stability of the global financial
system,’’ according to The Economist (August 16, 2007). The speech by Dr. Nishimura,
along with the press commentaries, suggests that retail investors in Japan have significant
market impact in currency trading. However, Tokyo is not a leading hub for currency
trading, not even for yen-related transactions.3 Its share of currency trading has been
declining relative to markets elsewhere. If indeed the ‘‘housewives of Tokyo’’ are replacing
the ‘‘gnomes of Zurich’’ and winning against futures traders in Chicago, it would suggest
that trading in Tokyo has greater price impact than trading in Europe and North America.
Is Tokyo a leading information hub for currency trading, particularly for the yen? How
does one measure the significance of Tokyo trading to the pricing of the Japanese yen
relative to trading in markets in Europe and North America?
This paper aims to address these questions. It proposes and compares two approaches to
estimate the contribution of a particular market to the price discovery of an asset traded in
multiple markets. The global financial liberalization has resulted in more assets than ever
before being traded in multiple markets around the world. As companies and economies
become more exposed to global risk factors, knowing where price discovery occurs often
sheds light on what factors affect asset prices. A cross-market comparison of price
discovery provides an estimate of the distribution of information and information-based
trading for the underlying asset. A market’s contribution to price discovery reflects its
efficiency in collecting and analyzing information, the quality of its regulation and
microstructure, as well as its liquidity, which are important for attracting informed
investors. With changes in global economic conditions and trading technology, the
significance of a market may change over time. The traditional link between an asset, e.g.,
the Australian dollar and a particular market, e.g., Sydney, may be weakening. While
officials in Japan are concerned with maintaining the global status of the Japanese financial
markets, markets in Shanghai are getting greater global attention. Understanding what
1
IMM stands for the International Monetary Market of the Chicago Mercantile Exchange, a major venue for
the trading of currency futures.
2
Dr. Nishimura is currently the Deputy Governor, Bank of Japan. His speech is published as Nishimura (2007).
3
The BIS survey shows that U.K. and U.S. account for 34% and 16.6% of global currency transactions
respectively. Japan’s share of currency trading has declined from 9.1% in 2001 to 6% in 2007 and is similar to that
of Switzerland (6.1%) and Singapore (5.8%). For yen-related transactions, Japan’s market share is 25%,
compared to 28.3% for the U.K. (BIS, 2007, Tables B.2 and E.4).
84 J. Wang, M. Yang / Journal of Financial Markets 14 (2011) 82–108

changes have taken place in the global pecking order is necessarily the first step in
understanding what factors are responsible for such changes.
In this paper we measure price discovery in multiple markets that are sequential in time
and non-overlapping (e.g., Tokyo and London). The case of parallel markets where
trading takes place simultaneously in different venues (e.g., the NYSE and the NASDAQ)
was first examined in the seminal study by Hasbrouck (1995). The information share
proposed by Hasbrouck (1995) measures the contribution of a particular market to the
changes in the efficient price of an asset. An alternative approach taken up by Booth, So,
and Tse (1999), Chu, Hsieh, and Tse (1999), and Harris, McInish, and Wood (2002), uses
the permanent-transitory decomposition of Gonzalo and Granger (1995). Both approaches
have been adopted by numerous studies of price discovery between parallel markets (e.g.,
spot versus derivative trading, floor versus electronic trading). Cross-market comparisons
of price discovery have spawned into ‘‘a mid-sized cottage industry’’ (Lehmann, 2002).
A special issue of the Journal of Financial Markets in 2002 was devoted to the comparison
between the two approaches. Pascual, Pascual-Fuster, and Climent (2006) expand the
Hasbrouck model to distinguish trade-related versus trade-unrelated information shocks.
New applications of the Hasbrouck and the Gonzalo-Granger models have continued to
flourish.4
In parallel markets, price discovery occurs simultaneously in multiple venues; therefore
it is difficult to achieve a clean separation of price innovations from different markets. This
is well recognized in the literature (e.g., Lehmann, 2002). Recently Yan and Zivot (2010)
use a structural cointegrated VAR to give new insights to the comparison between the
Hasbrouck and Gonzalo-Granger models. Yan and Zivot (2007) propose to use
the cumulative impulse responses from the structural cointegrated VAR to better capture
the dynamics of price discovery.
Because the existing models are designed only for parallel markets, international
comparisons of price discovery are often limited to one or two hours of overlapping
trading time (e.g., Hupperets and Menkveld, 2002; Grammig, Melvin, and Schlag, 2005;
Pascual, Pascual-Fuster, and Climent, 2006). In most cases, the closing of the home market
overlaps with the opening of the overseas market, often the NYSE. The small overlapping
hours may lead to bias against the newly-opened market as the newly-arrived traders learn
from past price movements (see Hsieh and Kleidon, 1996). Menkveld, Koopman, and
Lucas (2007) develop a state-space model that takes into account prices from non-
overlapping trading periods for partially overlapping markets. Currently there is no model
designed for comparing price discovery across non-overlapping markets (e.g., Tokyo vs.
London or Shanghai vs. New York). Studies of non-overlapping markets (e.g., Lieberman,
Ben-Zion, and Hauser, 1999; Agarwal, Liu, and Rhee, 2007) focus on Granger causality
between the prices without measuring their contributions to the efficient price.
We propose two approaches to measure price discovery in sequential non-overlapping
markets. The basic logic behind the two approaches is the same as the Hasbrouck model:
Information flow is measured by the variation in the efficient price of an asset. Changes in
the efficient price are identified as the random-walk component of the observed returns.
The information share of a particular market is its share in the total variance of the
efficient price in a trading day. The two approaches differ in how the variance of the

4
See for example Chakravarty, Gulen, and Mayhew (2004), Covrig, Ding, and Low (2004), Figuerola-Ferretti
and Gonzalo (2010), Cao, Hanscah, and Wang (2008), and Harris, McInish, and Wood (2008).
 J. Wang, M. Yang / Journal of Financial Markets 14 (2011) 82–108 85

efficient price is estimated. The first approach is based on the recent advances in estimating
the integrated variance using high-frequency observations. We use the two-scales (TS)
estimator of Zhang, Mykland, and Aı̈t-Sahalia (2005), which is a consistent estimator of
the integrated variance. The variance of the efficient price is estimated as the mean of the
TS estimator. In addition to the information share, the ratio of the TS estimator to the
realized variance provides a measure for the pricing efficiency of a market. Our second
approach is based on a structural vector autoregressive (VAR) model for the open-to-close
returns of sequential markets. This is particularly useful when intraday prices are not
available, or when measuring information flows during non-trading periods (e.g.,
overnight). Studies have shown significant information flow during non-trading periods.
Our measure of the information contribution of each market is based on the market-
specific shocks and has a clean interpretation. The structural VAR model has several
advantages: it is very easy to implement, its data requirement is not onerous, and the
empirical results do not depend on intraday sampling frequency and the noise-filtering
method used. The drawback of the structural VAR is that the intraday observations are
not fully exploited for estimating the variance of the efficient price change. In our sample,
the estimated information shares from the structural VAR are similar to those from
realized variance. As discussed in Section 3, combining our methods with those for parallel
markets allows researchers to study the general case of partially overlapping markets.
Our methods are applied to the foreign exchange markets, which trade continuously around
the clock. By estimating the information shares across markets and time zones, we provide
new evidence on exchange rate price discovery. There is indirect evidence that some markets
are more important than others in currency trading. Ito, Lyons, and Melvin (1998) and Covrig
and Melvin (2002) provide evidence of private information in currency trading and suggest
that Tokyo may know more about the yen than other markets. However the findings are
disputed by Andersen, Bollerslev, and Das (2001). We compare price discovery across global
markets for AUD, JPY, EUR, and GBP, all against USD, for an eight-year period from
January 1996 to December 2003. These are the top-four currency pairs in trading value,
representing 58% of global currency trading (Bank for International Settlements (BIS), 2007).
A 24-hour day is divided into four sequential markets (see Table 1): ‘‘Asia,’’ ‘‘Europe,’’ the
overlap between London and New York City labeled as ‘‘LondonþNYC,’’ and ‘‘U.S.,’’ which
covers trading in the Americas. The information shares of these markets are estimated using
realized variance, as well as the structural VAR model. Sub-period analyses reveal changes in
the contribution of each market over the eight years. The findings are summarized below:

 ‘‘Europe’’ and ‘‘LondonþNYC’’ have the highest pricing efficiency: over 90% of the
price variations are changes in the efficient price. Their combined information shares are
40–42% for AUD and JPY and 51–53% for EUR and GBP. ‘‘Asia’’ has the lowest
pricing efficiency among the four markets. Its information shares are 28–33% for AUD
and JPY and 16–17% for EUR and GBP. ‘‘U.S.’’ takes up the remaining shares.
 Over the sample period, price discovery appears to become more concentrated in
‘‘Europe,’’ particularly in the ‘‘LondonþNYC’’ trading hours. ‘‘Asia’’ lost shares in
Asian currencies while ‘‘U.S.’’ remained stable. For JPY, the information share of
‘‘Asia’’ declined from 33% in 1996–1999 to 28% in 2000–2003. ‘‘Europe’’ increased its
share from 28% to 31% and that of ‘‘LondonþNYC’’ rose from 12% to 15%. The
findings suggest that Tokyo is not a leading information hub for JPY. The significance
of ‘‘Mrs. Watanabe’’ may have been overstated.
86 J. Wang, M. Yang / Journal of Financial Markets 14 (2011) 82–108

 The two-hour ‘‘LondonþNYC’’ period is highly significant for currency trading. It has
the highest per-hour information shares for all currencies. It gained information shares
in all four currencies over the sample period.
 Different measures of information shares give similar estimates of the significance of a
market. However the information share of a market can be very different from its share
of trading volume. For example, ‘‘Asia’’ accounts for over half of AUD trading but less
than one-third of AUD information share. In general, the share of the trading volume
of the home market overstates its information share.
Although we do not examine what affects a market’s information share, evidence from
microstructure studies of the foreign exchange markets offers some clues. Since the trading
platform is the same, the difference between markets is in the number and characteristics of
market participants. Studies have shown that order flows are the critical link between
exchange rate changes and economic fundamentals (Evans and Lyons, 2002a, 2005, 2007,
2008); order flows from financial institutions have greater information content than other
investors (Bjonnes, Rime, and Solheim, 2005; Carpenter and Wang, 2007); and there are
cross-market information flows between currencies (Evans and Lyons, 2002b; Yan and
Zivot, 2007). Therefore a market’s information share depends on the quantity and the
quality of its order flows. Such order flows tend to come from financial institutions with a
large client base and substantial research capacity.
The paper is organized as follows: Section 2 presents the two approaches for measuring
price discovery in sequential markets. Section 3 explains the data and trading statistics in
the foreign exchange markets. The estimation and empirical findings in the currency
markets are discussed in Section 4. Section 5 is a brief summary.

2. Measuring price discovery in sequential markets

Consider a single asset traded in two adjoining and non-overlapping markets during a
24-hour trading day. The scenario is depicted in Fig. 1: market 1 opens at the beginning of
day t and closes at the point when market 2 opens, while market 2 closes at the beginning
of day tþ1. As discussed shortly, the setup can be easily modified to allow overlapping
markets or non-trading periods. The closing prices of markets 1 and 2 are p1,t and p2,t,
respectively. The end-of-day price is ptp2,t. All prices are logarithms. The open-to-close
returns of the markets are r1,t=p1,tp2,t1 and r2,t=p2,tp1,t, and are subject to market-
specific price innovations Z1,t and Z2,t. In terms of notation, the first subscript i (i=1, 2)
indicates the market and the second subscript t (t=1, 2, 3,y) indicate the (24-hour) day.
For the same t, Z1,t and p1,t always occur before Z2,t and p2,t. The innovations Z1,t and Z2,t

1,t p1,t

Market 1 2,t p2,t

Market 2

Day t-1 Day t Day t+1

Trading time

Fig. 1. Two sequential non-overlapping markets.

 J. Wang, M. Yang / Journal of Financial Markets 14 (2011) 82–108 87

represent ‘‘news’’ and are uncorrelated with each other and over time. They may contain
permanent shocks reflecting unexpected changes in economic fundamentals, and transitory
shocks resulting from short-term mispricing, liquidity shocks, or microstructure factors
(e.g., bid-ask bounce or inventory control).
The closing prices pi,t of each market can be written as pi,t=mi,tþui,t, i=1,2, where mi,t is the
efficient price reflecting new information on economic fundamentals, and ui,t is a noise term
resulting from transitory factors. Changes in the efficient price in markets 1 and 2 are
Dm1,t=m1,tm2,t1=h1Z1,t and Dm2,t=m2,tm1,t=h2Z2,t, where h1 and h2 are constants and can
be estimated using daily observations, see Eq. (7). With respect to the information set available
at the opening of market i on day t, Dmi,t is a martingale difference. Therefore, Dm1,t and Dm2,t
are uncorrelated with each other and over time. They capture the information components in
price innovations Z1,t and Z2,t. The change of the efficient price over day t is Dmt=m2,t–
m2,t1=Dm1,tþDm2,t. Following Hasbrouck (1995), information flow in market i is measured
by the variance of Dmi,t. Therefore the information share of market i is defined as
varðDmi;t Þ varðDmi;t Þ
IS i ¼ ¼ ; i ¼ 1; 2: ð1Þ
varðDmt Þ varðDm1;t Þ þ varðDm2;t Þ
When the efficient price is treated as a continuous-time process, the conditional variance of its
change over a time interval is measured by the integrated variance (the instantaneous variance
integrated over the time interval), see Andersen and Benzoni (2008). The integrated variance
may be estimated for each day by using intraday observations. The variance of Dmi,t can be
expressed as the unconditional mean of the integrated variance of market i on day t.
In this section, we propose two ways to estimate var(Dmi,t). When intraday prices are
available, it can be estimated by the mean value of an estimator of the integrated variance.
This is discussed in Section 2.1. If intraday prices are not available, or one of the markets is a
non-trading period (e.g., overnight), var(Dmi,t) can be estimated using the structural VAR
model presented in Section 2.2. We use the two-market setup to demonstrate the
methodology and note that it can be easily generalized to any number of sequential markets.
If there is a non-trading period between two markets (e.g., Hong Kong and New York), one
can estimate the information shares of the three periods, including the non-trading period in
the middle. Similarly, when two markets are partially overlapping (e.g., Europe and the U.S.),
they can be divided into three periods with the overlapping period in the middle. Our model
can be used to estimate price discovery in the three sequential periods.5 Price discovery in the
overlapping period can be decomposed using models for parallel markets.

2.1. Information shares measured by realized variance

Our first approach is based on the fast expanding literature on realized variance.6
Market i can be divided into Mi sampling intervals. Return in market i on day t over
sampling interval s can be written as ri,t,s=Dmi,t,sþDui,t,s, where Dmi,t,s is the change in the
efficient price over the sampling interval and represents a permanent shift in the asset
value, and Dui,t,s captures intraday
P i noise. When sampled at high frequency, i.e., Mi-N,
the sum of squared Dmi,t,s, M s¼1 Dm 2
i;t;s , converges to the integrated variance in market i on

5
This is the approach taken in our empirical analysis of Europe and the U.S.
6
We thank the editor and the referee for suggesting the use of realized variance for information shares. For a
detailed discussion on realized variance, see Andersen and Benzoni (2008) and the references therein.
88 J. Wang, M. Yang / Journal of Financial Markets 14 (2011) 82–108

P i 2
day t. However, the realized variance (RV), defined as RVi;t  M s¼1 ri;t;s , is a biased and
inconsistent estimator of the integrated variance in the presence of the noise term ui,t,s.
Since the efficient price and the noise term are not observable, many studies have
proposed ways to reduce or remove the impact of the noise term on the estimation of the
integrated variance (e.g., Aı̈t-Sahalia, Mykland, and Zhang, 2005; Barndorff-Nielsen,
Hansen, Lunde, and Shephard, 2005; Zhang, Mykland, and Aı̈t-Sahalia, 2005; Bandi and
Russell, 2006). The two-scales (TS) estimator of Zhang, Mykland, and Aı̈t-Sahalia (2005)
appears to be the first consistent estimator of the integrated variance. Aı̈t-Sahalia and
Mancini (2008) show that the TS estimator unambiguously outperforms the RV in out-of-
sample forecasting comparisons. Barndorff-Nielsen, Hansen, Lunde, and Shephard (2005)
argue for the merits of the kernel-based estimators and show that the TS estimator can be
expressed as a kernel-based estimator
 XMi XHi  MXi h
Mi Hi h vi;t;Hi þ1
E
RVi;t ¼ 1 r þ
2
1 ðri;t;sh ri;t;s þ ri;t;sþh ri;t;s Þ ;
Mi ðHi þ 1Þ s¼1 i;t;s h¼1 Hi þ 1 s¼hþ1 Hi þ 1
ð2Þ
where Hi is the number of ‘‘autocovariances.’’ The adjustment factor vi;t;Hiþ1 is recursively
defined by vi,t,10 and vi;t;h  vi;t;h1 þ ðri;t;1 þ    þ ri;t;h1 Þ2 þ ðri;t;Mi h2 þ    þ ri;t;Mi Þ2
for h=2, y, Hiþ1. The TS estimator is in fact a linear combination of the standard RV at
two different frequencies (two-scales): the highest frequency possible and a lower
frequency. Since the RV consistently estimates the noise variance as sampling frequency
goes to infinity, the RV sampled at the high frequency is a good approximation to the noise
variance. At the low frequency, many feasible RVs may be computed (e.g., with 1-minute
returns there are five ways to construct 5-minute RV). The estimation error of the average
of these RVs at the low frequency depends on the noise variance. The linear combination
of the average of the low-frequency RVs and the high-frequency RV then serves to correct
the impact of the noise term on the average of the low-frequency RVs and effectively lead
to a consistent estimator of the integrated variance.
E
The TS estimators RVi;t for market i on day t can be treated as a stationary time series
with possibly strong autocorrelations (e.g., Andersen, Bollerslev, Diebold, and Labys,
2003; Corsi, 2009; Aı̈t-Sahalia and Mancini, 2008). The unconditional variance of the effi-
cient price var(Dmi,t) may be estimated by the mean of the TS estimator, RViE ¼ EðRVi;t E
Þ.
Then the information share in (1) becomes
RViE
ISi ðRV E Þ ¼ ; i ¼ 1; 2: ð3Þ
RV1E þ RV2E

The distribution of information shares across markets is denoted as IS(RVE). For

comparison, we also calculate the information shares based on the standard RV, IS(RV).7
The mean of the TS estimator, RViE , may be estimated by the sample mean of the time
E
series RVi;t . But the estimation error is not straightforward to quantify given that the time
E
series RVi;t is generally strongly autocorrelated. Following the insight of Corsi (2009) that
7
An alternative approach to the non-parametric estimator is to explicitly model the intraday return as an
autoregressive process. The intraday changes of the efficient price, extracted from the Beverage-Nelson (1981)
decomposition, are then squared and summed to estimate the integrated variance. In our empirical application,
this approach leads to results (not reported in this paper) similar to those based on the TS estimator.
 J. Wang, M. Yang / Journal of Financial Markets 14 (2011) 82–108 89

the realized variance series can be described as a heterogeneous AR process (which is a

restricted parsimonious AR), we use an unrestricted VAR to model the time series
E 0
behavior of the vector V t ¼ ½RV1;t
E
; RV2;t  and estimate the mean values.8 Specifically, the
VAR is given by GðLÞV t ¼ a þ errort , where G(L) is the AR lag polynomial matrix and a is
a constant vector. The VAR is estimated by OLS. The mean of Vt is given by G(1)1a, with
plug-in OLS estimates, and its components are used to compute (3). Further, a parametric
bootstrap with block-error draws is use to quantify the estimation errors in (3).9 The
bootstrap procedure is given in Appendix A.

2.2. A structural VAR model for open-to-close returns

Asset returns over non-trading periods, e.g. overnights and weekends, have been
examined since French (1980). Recent studies have shown substantial overnight
information flow.10 If intraday prices are not available for a market, or one wants to
compare price discovery between trading and non-trading periods, the above RV-based
method is no longer feasible. In this case, the information shares of the two markets can be
estimated from a structural VAR model using only open-to-close returns.
The basic idea behind the structural VAR is similar to the RV-based approach in Section
2.1. The goal is to estimation the variance of the efficient price changes for the sequential
markets. We use a structural VAR to identify the random-walk component in the open-to-
close returns and estimate its variance. Let rt=[r1,t, r2,t]0 be the vector of open-to-close
returns of the two markets in Fig. 1. The dynamics of rt is modeled as a structural vector
autoregressive process with K lags:11
X
K
B0 rt ¼ Bk rtk þ gt ð4Þ
k¼1

where gt=[Z1,t, Z2,t]0 . As discussed above, the structural shocks Z1,t and Z2,t capture the
market-specific price innovations and are uncorrelated. The variance of the structural
shocks is normalized to one. Therefore E(gt)=0; E(gtg0 tk)=0) for ka0; 0
! E(gtg t)=I,
b11 0
a 2  2 identity matrix.12 B0 is a lower triangular matrix B0 ¼ because the
b21 b22
markets are sequential: within the same trading day t, r1,t affects r2,t but not vice versa. The
impact of market 2 on market 1 is captured by the lagged returns on the right hand of (4).
8
The estimated mean values from VAR are almost the same as the sample means in our empirical study.
9
The information shares may be alternatively estimated from the sample mean of Vt with the standard errors
being constructed via direct block-bootstrapping Vt. Given that Vt is strongly autocorrelated, this approach may
be more sensitive to the block size than the parametric bootstrap with block-error draws.
10
See for example Cao, Ghysels, and Hatheway (2000), Davies (2003), Barclay and Hendershott (2003, 2004,
2008). Cliff, Cooper, and Gulen (2008) report a striking finding that the US equity premium in recent years is
solely due to overnight returns.
11
We do not consider cross-currency impact, e.g. from JPY returns to AUD returns. Such consideration
requires a high-dimensional VAR model of parallel trading of different currencies. The main difficulty is the
identification of currency-specific shocks. The lack of a cross-currency model is a limitation of the current study as
it can potentially reveal a common source of price discovery for multiple currencies. Cross-currency information
flows have been documented by Evans and Lyons (2002b) and Yan and Zivot (2007).
12
An alternative and equivalent parameterization is to normalize the diagonal elements of B0 as unity and leave
the variance of Zt as a positive diagonal matrix.
90 J. Wang, M. Yang / Journal of Financial Markets 14 (2011) 82–108

In the real world, the sequence of the markets is set by their trading hours in calendar time.
In Appendix B, we show that the model outcomes are invariant to the choice of the start of
the 24-hour trading cycle.
Eq. (4) can be written in the reduced form
X
K
rt ¼ Ak rtk þ et or AðLÞrt ¼ et ; ð5Þ
k¼1

where Ak ¼ B1
0 Bk ; AðLÞ ¼ IA1 L    AK L ; L is the lag operator; the roots of
K

9A(z)9=0 are outside the unit circle. The reduced-form shock et satisfies et ¼ B1 0 gt ¼
!
b11 0
gt with E(et)=0, (E(ete0 tk)=0 for ka0, Eðet e0 t Þ ¼ ðB0 0 B0 Þ1 ¼ X. The lag
b12 b22
polynomial A(L) and the covariance matrix X can be estimated using least squares. Since
B0 is lower triangular and X is symmetric, the elements of B0 are exactly identified by
ðB0 0 B0 Þ1 ¼ X and can be estimated by using the lower triangular Cholesky factor of the
least squares estimator of X.
The VAR in (5) has a moving average representation in the form of Beverage-Nelson
(1981) decomposition
X
1
1 1
rt ¼ AðLÞ et ¼ Að1Þ et þ CðLÞðet et1 Þ with CðLÞ ¼ Cj Lj
j¼0

where Að1Þ ¼ IA1     AK , CðLÞ ¼ ½AðLÞ1 Að1Þ1 =ð1LÞ and Cj converges to zero
0
exponentially as j increases. The daily return (over 24 hours) is obviously r1,tþr2,t= rt,
where is a vector of ones. The end-of-day price pt is the sum of 0 rt over t=1, y, t and
may be written as13
X
t X
t
pt ¼ p0 þ 0 Að1Þ1 et þ et ¼ p0 þ 0 Að1Þ1 B1
0 gt þ e t
t¼1 t¼1
X
1 X
1
0 0 0
et ¼ CðLÞet ¼ Cj etj ¼ Cj B1
0 gtj
j¼0 j¼0

where p0 is determined by the initial conditions at t=0. The term et represents pricing
errors and is stationary. The efficient price is defined as
Xt
mt ¼ lim Eðptþq 9F t Þ ¼ p0 þ 0 Að1Þ1 B1
0 gt ð6Þ
q-1
t¼1

where Ft is the information set at the end of day t. The daily change in the efficient price
Dmt ¼ mt mt1 ¼ 0 Að1Þ1 B1 0
0 gt ¼ h gt ¼ h1 Z1;t þ h2 Z2;t ð7Þ
is a combination of the structural shocks in two markets, and h0 ¼ ½h1 ; h2  ¼ 0 Að1Þ1 B1
0 is
the vector of the impact coefficients of the two shocks. The h coefficients here incorporate
the long-run effects A(1)1, equivalent to the cumulative impulse response of infinite steps.
13
Note that unlike rt, which is a 2  1 vector, the end-of-day price pt is a scalar and equals p2,t.
 J. Wang, M. Yang / Journal of Financial Markets 14 (2011) 82–108 91

Based on the structural VAR, the information share14 in (1) becomes

varðhi Zi;t Þ h2
ISi ðSVARÞ ¼ ¼ 2 i 2; i ¼ 1; 2: ð8Þ
varðDmt Þ h1 þ h2

The cross-market distribution of information shares is denoted as IS(SVAR).

The IS(SVAR) is based on the impact of structural shocks that are market-specific, while
most existing models are based on reduced-form shocks. From (5), the reduced-form
shocks e1,t and e2,t are functions of the independent structural shocks: e1;t ¼ b11 Z1;t and
e2;t ¼ b21 Z1;t þ b22 Z2;t where bij are the elements of B1 0 . Eq. (7) implies that Dmt ¼
Að1Þ1 et ¼ d0 et ¼ d1 e1;t þ d2 e2;t , and varðDmt Þ ¼ d21 varðe1;t Þ þ d22 varðe2;t Þ þ 2d1 d2 covðe1;t ; e2;t Þ.
The contributions of the reduced-form shocks e1,t and e2,t cannot be unambiguously
separated unless they are uncorrelated. Therefore, using reduced-form VAR to compute
information shares is appropriate only when the correlations among the reduced-form
shocks can be ignored. The structural model outlined above provides a clean separation of
price innovations in sequential markets, while the structural model of Yan and Zivot
(2007) provide a clean separation between permanent and transitory shocks in parallel
markets. Compared to the state-space model of Menkveld, Koopman, and Lucas (2007),
our model is extremely easy to implement: least squares estimation of (4) and the Cholesky
factorization of ˆX are sufficient for the measure in (8). The estimation errors of the
information shares are quantified using the bootstrap procedure detailed in Appendix A.

3. Global currency markets and data summary

The global currency markets can be treated as the sequential markets described in Section 2.
A common asset, e.g. the Japanese yen, is traded 24 hours a day across Asia, Europe, and the
United States time zones. The trading platform is identical across all markets. While the
determinants of currency values are not well understood,15 they are more susceptible to global
risk factors than most other assets. For example, the link between the value of the Australian
dollar and commodity prices is widely acknowledged by academics and market practitioners.
Greater volatility in commodity prices may shift the price discovery of the Australian dollar
from Sydney to London or New York where commodity trading is concentrated. We implement
the models in Section 2 to estimate the information shares of different markets in currency
trading and explore how the significance of each market has changed over time.
The exchange rates we study are AUD, JPY, EUR, and GBP against USD. Our data
source is the Reuters’ intraday foreign exchange quotes for these rates from 1 January 1996
(1 January 1999 for EUR) to 31 December 2003. As conventional in studies of currency
trading, weekends are removed because of thin trading. We also remove days with large
gaps (over 4 hours) in quote arrivals, which can be the result of system stoppage or
holidays in parts of the world. On October 7 and 8, 1998, JPY had ‘‘once-in-a-generation’’
volatility, and both AUD and GBP experienced high volatility.16 These days are treated as
14
An alternative measure, termed the component share (CS), is often used by studies based on the Gonzalo-Granger
(1995) model. A similar measure for our structural model is CSi=hi/(h1þh2), i=1,2.
15
The disconnection between exchange rates and domestic macroeconomic fundamentals are well established
since Meese and Rogoff (1983). See Lyons (2001) for the new micro exchange rate economics.
16
On October 7, 1998, JPY jumped from around 130 to 120 per USD in one day. See Cai, Cheung, Lee, and
Melvin (2001) for events surrounding these days.
92 J. Wang, M. Yang / Journal of Financial Markets 14 (2011) 82–108

Table 1
Definitions of four sequential markets.
This table divides a 24-hour calendar day into four sequential markets: ‘‘Asia’’ from 23 to 6 GMT, ‘‘Europe’’
from 7 to 12 GMT, ‘‘LondonþNYC’’ from 13 to 14 GMT, and ‘‘U.S.’’ from 15 to 22 GMT. The bold letters
denote local trading hours.

Markets GMT Sydney Tokyo Hong Kong/ Frankfort/Zurich London New York San Francisco
Singapore

0 10 9 8 2 1 20 17
A 1 11 10 9 3 2 21 18
S 2 12 11 10 4 3 22 19
I 3 13 12 11 5 4 23 20
A 4 14 13 12 6 5 0 21
5 15 14 13 7 6 1 22
6 16 15 14 8 7 2 23

E 7 17 16 15 9 8 3 0
U 8 18 17 16 10 9 4 1
R 9 19 18 17 11 10 5 2
O 10 20 19 18 12 11 6 3
P 11 21 20 19 13 12 7 4
E 12 22 21 20 14 13 8 5

London 13 23 22 21 15 14 9 6
þNYC 14 0 23 22 16 15 10 7

15 1 0 23 17 16 11 8
16 2 1 0 18 17 12 9
U 17 3 2 1 19 18 13 10
S 18 4 3 2 20 19 14 11
19 5 4 3 21 20 15 12
20 6 5 4 22 21 16 13
21 7 6 5 23 22 17 14
22 8 7 6 0 23 18 15

ASIA 23 9 8 7 1 0 19 16

outliers and are removed from our analyses. This leaves us with 1884 days for AUD, 1902
days for JPY, 1189 days for EUR, and 1879 days for GBP.
We divide the 24 hours into four sequential markets based on their time zones. In
addition to the three conventional time zones in Asia, Europe, and the Americas, the
overlapping trading hours in London and New York City are singled out as a separate
market due to its special significance in currency trading. The four markets are defined in
Table 1, which lists the local time in various markets relative to the Greenwich Mean Time
(GMT). The bold letters are local trading hours from 9 am to 4 pm local time. Market 1 is
the Asian trading hours from the start of 23 GMT to the end of 6 GMT in the next day,
and is labeled as ‘‘Asia’’. Market 2 is the European trading from the start of 7 GMT to the
end of 12 GMT. It covers most of the trading hours in Frankfurt and Zurich and is labeled
as ‘‘Europe’’. Market 3 is the overlap of London afternoon and New York morning
trading from 13 to 14 GMT and is labeled as ‘‘LondonþNYC’’. Market 4 is from the start
 J. Wang, M. Yang / Journal of Financial Markets 14 (2011) 82–108 93

Table 2
Average daily transactions.
This table reports the average daily transactions in April 2007 in top 10 foreign exchange markets. It is
constructed from Bank for International Settlements (BIS), (2007, Table E.5). Transactions include spot, outright
forward, and swap transactions against USD and are measured in billion USD.

AUD JPY EUR GBP

Value Percent Value Percent Value Percent Value Percent

Australia 76.7 33.4 13.5 2.5 23.5 2.2 13.5 3.0

Denmark 0.46 0.2 1.89 0.4 20.9 1.9 1.74 0.4
France 5.89 2.6 13.0 2.4 48.2 4.4 9.92 2.2
Germany 1.25 0.5 9.29 1.7 43.0 3.9 7.03 1.6
Hong Kong 14.0 6.1 16.4 3.1 20.2 1.9 12.6 2.8
Japan 10.6 4.6 138.8 25.8 25.7 2.4 7.62 1.7
Singapore 15.7 6.8 43.1 8.0 47.9 4.4 21.2 4.7
Switzerland 6.12 2.7 20.7 3.9 74.0 6.8 28.6 6.3
United Kingdom 55.9 24.3 153.6 28.6 443.6 40.7 240.3 53.3
United States 30.9 13.5 94.0 17.5 179.1 16.4 77.1 17.1
Asia in Top 10 117.0 51.0 211.8 39.4 117.3 10.7 54.9 12.2
Europe in Top 10 69.6 30.3 198.5 36.9 629.7 57.7 287.6 63.8
Top 10 217.5 94.7 504.3 93.8 926.1 84.9 419.6 93.1
Global total 229.6 100 537.5 100 1091.2 100 450.8 100

of 15 GMT to the end of 22 GMT and covers the Americas excluding ‘‘LondonþNYC’’.
Since the United States is the only one in the top-ten markets (Table 2), market 4 is labeled
as ‘‘U.S.’’.
Table 2 reports the average daily trading values of the four currency pairs in top-10
currency markets. It is constructed from the triennial survey conducted by the Bank for
International Settlements (BIS), (2007, Table E.5). Although the reported trading values
are not divided into the four markets defined in Table 1, it gives a broad picture of the
distribution of the trading values over a 24-hour trading day. Five of the top 10 are in
Europe and four are in Asia. AUD trading is more concentrated in Asia (51%),
particularly in Australia (33.4%). There is more JPY-USD trading in U.K. (28.6%) than in
Japan (25.8%), as in the case of all JPY-related transactions. JPY trades are more evenly
split between Asia (39.4%) and Europe (36.9%). EUR and GBP are dominated by Europe
and U.K. in particular. Asia’s trading shares in EUR and GBP are much lower than those
of U.S. These ten markets account for over 93% of world trading of AUD, JPY, and GBP,
and 84.9% of world trading of EUR.

4. Information shares in currency trading

This section reports the empirical estimation of the information shares of the four
sequential markets: ‘‘Asia’’, ‘‘Europe’’, ‘‘LondonþNYC’’, and ‘‘U.S.’’. The first two sub-
sections report the estimated information shares based on realized volatility and the
structural VAR model, respectively. We then present evidence on the information shares of
per-hour trading and the changes of information shares over the eight-year sample period.
94 J. Wang, M. Yang / Journal of Financial Markets 14 (2011) 82–108

0.8
JPY
0.7

Realized Volatility (x104)

0.6
0.5
0.4 AUD EUR
0.3
0.2 GBP
0.1
0.0
0 10 20 30 40 50 60
Sampling Intervalin Minute

Fig. 2. Signature plots of realized volatility.

4.1. Information shares based on realized variance: IS(RVE)

To estimate RV and RVE, intraday returns at various sampling intervals are

constructed. We sample the intraday Reuters quotes at 1, 2, 5, 10, 15, 20, 30, and 60
minutes intervals. The midpoint of the bid and ask quotes at the end of each sampling
interval is computed as the linear interpolation of the mid-quotes immediately before and
after the end of the interval.17 The return in market i on day t at time s, ri,t,s, is calculated as
100 times the log ratio of the prices at time s and s1. The daily realized variance for
market i, RVi,t, i=1,y,4, is the sum of the squared ri,t,s over its trading hours.
Fig. 2 presents the average daily RV at different sampling intervals, termed the signature
plot. The pattern of RV at different sampling intervals is very similar to those reported in
other studies, e.g. Hansen and Lunde (2006). As discussed in Section 2.1, the estimated RV
is largely driven by noise when sampled at ultra high frequencies. For example, at 1-minute
sampling interval the realized variance has a high noise component. The average RV
declines as the sampling interval increases. It becomes stable when the sampling interval is
higher than 30 minutes for AUD, JPY, EUR, and 15 minutes for GBP. We use the
30-minute RV, denoted as RV30m, as the daily RV. As we see later, there is still a significant
noise component in RV even with 30-minute sampling.
The two-scales estimator, RViE , calls for the use of the highest sampling frequency. In
E
our case, it is estimated using 1-minute sampling and is denoted as RV1m . The number of
the ‘‘autocovariance’’ terms, Hi in Eq. (2), is set at 6, which is approximately the value
computed from the popular formula of Newey and West (1994) for the Bartlett kernel; see
also Hansen and Lunde (2006, p139).
With the daily RVi; tEPseries (market i on day t), the daily information share may be
defined as ISi;t ¼ RVi;tE
= j RVj;t
E 18
. The time series plots of the daily RVi;tE
and ISi,t for JPY

17
One drawback of the linear interpolation is that when the sampling interval goes to zero, the return and the
realized variance reduce to zero (Hanson and Lunde, 2006). An alternative approach is to use the mid-quote
immediately before the end of an interval. The drawback here is the possibility of stale price. Both methods are
widely used in the literature. We use the linear interpolation because we do not sample at ultra high frequencies
and we are concerned of stale price, especially for AUD.
18
We thank the referee for suggesting this and the use of unit-root tests to check stationarity.
 J. Wang, M. Yang / Journal of Financial Markets 14 (2011) 82–108 95

Asia London+NYC

3 1.5

2 1.0

1 0.5

0 0.0
1996 1998 2000 2002 2004 1996 1998 2000 2002 2004
Europe U.S.
1.5
4
3 1.0
2
0.5
1
0 0.0
1996 1998 2000 2002 2004 1996 1998 2000 2002 2004

Asia London+NYC

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0.0 0.0
1996 1998 2000 2002 2004 1996 1998 2000 2002 2004
Europe U.S.
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0.0
1996 1998 2000 2002 2004 1996 1998 2000 2002 2004

E
Fig. 3. Plots of daily RV1m E
and ISðRV1m Þ for JPY: (a) dailyRV1m
E E
for JPY and (b) daily ISðRV1m Þ for JPY.

are presented in Fig. 3, where each daily series fluctuates at an approximately constant
level and there is no visible trend. The plots for other currencies are not presented because
E
of similarity. The sample statistics for RVi;t and ISi,t are presented in Table 3. The
augmented Dickey-Fuller test convincingly rejects the null hypothesis of a unit root in any
of the daily series, whereas the Ljung-Box Q statistic suggests that all daily series have
strong autocorrelations. Because of the autocorrelations, the sample standard deviations in
Table 3 are not suitable for statistical inferences.
E
Table 4 reports the estimated mean values of daily RV1m and RV30m for the four
E
markets. Since RV1m and RV30m are highly persistent, the mean values are estimated from
a VAR model and their standard errors are estimated from the bootstrap procedure in
Appendix A with 1000 replications. The lag length of the VAR, GðLÞV t ¼ a þ errort , is
selected by the Akaike Information Criterion (AIC) from a maximum 32, where Vt is the
96 J. Wang, M. Yang / Journal of Financial Markets 14 (2011) 82–108

Table 3
E
Summary of sample statistics of daily RV1m and information share.
This table reports the sample mean, sample standard deviation (Stddev), Ljung-Box Q statistic for 10 lags
(QLB(10)), and augmented Dickey-Fuller unit-root test statistic (ADFstat) of daily realized variances and
information shares. The 1% critical value for ADFstat is 3.96. Here RV1m E
is the daily TS estimator of the
E E
integrated variance at 1-minute sampling interval and IS(RV1m ) is the daily information share based on RV1m .

Asia Europe LondonþNYC U.S.

AUD RV E1m
Mean 0.127 0.127 0.055 0.141
Stddev 0.137 0.128 0.074 0.148
QLB(10) 1128 1603 699 857
ADFstat 11.5 10.6 8.93 10.8
IS(RV E1m )
Mean 0.283 0.285 0.119 0.314
Stddev 0.129 0.114 0.076 0.131
QLB(10) 349 38.0 30.5 84.2
ADFstat 13.3 30.0 29.5 14.2
JPY RV E1m
Mean 0.152 0.135 0.058 0.116
Stddev 0.238 0.196 0.080 0.125
QLB(10) 1126 1281 321 1548
ADFstat 12.6 11.0 14.1 9.5
IS(RV E1m )
Mean 0.304 0.297 0.134 0.265
Stddev 0.139 0.118 0.089 0.124
QLB(10) 194 139 65.7 98.6
ADFstat 26.0 13.7 28.1 21.0
EUR RV E1m
Mean 0.072 0.142 0.069 0.133
Stddev 0.107 0.236 0.068 0.121
QLB(10) 62.7 134 171 296
ADFstat 21.1 9.98 12.6 14.8
IS(RV E1m )
Mean 0.177 0.329 0.167 0.327
Stddev 0.103 0.129 0.098 0.135
QLB(10) 132 159 36.4 206
ADFstat 16.0 13.0 22.8 12.4
GBP RV E1m
Mean 0.035 0.085 0.033 0.072
Stddev 0.037 0.077 0.029 0.055
QLB(10) 377 1021 348.6 558
ADFstat 15.6 10.5 14.2 11.9
IS(RV E1m )
Mean 0.160 0.370 0.150 0.320
Stddev 0.090 0.120 0.080 0.130
QLB(10) 179 204 45.3 307
ADFstat 12.1 20.3 30.4 13.4

vector of RVs of the four markets on day t. For (AUD, JPY, EUR, GBP), the selected lag
E E
lengths are (9, 19, 23, 8) for RV1m and (8, 8, 5, 8) for RV30m, respectively. As RV1m is a
reasonable proxy for the information flow (the variance of the efficient price change) and
RV30m contains both information and noise, the ratio RV1m E
=RV30m provides a measure of
 J. Wang, M. Yang / Journal of Financial Markets 14 (2011) 82–108 97

Table 4
Estimated mean of daily realized variance.
This table reports the estimated mean of daily realized variance and the associated standard error, using the
method of Appendix A. RV1m E
=RV30m is the ratio of the estimated means of RV1m
E
and RV30m, respectively. RV1m E

is the TS estimator of the integrated variance at 1-minute sampling interval. RV30m is the realized variance of the
observed price at 30-minute sampling interval. The intraday returns are the difference in log prices over the
sampling interval, multiplied by 100.

Asia Europe LondonþNYC U.S.

AUD
E 0.127 0.128 0.055 0.141
RV1m
St. err. 0.011 0.011 0.005 0.011
RV30m 0.182 0.141 0.063 0.166
St. err. 0.018 0.013 0.005 0.014
RV1mE
=RV30m (%) 70 91 87 85

JPY
E 0.151 0.134 0.058 0.114
RV1m
St. err. 0.021 0.017 0.005 0.013
RV30m 0.203 0.143 0.065 0.14
St. err. 0.024 0.016 0.005 0.012
RV1mE
=RV30m (%) 74 94 89 81

EUR
E 0.070 0.142 0.069 0.131
RV1m
St. err. 0.007 0.019 0.006 0.014
RV30m 0.099 0.153 0.075 0.165
St. err. 0.010 0.013 0.005 0.010
RV1mE
=RV30m (%) 71 93 92 79

GBP
E 0.035 0.085 0.033 0.072
RV1m
St. err. 0.002 0.005 0.002 0.003
RV30m 0.053 0.095 0.039 0.088
St. err. 0.005 0.006 0.002 0.004
RV1mE
=RV30m (%) 66 89 85 82

market-specific pricing efficiency. Table 4 shows that for all currencies and all markets, the
ratio RV1mE
=RV30m are less than one, implying that RV30m contains considerable noise,
ranging from 6% to 34%. Using this measure for the pricing efficiency, there is a
remarkably consistent ordering in the pricing efficiency: ‘‘Europe’’ has the highest
efficiency for all four currencies, followed by ‘‘LondonþNYC’’ and ‘‘U.S.’’. Asia has the
lowest pricing efficiency across the four markets.
E
The estimated information shares ISi(RV1m ) are reported in Table 5. As in Table 4, the
associated standard errors are calculated from the bootstrap. All of the estimated
information shares are statistically different from zero. There are several features in the
table. First, if the 24 hours are divided into three 8-hour time zones, then the European
time zone, which includes ‘‘Europe’’ and ‘‘LondonþNYC’’ defined in Table 1, dominates
price discovery in currency trading. The combined information shares of ‘‘Europe’’ and
‘‘LondonþNYC’’ range from 40.6% for AUD to 52.8% for GBP. The two-hour
‘‘LondonþNYC’’ trading accounts for a significant portion of price discovery in this time
98 J. Wang, M. Yang / Journal of Financial Markets 14 (2011) 82–108

Table 5
Information shares based on TS estimator.
This table reports the estimated information shares based on the TS estimator of the integrated variance at
E
1-minute sampling interval, ISðRV1m Þ. The estimation is based on equation (3) and the bootstrap procedure that
follows.

Asia Europe LondonþNYC U.S.

AUD
ISðRV1m E
Þ (%) 28.2 28.3 12.2 31.3
St. err. 0.007 0.005 0.004 0.007

JPY
ISðRV1m E
Þ (%) 33.1 29.3 12.6 25.0
St. err. 0.011 0.008 0.006 0.007

EUR
ISðRV1m E
Þ (%) 17.0 34.5 16.7 31.8
St. err. 0.008 0.014 0.007 0.011

GBP
ISðRV1m E
Þ (%) 15.6 37.8 14.7 31.9
St. err. 0.005 0.008 0.003 0.008

zone. ‘‘Europe’’ generally has greater information shares than ‘‘U.S.’’. The exception is
AUD which may reflect price discovery in AUD futures contracts traded on the Chicago
Mercantile Exchange.
Second, the information shares of ‘‘Asia’’ are 28.2% for AUD and 33.1% for JPY.
Therefore around 70% of permanent price changes in these currencies occur outside of
Asia. It seems unlikely that Tokyo knows more about the yen than others. Our results do
not support a significant role from Japanese retail investors. The information shares of
‘‘Asia’’ are much lower for EUR and GBP.
Third, a market’s high pricing efficiency may not translate into high information shares.
If a market with high pricing efficiency has low information flow, it may not have high
information shares. Even though the pricing efficiency of ‘‘Asia’’ is similar for AUD and
EUR (Table 4), the low information share of ‘‘Asia’’ in EUR trading may be the result of
low information flow regarding EUR in Asia time zone.
Fourth, the information shares of a market are generally different from its share of
trading values in Table 2. For example, Asia accounts for 51% of AUD trading but only
28% in AUD information flow. In general, trading volume tends to overstate the
significance of the home market in the pricing of local currencies.

4.2. Information shares based on the structural VAR: IS(SVAR)

For each market defined in Table 1, the open-to-close returns are calculated from the
Reuters’ intraday quotes. For example, ‘‘Europe’’ opens at the start of 7 GMT and closes
at the end of 12 GMT. The opening price is the interpolation of the mid-quotes
immediately before and after 7 GMT. The closing price is the opening price of
‘‘LondonþNYC’’, which is the interpolation of the mid-quotes immediately before and
after 13 GMT. The open-to-close return for ‘‘Europe’’ is the difference of the two log
 J. Wang, M. Yang / Journal of Financial Markets 14 (2011) 82–108 99

Table 6
Summary of open-to-close returns.
The definitions of markets are given in Table 1. Open-to-close returns in the four markets are defined as
100  [ln(QLB)–ln(Popen)]. QLB(10) is the Ljung-Box statistic for 10 lags. The asterisknindicates significant at 5%
level.

Asia Europe LondonþNYC U.S.

AUD
Mean 0.007 0.006 0.001 0.005
Stddev. 0.361 0.360 0.254 0.348
Skewness 0.246 0.270 0.117 0.395
Kurtosis 7.60 11.44 6.90 5.82
QLB(10) 11.09 13.64 4.80 23.20n
Correlation
Europe 0.010
LDN-NYC 0.068n 0.073n
U.S. 0.061n 0.029 0.007
JPY
Mean 0.006 0.009 0.002 0.018
Stddev. 0.424 0.392 0.262 0.323
Skewness 0.034 0.640 0.503 0.048
Kurtosis 8.95 8.70 11.40 7.16
QLB(10) 31.03n 13.33 24.29n 13.54
Correlation
Europe 0.018
LDN-NYC 0.017 0.0006
U.S. 0.005 0.084n 0.015
EUR
Mean 0.006 0.040 0.016 0.029
Stddev. 0.260 0.415 0.283 0.373
Skewness 0.646 0.579 0.019 0.051
Kurtosis 8.45 7.12 4.83 5.52
QLB(10) 13.64 10.30 18.50n 21.56n
Correlation
Europe 0.068n
LDN-NYC 0.003 0.081n
U.S. 0.027 0.012 0.091n
GBP
Mean 0.006 0.019 0.016 0.020
Stddev. 0.178 0.299 0.197 0.264
Skewness 0.287 0.065 0.035 0.118
Kurtosis 8.42 5.07 4.90 6.15
QLB(10) 5.54 13.65 5.04 10.66
Correlation
Europe 0.064n
LDN-NYC 0.003 0.042n
U.S. 0.031 0.004 0.073n
100 J. Wang, M. Yang / Journal of Financial Markets 14 (2011) 82–108

prices. Table 6 reports the summary statistics. Similar to Table 4, AUD and JPY have high
volatility during Asia trading while EUR and GBP have high volatility during the
European market.19 EUR and GBP have greater volatility during the two-hour
‘‘LondonþNYC’’ than during the eight-hour trading in ‘‘Asia’’. Both show high levels
of skewness and kurtosis in Asia. The Ljung-Box Q-statistic shows that JPY and EUR
have significant same-market autocorrelations while GBP shows no autocorrelation.
Returns in the LondonþNYC period are often negatively correlated with returns in other
markets. JPY returns in Asia have strong autocorrelation but no cross-market correlation.
When estimating the structural VAR model, the number of lags is determined by the
AIC criterion, and is one for AUD and GBP, five for JPY, and two for EUR. The reduced-
form VAR of Eq. (5) is estimated via least squares. The structural coefficient B1 0 is
obtained from the Cholesky factorization of the estimated covariance matrix X. The price
impact coefficients are h0 ¼ ½h1 ; h2 ; h3 ; h4  ¼ 0 Að1Þ1 B1
0 . The IS(SVAR) is calculated from
(8). Standard errors are based on bootstraps with 1000 replications.
The results of the structural VAR are presented in four segments in Table 7 for AUD,
JPY, EUR, and GBP, respectively. The top panel of each segment presents the estimated
structural coefficients matrix B1 0 . For all four exchange rates, at least one of the off-
1
diagonal elements of B0 is statistically different from zero.20 Therefore the covariance
matrix X is not diagonal, and the reduced-form shocks et in Eq. (5) are correlated across
markets. As discussed before, it is critical to use the structural model in this case in order to
separate market-specific price innovations.
The bottom panel of each segment in Table 7 reports the estimated information shares
and statistical inference. Across the four markets, ‘‘Asia’’ has the highest information
shares for AUD and JPY, and ‘‘Europe’’ has the highest information shares for EUR and
GBP. However the combined ‘‘Europe’’ and ‘‘LondonþNYC’’ trading period has the
highest information shares for all four currencies. ‘‘U.S.’’ has much higher information
shares than ‘‘Asia’’ in EUR and GBP. The standard errors are estimated from the
bootstrap procedure in Appendix A. Even though some of the estimated information
shares are small, e.g. 8.1% for EUR in ‘‘Asia’’, all of them are significantly greater than
zero. The 90% confidence intervals are narrow enough to allow for statistical comparisons
between information shares of different markets.
Table 8 compares the estimated ISðRV1m E
Þ with IS(SVAR). Comparisons are based on
the mean absolute difference (MAD) between the estimated information shares of the same
market and the Spearman’s rank correlation across markets (Rank Cor.). Except for
AUD, cross-market rankings based on ISðRV1m E
Þ and IS(SVAR) are similar and rank
correlations are high at 0.8 or 1. The MAD for AUD is reasonably small; therefore the low
rank correlation for AUD is driven by small numerical differences in the values across
markets that are of similar magnitudes. Overall IS(SVAR) and IS(RVE) are broadly
E
consistent with each other. Since RV1m utilizes the intraday high-frequency data and is
based on a consistent estimator of the daily market-specific integrated variance, ISðRV1m E
Þ
appears more desirable than IS(SVAR) and should be used when high-frequency data are
available. However, if high-frequency data are not available, or one of the markets is a
non-trading period, the structural VAR based IS(SVAR) is the only feasible choice.

19
Note that Table 4 reports the realized variance while Table 6 reports the sample standard deviations of open-
to-close returns.
20
The standard errors of B1
0 coefficients are not reported here to conserve space.
 J. Wang, M. Yang / Journal of Financial Markets 14 (2011) 82–108 101

Table 7
Information shares based on the structural VAR model.
The definitions of markets are given in Table 1. IS(SVAR) is the information share estimated from the
structural VAR model. The number of lags K is determined by the Akaike Information Criterion.

Asia Europe LondonþNYC U.S.

AUD
Structural coefficients B1
0 with K=1 lag
Asia 0.359n
Europe 0.003 0.360n
LondonþNYC 0.016n 0.018n 0.253n
America 0.021n 0.009 0.003 0.346n
IS(SVAR) (%) 33.1 29.6 13.1 24.3
St. err. 0.030 0.028 0.020 0.023
Lower 5% 0.282 0.249 0.100 0.207
Upper 5% 0.381 0.340 0.165 0.283
JPY
Structural coefficients B1
0 with K=5 lags
Asia 0.416n
Europe 0.004 0.386n
LondonþNYC 0.005 0.001 0.259n
U.S. 0.006 0.026n 0.005 0.319n
IS(SVAR) (%) 30.9 27.9 20.3 20.9
St. err. 0.060 0.044 0.040 0.051
Lower 5% 0.214 0.210 0.143 0.131
Upper 5% 0.408 0.352 0.271 0.296

EUR
Structural coefficients B1
0 with K=2 lags
Asia 0.256n
Europe 0.026n 0.412n
LondonþNYC 0.001 0.023n 0.281n
U.S. 0.007 0.002 0.029n 0.365n

IS(SVAR) (%) 8.1 44.3 21.0 26.6

St. err. 0.036 0.042 0.031 0.038
Lower 5% 0.033 0.366 0.163 0.207
Upper 5% 0.148 0.504 0.269 0.336

GBP
Structural coefficients B1
0 with K=1 lag
Asia 0.176n
Europe 0.017n 0.297n
LondonþNYC 0.001 0.008 0.196n
U.S. 0.009 0.001 0.020n 0.263n

IS(SVAR) (%) 13.6 38.8 17.3 30.2

St. err. 0.020 0.025 0.024 0.024
Lower 5% 0.106 0.349 0.138 0.260
Upper 5% 0.170 0.432 0.215 0.339
102 J. Wang, M. Yang / Journal of Financial Markets 14 (2011) 82–108

Table 8
Comparing information share measures.
This table compares ISðRV1mE
Þ, the information shares based on the TS estimator of the integrated variance at
1-minute sampling interval, with IS(SVAR), the information shares based on the structural VAR model for the
open-to-close returns. ‘‘MAD’’ is the mean absolute difference between ISðRV1mE
Þ and IS(SVAR). ‘‘Rank cor.’’ is
the Spearman’s rank correlation across markets.

MAD (%) Rank cor.

AUD 3.5 0.2

JPY 3.9 1
EUR 6.9 0.8
GBP 1.7 0.8

Table 9
Information shares per trading hour.
This table compares ISðRV1mE
Þ, the information shares based on the TS estimator of the integrated variance at
1-minute sampling interval, with IS(SVAR), the information shares based on the structural VAR model for the
open-to-close returns. ‘‘MAD’’ is the mean absolute difference between ISðRV1mE
Þ and IS(SVAR). ‘‘Rank Cor.’’ is
the Spearman’s rank correlation across markets.

Asia (%) Europe (%) LondonþNYC (%) U.S. (%) MAD (%) Rank cor.

AUD
E
ISðRV1m Þ 3.5 4.7 6.1 3.9
IS(SVAR) 4.1 4.9 6.6 3.0 0.53 0.8
JPY
E
ISðRV1m Þ 4.1 4.9 6.3 3.1
IS(SVAR) 3.9 4.7 10.2 2.6 1.22 1
EUR
E
ISðRV1m Þ 2.1 5.8 8.4 4.0
IS(SVAR) 1.0 7.4 10.5 3.3 1.36 1
GBP
E
ISðRV1m Þ 2.0 6.3 7.4 4.0
IS(SVAR) 1.7 6.5 8.7 3.8 0.46 1

4.3. Price discovery per trading hour

The four markets in Table 1 have different trading hours. A market may have a high
information share simply because of its long trading hours. An alternative comparison of
information flows is based on the per-hour information shares reported in Table 9. The
table highlights the dominance of ‘‘LondonþNYC’’ in all currencies but particularly EUR
and GBP. ‘‘Europe’’ has the second highest per-hour information shares in all four
currencies. ‘‘Asia’’ beats ‘‘U.S.’’ only in JPY. On a per-hour basis, not only the MADs
between ISðRV1m E
Þ and IS(SVAR) are smaller, their rank correlations are much higher.
4.4. Historical trend in price discovery

The historical evolution of the information shares in currency trading is examined in

Table 10. We compare the average information shares in the first four years (three years for
 J. Wang, M. Yang / Journal of Financial Markets 14 (2011) 82–108 103

Table 10
Sub-period information share.
E
This table reports the annual estimates of ISðRV1m Þ, the information shares based on the TS estimator of the
integrated variance at 1-minute sampling interval. N is the number of observations in each year.

Asia (%) Europe (%) LondonþNYC (%) U.S. (%) N

AUD
1996 36 24 11 29 240
1997 29 30 11 31 232
1998 28 29 12 32 237
1999 26 29 11 33 249
Average 30 28 11 31
2000 30 26 12 32 247
2001 26 31 12 31 229
2002 29 29 12 31 228
2003 22 30 15 33 222
Average 27 29 13 32

JPY
1996 31 29 13 27 243
1997 32 28 12 27 242
1998 36 26 12 26 240
1999 33 30 12 26 250
Average 33 28 12 27

2000 27 34 14 25 245
2001 30 32 14 24 228
2002 30 31 15 25 232
2003 25 28 16 31 222
Average 28 31 15 26
EUR
1999 20 33 15 32 252
2000 15 35 16 34 247
2001 18 34 17 31 233
Average 18 34 16 32

2002 17 31 18 34 233
2003 19 31 18 32 224
Average 18 31 18 33
GBP
1996 18 33 14 35 236
1997 18 38 13 31 229
1998 17 37 14 32 237
1999 15 39 14 32 245
Average 17 37 14 32

2000 13 39 15 34 245
2001 14 39 17 31 231
2002 16 35 16 33 229
2003 17 37 16 30 227
Average 15 37 16 32
104 J. Wang, M. Yang / Journal of Financial Markets 14 (2011) 82–108

EUR) and in the last four years (two years for EUR). ‘‘Asia’’ lost information shares over
the sample period. Its information share in AUD dropped from 30% in the first half to
27% in the second half. Its share in JPY pricing dropped from 33% to 28%, which is the
largest change across all currencies and markets. This may be the result of the decade-long
recession in Japan during the 1990s. Traditional market centers in Europe and the United
States gained information shares. ‘‘Europe’’ gained in information share in JPY at the
expense of ‘‘Asia’’. It lost shares in EUR to ‘‘LondonþNYC’’. The information shares of
‘‘LondonþNYC’’ increased across all four currencies. The information shares of ‘‘U.S.’’
remained stable over the sample period. Contrary to the perception of rising importance of
Hong Kong and Singapore, there is evidence of concentration of price discovery in the
traditional market centers. The price discovery of AUD and JPY is more concentrated in
Europe. The price discovery of EUR and GBP is more concentrated in the overlapping
hours between London and New York.

5. Final remarks

Two simple methods for comparing price discovery in sequential markets are introduced
and applied to the 24-hour foreign exchange trading. We present new evidence on the
information shares across markets in different time zones, and how the information shares
have changed over the eight-year sample period. Our model for sequential markets can be
used in conjunction with models for parallel markets to compare price discovery in
partially overlapping markets. It can also be used to measure the information shares of
intraday trading hour by hour and explore related microstructure issues. Future research
should explore what determines the information share of a market and what drive its
changes.

Appendix A. Bootstrap procedure

Consider a general VAR model U(L)yt=cþet, where U(L) is the AR lag polynomial
matrix and yt, et and c are vectors of the same dimension. Here c is constant and et is a
martingale difference process. We are interested in a function F of the parameters in U(L),
c and the variance of et. For our application the F function is the information shares
defined in (3) or (8) or the mean of the TS estimator. While the VAR model is easily
estimated by OLS, the standard error of the estimated F is generally unavailable in
analytical forms. For financial time series, the et in the VAR usually exhibits dependence in
the second or higher moments, such as GARCH effects. To take this data feature into
account, we use a parametric bootstrap with block residual draws to quantify the sampling
variation in the estimated F. The bootstrap procedure is described as follows.

S1. Determine the lag length K by AIC, estimated the VAR model and estimate the
function of interest F, save the estimates ˆUðLÞ and ˆc and the residual series
fˆeKþ 1; . . .;ˆen g;
S2. Generate artificial yt data fyKþ1 ; . . .; yn g from the VAR by using ˆUðLÞ, ˆc and block
random draws from fˆeKþ 1; . . .;ˆen g;
S3. Estimate the VAR and the corresponding function of interest F with the artificial
data fyKþ1 ; . . .; yn g;
 J. Wang, M. Yang / Journal of Financial Markets 14 (2011) 82–108 105

S4. Repeat S2 and S3 many times to construct an empirical distribution, hence the
standard errors, for the estimated F.

The block draws in S2 are designed to take into account the possible dependence
structure in the et series. The size of the block needs to be chosen appropriately: a too-small
size destroys the dependence structure in the et series; a too-large size leads to a loss of
variation in draws. Empirically we find that the results are not sensitive to the block size.
The reported results are based on a block size approximately equal the square root of the
sample size. The initial values for yt in S2 are obtained by randomly drawing a K-block
from the original series {y1, y, yn}. Given that our sample sizes are reasonably large, the
above bootstrap procedure is adequate for estimating standard errors of information
shares. Berkowitz and Kilian (2000) provide a review of time series bootstrapping.

Appendix B. Rotation of markets

The structural VAR in (4) may be written as

BðLÞrt ¼ gt ; BðLÞ ¼ B0 AðLÞ ¼ B0 B1 L    BK LK ;

" ð0Þ #
b11 0
B0 ¼ ð0Þ ; ðB:1Þ
b21 bð0Þ
22

where Bk ¼ B0 Ak . The impact coefficient vector is given by h ¼ ½h1 ; h2  ¼ 0 Bð1Þ1 . We

show that the information measure in (8) is invariant to the choice of market 1, by
verifying that a rotation of the positions of the two markets causes exactly the same
rotation of the elements in the h-vector. Let bij(L) be the (i, j)-element of B(L) and bðkÞ
ij the
(i, j)-element of Bk. We re-write (B.1) as
b21 ðLÞr1;t þ b22 ðLÞr2;t ¼ Z2;t ð2nd equation for tÞ
ðB:2Þ
b11 ðLÞr1;tþ1 þ ½b12 ðLÞ=Lr2;t ¼ Z1;tþ1 ð1st equation for t þ 1Þ

where b12(L)/L is a lag polynomial of order K1. Suppose that the original market 2 is
shifted to the first position (i.e., the start of a day is now shifted to the start of the original
market 2). Denote the open-to-close returns of the new setting as R1,t=r2,t and R2,t=r1,tþ1.
Denote the corresponding structural shocks as z1,t=Z2,t and z2,t=Z1,tþ1. The new system,
according to (B.2), has the following VAR representation:
b22 ðLÞR1;t þ ½b21 ðLÞLR2;t ¼ z1;t ;
ðB:3Þ
½b12 ðLÞ=LR1;t þ b11 ðLÞR2;t ¼ z2;t ;

where the matrix lag polynomial is

" # " # " #
b22 ðLÞ b21 ðLÞL b22 ðLÞ b21 ðLÞ bð0Þ
22 0
with ¼ :
b12 ðLÞ=L b11 ðLÞ b12 ðLÞ=L b11 ðLÞ
L¼0
bð1Þ
12 bð0Þ
11

Hence the general structural of (B.3) is the same as that of (B.1), although the lag lengths
of (B.3) and (B.1) are different. Now the information measures based on (B.3) are
106 J. Wang, M. Yang / Journal of Financial Markets 14 (2011) 82–108

computed from its own impact coefficient vector

" #1 " #1
0
b 22 ðLÞ b 21 ðLÞL b 22 ð1Þ b 21 ð1Þ
¼ 0 ¼ ½h2 ; h1  ðB:4Þ
b12 ðLÞ=L b11 ðLÞ b12 ð1Þ b11 ð1Þ
L¼1

which is exactly the rotated h-vector of the original system. This result can be readily
extended to the cases with m sequential markets by recognizing that a shift of the start of
the 24-hour trading cycle is equivalent to pre-multiplying the rotation matrix
" #
0 Im1
TðLÞ ¼
1=L 00

to the VAR in (B.1), where Im1 is the identity matrix of size m1. The impact coefficient
vector of the shifted system is given by h[T(1)]1, which is exactly the shifted version of the
original h-vector.

References

Agarwal, S., Liu, C., Rhee, S.G., 2007. Where does price discovery occur for stocks traded in multiple markets?
Evidence from Hong Kong and London. Journal of International Money and Finance 26, 46–63.
Aı̈t-Sahalia, Y., Mykland, P.A., Zhang, L., 2005, How often to sample a continuous time process in the presence
of market microstructure noise. Review of Financial Studies 18, 351-416.
Aı̈t-Sahalia, Y., Mancini, L., 2008. Out of sample forecasts of quadratic variation. Journal of Econometrics 147,
17–33.
Andersen, T.G., Bollerslev, T., Das, A., 2001. Variance ratio statistics and high frequency data: testing for
changes in intraday volatility patterns. Journal of Finance 56, 305–327.
Andersen, T.G., Bollerslev, T., Diebold, F.X., Labys, P., 2003. Modeling and forecasting realized volatility.
Econometrica 71, 579–625.
Andersen, T.G., Benzoni, L., 2008. Realized volatility, Chapter prepared for the Handbook of Financial Time
Series. Springer Verlag.
Bandi, F.M., Russell, J.R., 2006. Separating microstructure noise from volatility. Journal of Financial Economics
79, 655–692.
Bank for International Settlements, 2007, Foreign Exchange and Derivatives Market Activity in 2007. Triennial
Central Bank Survey.
Barclay, M., Hendershott, T., 2003. Price discovery and trading after hours. Review of Financial Studies 16,
1041–1073.
Barclay, M., Hendershott, T., 2004. Liquidity externalities and adverse selection: evidence from trading after
hours. Journal of Finance 59, 681–710.
Barclay, M., Hendershott, T., 2008. A comparison of trading and non-trading mechanisms for price discovery.
Journal of Empirical Finance 15, 839–849.
Barndorff-Nielsen, O.E., Hansen, P.R., Lunde, A., Shephard, N., 2005. Regular and modeified kernel-based
estimators of integrated variance: the case of independent noise. Working Paper.
Berkowitz, J., Kilian, L., 2000. Recent developments in bootstrapping time series. Econometric Reviews 19, 1–48.
Beveridge, S., Nelson, C.R., 1981. A new approach to the decomposition of economic time series into permanent
and transitory components with particular attention to the measurement of the ‘‘Business Cycle’’. Journal of
Monetary Economics 7, 151–174.
Bjonnes, G., Rime, D., Solheim, H., 2005. Volume and volatility in the FX market: does it matter who you are?
In: Paul De, Grauwe (Ed.), Exchange Rate Economics: Where do we Stand? MIT Press.
Booth, G.G., So, R., Tse, Y., 1999. Price discovery in the German equity derivatives markets. Journal of Futures
Markets 19, 619–643.
Cai, J., Cheung, Y.L., Lee, R., Melvin, M., 2001. Once-in-a-generation Yen volatility in 1998: fundamentals,
intervention, or order flow?. Journal of International Money and Finance 20, 327–347.
Cao, C., Ghysels, E., Hatheway, F., 2000. Price discovery without trading: evidence from the Nasdaq preopening.
Journal of Finance 55, 1339–1365.
 J. Wang, M. Yang / Journal of Financial Markets 14 (2011) 82–108 107

Cao, C., Hanscah, O., Wang, X., 2008. The information content of an open limit order book. Journal of Futures
Markets 29, 16–41.
Carpenter, A., Wang, J., 2007. Herding and the information content of trades in the Australian dollar market.
Pacific-Basin Finance Journal 15, 173–194.
Chakravarty, S., Gulen, H., Mayhew, S., 2004. Informed trading in stock and option markets. Journal of Finance
59, 1235–1257.
Chu, Q., Hsieh, W., Tse, Y., 1999. Price discovery on the S&P 500 index markets: an analysis of spot index: Index
Futures and SPDRs. International Review of Financial Analysis 8, 21–34.
Cliff, M., Cooper, M.J., Gulen, H., 2008, Return Differences between Trading and Non-trading Hours: Like
Night and Day. Manuscript.
Corsi, F., 2009. A simple approximate long-memory model of realized volatility. Journal of Financial
Econometrics 7, 174–196.
Covrig, V., Ding, D.K., Low, B.S., 2004. The contribution of a satellite market to price discovery: evidence from
the Singapore exchange. Journal of Futures Markets 24, 981–1004.
Covrig, V., Melvin, M., 2002. Asymmetric information and price discovery in the FX market: does Tokyo know
more about the Yen? Journal of Empirical Finance 9, 271–285.
Davies, R., 2003. The Toronto stock exchange preopening session. Journal of Financial Markets 6, 491–516.
Evans, M., Lyons, R., 2002a. Order flow and exchange rate dynamics. Journal of Political Economy 110, 170–180.
Evans, M., Lyons, R., 2002b. Informational integration and FX trading. Journal of International Money and
Finance 21, 807–831.
Evans, M., Lyons, R., 2005, Meese-Rogoff redux: micro-based exchange rate forecasting. American Economic
Review P&P, May.
Evans, M., Lyons, R., 2007, Exchange rate fundamentals and order flow. NBER Working Paper No. W13151.
Evans, M., Lyons, R., 2008. How is macro news transmitted to exchange rates? Journal of Financial Economics
88, 26–50.
Figuerola-Ferretti, I., Gonzalo, J., 2010. Modelling and measuring price discovery in commodity markets. Journal
of Econometrics 158, 95–107.
French, K.R., 1980. Stock returns and the weekend effect. Journal of Financial Economics 8, 55–69.
Gonzalo, J., Granger, C., 1995. Estimation of common long-memory components in cointegrated systems.
Journal of Business and Economic Statistics 13, 27–35.
Grammig, J., Melvin, M., Schlag, C., 2005. Internationally cross-listed stock prices during overlapping trading
hours: price discovery and exchange rate effects. Journal of Empirical Finance 12, 139–164.
Hansen, P.R., Lunde, A., 2006. Realized variance and market microstructure noise. Journal of Business and
Economic Statistics 24, 127–218.
Harris, F., McInish, T.H., Wood, R.A., 2002. Security price adjustment across exchanges: an investigation of
common factor components for dow stocks. Journal of Financial Markets 5, 277–308.
Harris, F., McInish, T.H., Wood, R.A., 2008. The impact of Reg NMS on price discovery: a unified approach to
CFS and IS measures. Working Paper, Wake Forest University.
Hasbrouck, J., 1995. One security, many markets: determining the contributions to price discovery. Journal of
Finance 50 (4), 1175–1199.
Hsieh, D., Kleidon, A.W., 1996. Bid-Ask spreads in foreign exchange markets: implications for models of
asymmetric information. In: Frankel, J., Galli, G., Giovannini, A. (Eds.), The Microstructure of Foreign
Exchange Markets. National Bureau of Economic Research, Cambridge.
Hupperets, E.C.J., Menkveld, A.J., 2002. Intraday analysis of market integration: Dutch blue chips traded in
Amsterdam and New York. Journal of Financial Markets 5, 57–82.
Ito, T., Lyons, R., Melvin, M., 1998. Is there private information in the FX market? The Tokyo experiment.
Journal of Finance 53, 1111–1130.
Lehmann, B., 2002. Some desiderata for the measurement of price discovery across markets. Journal of Financial
Markets 5, 259–276.
Lieberman, O., Ben-Zion, U., Hauser, S., 1999. A characterization of the price behavior of international dual
stocks: an error correction approach. Journal of International Money and Finance 18, 289–304.
Lyons, R., 2001. In: The Microstructure Approach to Exchange Rates. MIT Press, Cambridge.
Meese, R.A., Rogoff, K., 1983. Empirical exchange rate models of the seventies: do they fit out of sample? Journal
of International Economics 14, 3–24.
Menkveld, A., Koopman, S., Lucas, A., 2007. Modeling round-the-clock price discovery for cross-listed stocks
using state space methods. Journal of Business & Economic Statistics 25, 213–225.
108 J. Wang, M. Yang / Journal of Financial Markets 14 (2011) 82–108

Newey, W.K., West, K.D., 1994. Automatic lag selection in covariance matrix estimation. Review of Economic
Studies 61, 631–653.
Nishimura, K., 2007. Increased diversity and deepened uncertainty: policy challenges in a zero-inflation economy.
International Finance 10 (3), 281–300.
Not-yet-desperate Housewives. Economist, August 16, 2007.
Pascual, R., Pascual-Fuster, B., Climent, F., 2006. Cross-listing, price discovery and the informativeness of the
trading process. Journal of Financial Markets 9, 144–161.
Yan, B., Zivot, E., 2007, The Dynamics of price discovery. Working Paper, University of Washington.
Yan, B., Zivot, E., 2010, A structural analysis of price discovery measures. Journal of Financial Markets 13, 1–19.
Zhang, L., Mykland, P., Aı̈t-Sahalia, Y., 2005. A tale of two time scales: determining integrated volatility with
noisy high-frequency data. Journal of the American Statistical Association 100, 1394–1411.