Upload
others
View
5
Download
0
Embed Size (px)
Citation preview
Are Volatility Over Volume Liquidity Proxies Useful For Global Or US Research?*
Kingsley Y. L. Fong** Craig W. Holden** Ondrej Tobek**
University of New South Wales Indiana University University of Cambridge
March 2018
Abstract
We examine a general class of volatility over volume liquidity proxies as computed from low frequency
(daily) data. We start from the Kyle and Obizhaeva (2016) hypothesis of transaction cost invariance to
identify a new volatility over volume liquidity proxy “VoV(%Spread)” for percent spread cost and a new
volatility over volume liquidity proxy “VoV(λ)” for the slope of the transaction cost function “λ”. We test
the monthly and daily versions of these new and existing liquidity proxies against liquidity benchmarks as
estimated from high frequency (intraday) data on both a global and US basis. We find that both the monthly
and daily versions of VoV(λ) dominate the equivalent versions of Amihud and other cost-per-dollar-volume
proxies on both a global and US basis. We also find that both the monthly and daily versions of
VoV(%Spread) dominate the equivalent versions of other percent-cost proxies for US studies that cover
pre-1993 years. In a case study, we find that our new VoV liquidity proxies yield different research
inferences than the best previous liquidity proxies from the prior literature. The success of our invariance-
based liquidity proxies across exchanges and over time supports the prediction of Kyle and Obizhaeva of a
specific functional form for transaction costs across exchanges and over time.
JEL classification: C15, G12, G20. Keywords: Liquidity, transaction costs, effective spread, price impact.
* We are thank seminar participants at the 2017 SFS Asia-Pacific Cavalcade conference and the 2017
Financial Research Network conference. We are solely responsible for any errors.
** Kingsley Y.L. Fong email: [email protected], Craig W. Holden emai: cholden@ indiana.edu, and
Ondrej Tobek email: [email protected]. Correspondence address: Craig W. Holden, Kelley School of
Business, Indiana University, 1309 E. Tenth St., Bloomington, IN 47405. Phone: 1-812-855-3383.
1
1. Introduction
We examine a general class of volatility over volume (VoV) liquidity proxies as computed from
low frequency (daily) data. This builds on a recent literature that proposes the use of low-frequency liquidity
proxies. The idea is that low-frequency liquidity proxies can do a good job of capturing liquidity
benchmarks as computed from high-frequency (intraday) data and at the same time provide enormous
computational savings compared to the use of high-frequency data. Fong, Holden, and Trzcinka (FHT)
(2017) show that these computational savings for global research are of the order of 1,000X to 10,000X
depending on the specific decade being analyzed.1
Kyle and Obizhaeva (2016) develop the empirical hypothesis of “market microstructure
invariance.” This includes the sub-hypothesis of “transaction cost invariance,” which means that transaction
costs per unit of business time have a constant distribution across all assets and over all time periods. We
start from the specific functional forms of their hypothesis to identify two new liquidity proxies: (1)
“VoV(%Spread)” which proxies for the percent effective spread and (2) “VoV(λ)” which proxies for the
slope of the transaction cost function (typically called λ).
Liquidity proxies are important because liquidity is connected in fundamental ways to asset pricing,
corporate finance, and market microstructure. A large literature on global and US research relies on monthly
liquidity proxies2 and daily liquidity proxies.3
1 FHT compute the ratio of high-frequency data points to low-frequency data points in their sample. Due to the
exponential increase in high-frequency data, this ratio rose from 42X in 1996 to 962X in 2007 to 12,616X in 2014. 2 Monthly liquidity proxies are used in Jain (2005), Lesmond (2005), Stahel (2005), Bailey, Karolyi, and Salva,
(2006), Levine and Schmukler (2006), Bekaert, Harvey, and Lundbland (2007), LaFond, Lang, and Skaife (2007),
Chan, Jain, and Xia (2008), Henkel (2008), Henkel, Jain, and Lundblad (2008), DeNicolo and Ivaschenko (2009),
Griffin, Kelly, and Nardari (2010), Hearn, Piesse, and Strange (2010), Clark (2011), Griffin, Hirschey, and Kelly
(2011), Lee (2011), Lang, Lins, and Maffett (2012), Liang and Wei (2012), Asness, Moskowitz, and Pedersen
(2013), Marshall, Nguyen, and Visaltanachoti (2013), and Bekaert, Harvey, Lundblad, Siegel (2014), Hearn (2014),
Karnaukh, Renaldo and Soderlind (2015), Orlov (2016), and Massa, Mataigne, Vermaelen, and Xu (2017). 3 Daily liquidity proxies are used in Bhattacharya, Daouk, Jorgenson, and Kehr (2000), Attig, Gadhoum, and Lang
(2003), Gomez-Puig (2006), Gersl and Komarkov (2009), Erten and Okay (2012), Karolyi, Lee, and van Dijk
2
Using a dataset very similar to FHT, we test the top-three monthly and daily versions of low-
frequency liquidity proxies considered by FHT plus several new proxies not considered there. We test all
of these proxies against the corresponding monthly and daily versions of high-frequency liquidity
benchmarks on a global basis and in the US. We test the performance of monthly and daily liquidity proxies
on five dimensions of performance that are crucial for asset pricing, corporate finance and market
microstructure applications: (1) higher average cross-sectional correlation with the benchmarks, (2) higher
average Spearman’s cross-sectional correlation with the benchmarks, (3) higher portfolio time-series
correlation of proxy first differences with benchmark first differences, (4) higher individual stock time-
series correlation of proxy first differences with benchmark first differences, and (5) lower average root
mean squared error (RMSE) with the benchmarks.
For global research, we find that monthly VoV(λ) dominates monthly Amihud and five other
monthly cost-per-dollar-volume proxies by a statistically and economically significant amount. We also
find that daily VoV(λ) dominates daily Amihud and three other daily cost-per-dollar-volume proxies by a
statistically and economically significant amount. For US studies covering 1993-present, we find that
monthly (daily) VoV(%Spread) ties monthly (daily) Closing Percent Quoted Spread. For US studies
including pre-1993 years, when Closing Percent Quoted Spread is not available, we find that monthly
(daily) VoV(%Spread) dominates all other monthly (daily) percent-cost proxies tested by a statistically and
economically significant amount.
Specifically, on a global basis we find that monthly (daily) VoV(λ) has the highest average cross-
sectional correlation, highest average Spearman’s cross-sectional correlation, highest portfolio time-series
correlation of first differences, and highest individual stock time-series correlation of first differences with
monthly (daily) high-frequency benchmark lambda for both time-periods analyzed. The gain in
performance is economically meaningful. Specifically, monthly VoV(λ) has an average correlation across
(2012), Beber and Pagano (2013), Lee, Tseng, and Yang, (2014), Hanselaar, Stulz, and van Dijk (2016), Lee,
Sapriza, and Wu (2016), and Boubakri, Saad, and Samet (2017).
3
the eight contests (i.e., four correlation measures X two time-periods) of 0.562 versus 0.464 for monthly
Amihud. Daily VoV(λ) has an average correlation across the eight contests of 0.534 versus 0.444 for daily
Amihud. We find that the dominant performance of monthly VoV() is robust by time-period, by size
quintiles, by volatility quintiles, net of controlling for size and volatility, and by exchange. We find that
none of the monthly (daily) cost-per-dollar-volume proxies captures the level of monthly (daily) lambda.
Qualitatively similar results hold for the US.
For US studies including pre-1993 years (for which data is available back to 1926), we find that
monthly (daily) VoV(%Spread) has the highest average cross-sectional correlation, highest average
Spearman’s cross-sectional correlation, highest portfolio time-series correlation of first differences, and
highest individual stock time-series correlation of first differences with monthly (daily) high-frequency
benchmark percent effective spread. The gain in performance is economically meaningful. Specifically,
monthly VoV(%Spread) has an average correlation across the four contests (i.e., four correlation measures)
of 0.705 versus 0.593 for the second-best monthly percent-cost proxy. Daily VoV(%Spread) has an average
correlation across the four contests of 0.584 versus 0.206 for the second-best daily percent-cost proxy. We
find that monthly (daily) VoV(%Spread) does the best job of capturing the level of monthly (daily) percent
effective spread.
Do our new VoV liquidity proxies yield different research inferences than the best previous
liquidity proxies from the prior literature? To address this question, we perform a case study that compares
daily VoV() and daily Amihud when analyzing the pattern of S&P 1500 illiquidity before, during, and
after the August 2007 “Quant Meltdown” event. We find that the daily VoV() proxy yields the inference
that S&P 1500 illiquidity is significantly higher than the base period in all three periods (before, during,
and after). By contrast, the daily Amihud proxy yields the inference that S&P 1500 illiquidity in not
significantly different than the base in all three periods. We conclude that our new VoV liquidity proxies
matter.
4
We contribute to two lines of literature. First, contribute to the liquidity proxy literature by
developing four new liquidity proxies (monthly and daily VoV(%Spread) and monthly and daily VoV())
and by comprehensively testing them against many alternatives on both a global and US basis. Compared
with FHT, we analyze their top monthly and daily percent-cost proxies, but also analyze several that they
do not: VoV(%Spread), Gibbs, and Close-High-Low. Further, we analyze their top monthly and daily cost-
per-dollar-volume proxies, but also analyze several that they do not: the VoV(), Gibbs Impact, and Close-
High-Low Impact. Despite the expanded competition, all four of our new proxies prove to be the best for
one or more purposes.
Second, the success of our invariance-based liquidity proxies across exchanges and over time
supports the prediction of Kyle and Obizhaeva of a specific functional form for transaction costs across
exchanges and over time. Furthermore, we find there is very limited performance gain and there is risk of
overfitting the data when we allow the data to determine the optimal parameter in the general class of
volatility over volume proxies from the predictions of the transaction cost invariance hypothesis.
The paper is organized as follows. Section 2 explains the volatility over volume liquidity proxies.
Section 3 describes the high-frequency liquidity benchmarks. Section 4 reviews the low-frequency proxies
form the existing literature. Section 5 describes the data. Sections 6 presents global performance results.
Section 7 presents US performance results. Section 8 performs a case study of whether the new liquidity
proxies yield different research inferences than the best previous liquidity proxies. Section 9 concludes.
Appendix A derives VoV(%Spread) and VoV() from transaction cost invariance formulas in Kyle and
Obizhaeva. Appendix B summarizes the formulas of low-frequency proxies from the existing literature. An
online supplemental appendix provides additional robustness checks by exchange.
2. Volatility Over Volume Liquidity Proxies
We examine a general class of volatility over volume liquidity proxies in which a function of
volatility is in the numerator and a function of volume is in the denominator. Let be the standard
deviation of daily returns and V be average daily dollar volume in real dollars (i.e., inflation-adjusted). We
5
define volatility over volume (VoV) class of liquidity proxies as liquidity proxies that have the following
form
b
c
aVoV
V
, (1)
where a, b, and c are strictly positive constants.
The key idea of Kyle and Obizhaeva is that markets for different assets run at different speeds,
which they call “business time.” Business time is the speed with which “bets” arrive at the market. A bet is
the total amount of risk transfer that is desired to be traded by a given trader. Thus, their key hypothesis of
“market microstructure invariance” is that all markets look the same when measured per unit of business
time. Specifically, this hypothesis has two sub-hypotheses. The first is “bet invariance,” which means that
bets per unit of business time have a constant distribution across all assets and over all time periods. The
second is “transaction cost invariance,” which means that transaction costs per unit of business time have
a constant distribution across all assets and over all time periods.
Transaction cost invariance supports alternative functional forms. Kyle and Obizhaeva consider
three alternative forms: (1) a constant percent quoted spread cost, (2) linear price impact costs, and (3)
square-root price impact costs. The appendix shows how the constant percent quoted spread cost case leads
to the following volatility over volume percent-cost proxy VoV(%Spread)
2
3
1
3
%VoV Spreada
V
, (2)
where a is a constant, 2
,3
b and 1
3c . In our empirical implementation, we set
2,
3b
1
3c , and
convert V to real dollars4 and then find the value of a that matches the mean value of monthly
4 We convert to real dollars as follows: Real Dollar Volume = Nominal Dollar Volume / CPI. For CPI we use the “Consumer Price Index for All Urban Consumers: All Items (CPIAUCNS)” available from the online FRED database provide by the Federal Reserve Bank of St Louis and normalize it to 1.0 on January 1, 2000 (i.e, we divide by the value of CPIAUCNS on 1/1/2000).
6
VoV(%Spread) with monthly percent effective spread in our US sample. We obtain parameter a = 8.0 and
use this parameter value for both the monthly and daily versions and when analyzing both the global and
US samples.5
The appendix also shows how the square-root price impact cost case leads to the following volatility
over volume cost-per-dollar-volume proxy VoV()
1
2
VoVa
V
, (3)
where a is a constant, 1,b and 1
2c . In our empirical implementation, we set 1,b and
1
2c and then
find the value of a that matches the mean value of VoV() with lambda in our global sample. We obtain
parameter a = 1 and use this parameter value for both the monthly and daily versions and when analyzing
both the global and US samples.6
The volatility over volume class of liquidity proxies encompasses a wide range of special cases.
When 1,a b c it bears a general resemblance to the well-known Amihud liquidity proxy, which is
given by
t
t
rAverage
V
, (4)
where tr is the daily return on day t , tV is the dollar volume on day t , and the average is computed over
all positive volume days. The numerator tr of Amihud is similar to numerator of the VoV class of
liquidity proxies with 1b , in that both are measures of volatility.
An important difference is that the VoV class of liquidity proxies is the ratio of two statistics,
whereas the Amihud liquidity proxy is the average of daily ratios which may include noisy values when
5 a = 8.0 is very close to 7.62 estimated in Kyle and Obizhaeva for fixed costs component of their linear version of cost function. 6 a = 1.0 is again very close to 1.2 estimated in Kyle and Obizhaeva for variable costs component of their square root version of cost function.
7
outlier low-volume days occur. This point is easily illustrated by a simple numerical example. Suppose that
the returns and dollar volume for a given stock are as shown in Table 1. On days 1-4, the dollar volume is
between $200,000 and $300,000, but on day 5a it was just $5,000, whereas in the alternative case on day
5b it is $250,000. The column labeled “Amihud ratio” is the daily ratio of absolute return over dollar
volume. Notice that the low volume on day 5a causes the Amihud ratio on day 5a to be much larger than
the other days. The Amihud liquidity proxy when computed over days 1 to 4 and 5a (and multiplied by
1,000,000) is 1.69. The analogous Amihud liquidity proxy computed over days 1 to 4 and 5b is 0.12. The
former is 14.2 times larger than the later. Thus, Amihud yields very noisy results when outlier low-volume
days occur. By contrast, VoV(λ) computed over days 1 to 4 and 5a is 76.52 and 68.33 when computed over
days 1 to 4 and 5b. It is robust to the outlier realization of low-volume days.
3. High-Frequency Liquidity Benchmarks
We test our new volatility over volume liquidity proxies and other liquidity proxies from the
existing literature against high-frequency liquidity benchmarks. Specifically, we following Fong, Holden,
and Trzcinka (2017) in distinguishing between the percent-cost benchmarks vs. cost-per-dollar-volume
benchmarks. A percent-cost benchmark measures the percentage cost of doing a small trade. By contrast, a
cost-per-dollar-volume benchmark measures the marginal transaction cost per US dollar of volume.
Our percent-cost benchmark is percent effective spread. The percent effective spread of a given
stock on the trade is defined as
Percent Effective Spreadk = 2Dk (ln(Pk) - ln(Mk)), (1)
where Dk is an indicator variable that equals +1 if the thk trade is a buy and -1 if the thk trade is a sell, Pk
is the price of the thk trade and Mk is the midpoint of the consolidated best bid and offer prevailing
immediately prior to the time of the thk trade, where the exact amount of time prior depends on the time-
stamp of each exchange. Next, we aggregate over either a day or month, which in either case we refer to as
period i. A stock’s Percent Effective Spreadi for period i is the volume-weighted average of Percent
Effective Spreadk computed over all trades in period i.
thk
8
Our cost-per-dollar-volume benchmark is the slope of the price function, which is typically called
. We follow Goyenko, Holden, and Trzcinka (2009) and Hasbrouck (2009) in calculating as the slope
coefficient of the regression
n n nr S u , (5)
as estimated over five minute intervals. For the thn five-minute interval, nr is the stock return, nS =
kn knkSign v v
is the signed square-root of volume measured in US dollars, knv is the signed volume
measured in US dollars of the thk trade in the thn five-minute interval, and nu is the error term.7
4. Low-Frequency Proxies from the Existing Literature
We analyze five monthly percent-cost low-frequency proxies from the existing literature. This
includes what FHT finds to be the top-three performing monthly percent-cost proxies: “Closing Percent
Quoted Spread” from Chung and Zhang (2014), “High-Low” from Corwin and Schultz (2012), and “FHT”
from Fong, Holden, and Trzcinka (2017). We also include two leading percent-cost proxies that were not
analyzed by FHT. FHT explain that they did not include the “Gibbs” proxy from Hasbrouck (2004) because
it is very computationally-intensive. “Close-High-Low” from Abdi and Ranaldo (2017) was developed after
the FHT paper.
We analyze six monthly cost-per-dollar-volume low-frequency proxies from the existing literature.
This includes ‘‘Amihud’’from Amihud (2002) plus five proxies based on the “Extended Amihud Class” of
Goyenko, Holden, and Trzcinka (2009), Section 5.2. These five proxies are found by dividing five of the
monthly percent-cost proxies mentioned above by the average US dollar value of daily volume.
Specifically, these five proxies are: Gibbs Impact, FHT Impact, High-Low Impact, Closing-High-Low
Impact, and Closing Percent Quoted Spread Impact.
7 Noss et al. (2017) run similar benchmarking exercise in FOREX markets and found similar results as in our study. They benchmark the proxies to steepness of order book throughout the day which can be computed without any measurement error and does not depend on functional specification. This adds further robustness to our findings.
9
Most of the monthly liquidity proxies cannot be computed on a daily basis, because they require
multiple daily observations to compute the proportion of zero return days, to run a regression, or compute
a maximum likelihood estimation. Therefore, we analyze three daily percent-cost low-frequency proxies
from the prior literature: “High-Low,” “Closing-High-Low,” and “Closing Percent Quoted Spread.”
Similarly, we analyze four daily cost-per-dollar-volume low-frequency proxies from the prior literature:
“High-Low Impact,” “Closing-High-Low Impact,” “Closing Percent Quoted Spread Impact,” and
“Amihud.”
5. Data
We obtain intraday trades, intraday quotes, and securities identifiers for non-US stocks from the
Thomson Reuters Tick History (TRTH) database, which is provided by the Securities Industry Research
Center of Asia-Pacific. We obtain other non-US data on returns, size, securities, individual security
information, and daily exchange rates from Datastream. We obtain intraday trade and quote data for US
stocks from the Trade and Quote database provided by the New York Stock Exchange. We obtain returns
and size data from the Center for Research in Security Prices (CRSP) and Compustat.
Our earlier sample covers 42 exchanges around the world from 1996 to 2007. It includes the leading
exchange by volume in 36 countries and three exchanges in China (the Hong Kong Stock Exchange,
Shanghai Stock Exchange, and Shenzhen Stock Exchange) and three exchanges in the US (the New York
Stock Exchange, American Stock Exchange, and NASDAQ). ). Given the large number of stocks and large
amount of data in the US market, we select a random sample of 400 firms out of the universe of all eligible
US firms in 1996, replace any firms that are ineligible in 1997 with randomly drawn firms out of the
universe of all eligible US firms in 1997, and so on rolling forward to 2007. Following the methodology of
Hasbrouck (2009), a stock must meet five criteria to be eligible: (1) it has to be a common stock, (2) it has
to be present on the first and last TAQ master file for the year, (3) it has to have the NYSE, AMEX or
NASDAQ as the primary listing exchange, (4) it does not change primary exchange, ticker symbol or cusip
over the year, and (5) has to be listed in CRSP.
10
Our later sample cover the same 42 exchanges from 2008-2014. We select a random sample of 30
stocks per exchange – stratified by size tercile – out of all eligible firms on a given exchange in 2008. Our
stratification is to randomly select 10 large stocks, 10 medium stocks, and 10 small stocks from each
exchange. We replace any firms that are ineligible in 2009 with a randomly drawn firm from the same
exchange and the same size tercile, and repeat this process rolling forward year-by-year to 2014.
We impose several filters in order to have reliable and consistent proxy estimates. First, we require
that a stock have at least five positive-volume days and 11 non-zero return days in the month. The daily
sample is based on the stock-days contained within the set of stock-months that have at least five positive-
volume days and 11 non-zero return days. Second, for Datastream we follow the recommendation of Ince
and Porter (2006) to remove any stock-month with extreme return reversal. Finally, we winsorize the
sample at both ends at 1% level. That is, we set the values of proxies that are above the 99th percentile value
equal to the 99th percentile value and set the values of proxies below 1st percentile value equal to the 1st
percentile value. Our resulting sample is slightly larger than the FHT dataset.
For non-US exchanges, we convert the local currency value of volume into US dollars equivalent
using the daily (average over the month) exchange rate for daily (monthly) proxies. For VoV(%Spread), we
adjust dollar volume for inflation by dividing by the US CPI, which in turn has been normalized to equal
1.0 on 1/1/2000. To illustrate the importance of adjusting dollar volume for inflation over long periods of
time, Figure 1 plots the VoV(%Spread) when dollar volume is adjusted for inflation, the VoV(%Spread)
when dollar volume is not adjusted for inflation, and Close-High-Low from 1926 to 2015. We compare to
the Close-High-Low proxy, because Abdi and Ranaldo (2017) show it to be the best percent-cost proxy
from the existing literature for US research when including pre-1993 years. We find that the inflation-
adjusted version of VoV(%Spread) is very close to the Close-High-Low proxy over the entire 90 year span,
11
whereas the unadjusted version of VoV(%Spread) is more than 2X off from 1926-1950 due to backwards-
in-time cumulative effect of inflation raised to the 1/3 power in the denominator of VoV(%Spread).8
For VoV(λ), we do not adjust dollar volume for inflation (i.e., we use nominal dollar volume). Our
reason for not doing so is because the RHS of the benchmark regression in equation (5) uses nominal dollar
volume and all of the competing cost-per-dollar-volume proxies use nominal dollar volume. In practical
terms, none of the VoV(λ) performance results are materially affected by this issue.
For the daily VoV proxies, we use the Parkinson (1980) volatility formula. In the case of a one-day
estimator (T=1), the formula becomes: 8 ln H L , where H is the daily high price and L is the
daily low price.9
Our earlier, high-frequency sample has 8.0 billion trades and 17.7 billion quotes. We compute the
corresponding benchmarks and proxies for 24,419 firms in 1,500,611 stock-months. Our later, high-
frequency sample has 1.8 billion trades and 14.7 billion quotes. We compute the corresponding benchmarks
and proxies for 3,087 firms in 84,789 stock-months.
Table 2 provides some descriptive statistics for the global and US samples of both monthly and
daily percent-cost proxies and benchmarks and both monthly and daily cost-per-dollar-volume proxies and
benchmarks. For the global sample, descriptive statistics are provided for both the earlier sample (1996-
2007) and the later sample (2008-2014). For the US sample, descriptive statistics are provided for the
combined sample period (1996-2014). Panel A shows that the means of all of the monthly percent-cost
proxies are in the same ballpark as the mean of percent effective spread and that this is true for both time
periods. Panel B shows that the means of all of the daily percent-cost proxies are also in the same ballpark
8 Another adjustment that we make to VoV(%Spread) is that we multiply the volume from Datastream in India by 10. The total volume for Indian exchanges in the Datastream is roughly 10 times lower than the value in the World Federation of Exchanges. 9 Following Corwin and Schultz (2012), we use the last valid high and low prices for a given stock for days where high is equal to low. We define valid high and low prices as prices on days where the stock was traded and high was not equal to low.
12
as the mean of daily percent effective spread for both time periods. Panel C shows that the same is true in
the US sample at both monthly and daily frequencies.
Panel D shows that the medians of all of the cost-per-dollar-volume proxies are far different than
the median of lambda. Panel E shows that medians of the daily versions of all of the cost-per-dollar-volume
proxies are far different than the median of lambda, except for the median of VoV() in the earlier sample,
which is reasonably close to the median of lambda in that period. Panel F shows that in the US sample the
medians of all of the cost-per-dollar-volume proxies are far different from the median of lambda at both
monthly and daily frequencies. In summary, the monthly and daily percent-cost proxies are roughly on the
same level as the benchmark, but the monthly and daily cost-per-dollar-volume proxies get the level wrong.
6. Global Performance Results
6.1 Monthly Cost-Per-Dollar-Volume Proxies
Table 3 shows the global performance of monthly cost-per-dollar-volume proxies compared to the
monthly cost-per-dollar-volume benchmark, lambda. Five panels report five performance dimensions
(average cross-sectional correlation, average Spearman’s cross-sectional correlation, portfolio time-series
correlation of first differences, individual stock time-series correlation of first differences, and average
RMSE) and a sixth panel provides a summary statistic across all of the correlation results. Monthly Amihud
is by far the most widely used cost-per-dollar-volume proxy from the existing literature. As a rough
indicator of its enormous influence, we note that Amihud (2002) has more than 5,000 Google scholar
citations. Thus, in the analysis below we often highlight the relative performance of our new cost-per-dollar
proxy, monthly VoV(), and monthly Amihud.
Panel A reports the average cross-sectional correlation of seven monthly cost-per-dollar-volume
proxies compared to monthly lambda for both time-periods. The average cross-sectional correlations are
computed in the spirit of Fama and MacBeth (1973) by calculating the cross-sectional correlation each
month across all firms on a given exchange and then calculating the average correlation value over all
exchange-months to get the global average. For the entire paper, we adopt the convention of placing a solid
13
box around the highest correlation in the row and a dashed box around any correlations that are statistically
indistinguishable from the highest correlation in the row at the 5% level.10 The idea is to determine the best
proxy on a given performance dimension and the full set proxies that are statistically indistinguishable from
the best. If a correlation is bold-faced, this means that it is statistically different from zero at the 5% level.11
We find that monthly VoV() dominates all six other monthly cost-per-dollar-volume proxies in
both time-periods. It has the highest correlations in both time-periods and is statistically significantly higher
than the correlations of any other proxy in both rows. There is economically large difference in performance
between monthly VoV() and the other monthly proxies.
Figure 2 shows the global average cross-sectional correlations between monthly cost-per-dollar-
volume proxies and monthly lambda over time. We see that monthly VoV()’s cross-section correlation is
higher than monthly Amihud’s cross-sectional correlation in every single month over the 19-year sample
period. In most months the gap in performance is large, which makes it an economically significantly
improvement as well.
Returning to Table 3, Panel B reports the average cross-sectional Spearman (rank-order) correlation
of monthly cost-per-dollar-volume proxies compared to monthly lambda. In the earlier period, monthly
VoV() ties with monthly Gibbs Impact for the highest Spearman correlation and is statistically
indistinguishable from the monthly Close-High-Low Impact’s Spearman correlation. In the later period,
monthly VoV() has the highest Spearman correlation and statistically distinguishable from the other
proxies, but the size of the difference is modest.
10 We test if two average cross-sectional correlations on the same row are statistically distinguishable by t-tests on the time-series of correlations in the spirit of Fama and MacBeth. Specifically, we calculate the cross-sectional correlation of each proxy for each month and then regress the correlations of one proxy on the correlations of another proxy. We assume that the time-series correlations of each proxy are i.i.d. over time, and test if the regression intercept is zero and the slope is one. Standard errors are adjusted for autocorrelation with a Newey-West correction using four lags. 11 We test if correlations are significantly different from zero using the following t-test 2 1⁄ , where r is the correlation coefficient and n is the degrees of freedom. See Swinscow (1997), chapter 11.
14
Panel C reports the portfolio time-series correlation of the first differences of monthly cost-per-
dollar-volume proxies compared to the first differences of monthly lambda. That is, we compute the
equally-weighted portfolio of each monthly cost-per-dollar volume proxy across all stocks in each month
and compute the equally-weighted portfolio of monthly lambda across all stocks in each month. Then we
compute the correlation of the proxy portfolio time-series first differences and the lambda portfolio time-
series first differences. This performance criteria may be an especially useful as a measure of fitness for
liquidity risk research. Monthly VoV() has the highest portfolio time-series correlation in both periods.
The monthly Close-High-Low, Closing Percent Quoted Spread, and Amihud correlations are statistically
indistinguishable from monthly VoV() in the later period, where we test whether portfolio time-series
correlations are statistically different from each other using a Fisher’s Z-test. There is economically large
difference in performance between monthly VoV() and monthly Amihud proxy in both periods, but the
difference is especially large in the later period.
Panel D reports the individual stock time-series correlation of the first differences of monthly cost-
per-dollar-volume proxies compared to the first differences of monthly lambda. We find that monthly
VoV() has the highest individual stock time-series correlation of the first differences in both periods and
is statistically significantly higher than the correlations of any other monthly proxy in both rows. There is
economically large difference in performance between monthly VoV() and monthly Amihud in both
periods.
Panel E reports the simple average of the correlations on the eight contests above (i.e., four panels
above X two time-periods per panel). This simple summary statistic allows us to get an overview of each
proxy’s correlation performance. We find that monthly VoV() has an average correlation of 0.562, which
compares to 0.464 for monthly Amihud. Thus, Monthly VoV() provides a large and economically
meaningful gain in correlation performance over monthly Amihud.
Figure 3 plots the global values of monthly VoV(), monthly Amihud, and monthly lambda by
quantiles (i.e., cutoffs of equal probability) of monthly lambda. Note that the y-axis is on a log-scale. Both
15
monthly VoV() and monthly Amihud get the rank ordering correct. That is, they both yield higher values
when monthly lambda is higher and lower values when it is lower. Monthly VoV() is close the actual
values of monthly lambda for high quantiles, but up to two orders of magnitude off for low quantiles.
Conversely, monthly Amihud is close to monthly lambda for low quantiles, but up to 1.5 orders of
magnitude off for high quantiles. Interestingly, monthly Amihud is much more noisy than monthly VoV().
Considering the illustration above that Amihud is sensitive to low-volume outliers, one would expect this
kind of noisy disturbances in Amihud.
Returning to Table 3, Panel F reports the average root mean squared error (RMSE) of monthly cost-
per-dollar-volume proxies compared monthly lambda and then divided by the median of monthly lambda.
The RMSE is calculated every month for given exchange and then averaged over all sample months. The
average RMSE is useful for determining if a particular liquidity proxy does a good job of capturing the
level of a liquidity benchmark, not just that it is correlated with a liquidity benchmark. Capturing the level
is important for many research purposes, such as computing the returns to a trading strategy net of a
correctly-scale proxy for transaction costs. In the case of RMSE comparisons, the solid box denoting the
best performance is placed around the lowest average RMSE in the row and the dashed box is placed around
the average RMSEs that are statistically indistinguishable from the lowest average RMSE in the row. We
test if average RMSEs are statistically indistinguishable from each other using paired t-test. Boldfaced
RMSE indicated that the ability of a proxy to predict the benchmark is statistically greater than zero at the
5% level.12
We find that monthly VoV() has the lowest ratio of average RMSE over median of monthly lambda
at 74 in the earlier period and 93 in the later period. The rest of the proxies have even larger ratios. This
means that the average error is nearly two orders of magnitude larger than the median of lambda itself.
Thus, none of the monthly liquidity proxies does a good job of capturing the level of monthly lambda.
12 Throughout the paper, we use the U-statistic of Theil (1996) to test if RMSEs are statistically significant. Under Theil’s convention if U2 = 1, then the proxy has no ability to predict the benchmark. We test if U2 is significantly less than 1 based on a F distribution, where the number of degrees of freedom is the sample size.
16
To summarize this sub-section, the global performance of monthly VoV() strongly dominates
monthly Amihud and the rest of the monthly cost-per-dollar-volume liquidity proxies. In Panels A-D, it has
the highest correlation with monthly lambda in all eight contests. These contests span cross-sectional and
time-series; regular and Spearman; and portfolio and individual stock correlations of first differences.
Monthly VoV()’s correlation is statistically significantly higher than the correlation of monthly Amihud
in all eight cases. Its average correlation across all eight contests provides a large and economically
meaningful gain in correlation performance. Further, monthly VoV() is much more robust (i.e. less noisy)
than monthly Amihud. However, none of the monthly proxies captures the level of monthly lambda.
6.2 Daily Cost-Per-Dollar-Volume Proxies
Table 4 shows the global performance of daily cost-per-dollar-volume proxies compared to daily
lambda. It reports the same five performance dimensions (average cross-sectional correlation, average
Spearman’s cross-sectional correlation, portfolio time-series correlation of first differences, individual
stock time-series correlation of first differences, and average RMSE) and provides a summary statistic
across all of the correlation results.
We find that daily VoV() has the highest correlation with daily lambda in all eight contests (panels
A-D X 2 time-periods per panel). In seven of the eight contests, its correlation is statistically significantly
higher than the correlations of all other cost-per-dollar-volume proxies and one case it is insignificantly
different than daily High-Low Impact. The gap between the performance of daily VoV() and the
performance of the second-place finisher on the row is moderate-to-large in all cases. Panel E reports that
daily VoV()’s average correlation across all eight contests is 0.534 compared to 0.444 for daily Amihud,
which is a large and economically meaningful gain in correlation performance.
Panel F reports the average RMSE divided by the median of daily lambda. The average error is
one-to-two orders of magnitude larger than the median of daily lambda. Thus, none of the daily liquidity
proxies does a good job of capturing the level of daily lambda.
6.3. Monthly Percent-Cost Proxies
17
Table 5 shows the global performance of monthly percent-cost proxies compared to the monthly
percent-cost benchmark, percent effective spread. It reports the same five performance dimensions and
provides a summary statistic across all of the correlation results.
We find that monthly Closing Percent Quoted Spread has the highest correlation in six of the eight
contests (specifically in both time-periods of Panels A, B, and D) and its correlation is statistically
significantly higher than that of any other monthly percent-cost proxy in those six cases. Monthly
VoV(%Spread) has the second highest correlation in those six cases and there is a large gap to the third
highest correlation in those six cases. Monthly VoV(%Spread) and monthly Closing Percent Quoted Spread
have the highest portfolio time-series correlation or are indistinguishable from the highest correlation in
both time-periods in Panel C. These correlation results are summarized in Panel E, which shows monthly
Closing Percent Quoted Spread with the highest average correlation at 0.784, monthly VoV(%Spread) is a
close second at 0.757, and there is a large gap to the third highest at 0.615.
Figure 4 shows the global average cross-sectional correlations between monthly percent-cost
proxies and monthly percent effective spread over time. We see that monthly Closing Percent Quoted
Spread has the highest cross-section correlation in all months (except for two) over the 19-year sample
period. Monthly VoV(%Spread) is a close second in all months (except for two) over 19 years. Then there
is a large drop in performance to the rest of the monthly percent-cost proxies.
Returning to Table 5, Panel F reports the average RMSEs. Monthly Close-High-Low and monthly
High-Low have the lowest average RMSE for the earlier and later periods, respectively. All of the average
RMSEs are statistically significant for all of the monthly percent-cost proxies, which in marked contrast to
analogous results for monthly cost-per-dollar-volume proxies. Essentially, all of the monthly percent-cost
proxies get the level right. The gap in performance between the VoV(%Spread) average RMSEs and the
lowest average RMSEs in a row is relatively small.
Figure 5 plots the global values of monthly VoV(%Spread), monthly Closing Percent Quoted
Spread, and monthly percent effective spread by quantiles (i.e., cutoffs of equal probability) of monthly
percent effective spread. Note that the y-axis is now on a linear-scale. Both monthly VoV(%Spread) and
18
monthly Closing Percent Quoted Spread are very close to monthly percent effective spread across-the-
board. Monthly VoV(%Spread) does slight better with high quantiles and monthly Closing Percent Quoted
Spread does slightly better with low quantiles, but both get the level right.
To summarize Table 5, monthly VoV(%Spread) performs very well and comes in a close second.
However, monthly Closing Percent Quoted Spread is the best monthly percent-cost proxy for global
research. This confirms the prior monthly recommendation of FHT, but does so over an expanded set of
test proxies.
6.4 Daily Percent-Cost Proxies
Table 6 shows the global performance of daily percent-cost proxies compared to daily percent-cost
benchmark, percent effective spread. It reports the same five performance dimensions and provides a
summary statistic across all of the correlation results.
We find that daily Closing Percent Closing Spread has the highest correlation with daily percent
effective spread in four of the eight contests in Panels A-D and is statistically significantly higher than any
other daily percent-cost proxy in those four cases. VoV(%Spread) has the highest correlation in the other
four contests and is statistically significantly higher than the other daily percent-cost proxy in those four
cases. Panel E reports that daily Closing Percent Quoted Spread’s average correlation across all eight
contests is 0.673 compared to 0.652 for daily VoV(%Spread), which is a small gap in performance. The
third place proxy finishes far behind with an average correlation across all eight contests of 0.302.
Panel F reports that daily VoV(%Spread) and daily Closing Percent Quoted Spread have the lowest
average RMSE in the earlier and later periods, respectively. All of the daily percent-cost proxies do a good
job of capturing the level, except for the case of VoV(%Spread) in the later period when it is statistically
insignificant.
To summarize Table 6, daily VoV(%Spread) performs very well, but comes in a close second. Daily
Closing Percent Quoted Spread is the best daily percent-cost proxy for global research, which confirms the
prior daily recommendation of FHT, but does so over an expanded set of test proxies.
6.5 Monthly Robustness Checks
19
In this subsection we check the robustness of our monthly results. Table 7 reports the global
performance of monthly cost-per-dollar-volume proxies with monthly lambda by size and volatility
quintiles and after controlling for size and volatility. Panels A and B report average cross-sectional
correlations broken out by size quintiles and by volatility quintiles, respectively. We find that monthly
VoV() has the highest average cross-sectional correlations for all five size quintiles and for all five
volatility quintiles. In all cases, it is significantly higher than any other monthly cost-per-dollar-volume
proxy and there is an economically meaningful gain in correlation performance.
Panels C and D report individual stock time-series correlations of first differences by size quintiles
and by volatility quintiles, respectively. We find that monthly VoV() has the highest individual stock time-
series correlations of first differences for all five size quintiles and for all five volatility quintiles. In all
cases, it is significantly higher than any other monthly cost-per-dollar-volume proxy and there is an
economically meaningful gain in correlation performance.
Panel E reports pooled partial correlations after controlling for log size and volatility. The
correlations are computed over the pooled sample (i.e., across all stocks and across all months). The partial
correlations are computed by regressing each monthly cost-per-dollar-volume proxy on monthly log size
and volatility. Next, monthly lambda is regressed on monthly log size and volatility. Finally, the residuals
from both regressions are collected and correlated with each other to yield the partial correlations. We find
that monthly VoV() has the highest pooled partial correlation and it statistically significantly higher than
any other monthly cost-per-dollar-volume proxy. Panel F reports pooled partial Spearman’s correlations
after controlling for log size and volatility. We find that monthly Gibbs Impact has the highest pooled
Spearman’s partial correlation and that monthly VoV() is statistically indistinguishable from it.
In summary, the dominant performance of monthly VoV() is robust by size and volatility
subsamples. Further, it is robust net of controlling for size and volatility.
Table 8 reports the global performance of monthly percent-cost proxies with monthly percent
effective spread by size and volatility quintiles and after controlling for size and volatility. Panels A and B
20
report average cross-sectional correlations by size quintiles and by volatility quintiles, respectively. We
find that monthly Closing Percent Quoted Spread has the highest average cross-sectional correlations for
all five size quintiles and for all five volatility quintiles. We also find that monthly VoV(%Spread) has the
second highest correlation in all cases and there is a large gap to the third highest correlation.
Panels C and D report individual stock time-series correlations of first differences by size quintiles
and by volatility quintiles, respectively. We find that monthly Closing Percent Quoted Spread has the
highest individual stock time-series correlations of first differences for all five size quintiles and for all five
volatility quintiles. Again we also find that monthly VoV(%Spread) has the second highest correlation in
all cases and there is a large gap to the third highest correlation.
Panel E reports pooled partial correlations after controlling for log size and volatility. We find that
monthly Closing Percent Quoted Spread has the highest pooled partial correlation and it statistically
significantly higher than any other monthly cost-per-dollar-volume proxy. Panel F reports pooled partial
Spearman’s correlations after controlling for log size and volatility. We find that monthly Closing Percent
Quoted Spread has the highest pooled Spearman’s partial correlation. We find that monthly VoV(%Spread)
has the second highest pooled partial correlation in both Panel E and F and there is a large gap to the third
highest correlation.
In summary, the first-place performance of monthly Closing Percent Quoted Spread and the
second-place performance of monthly VoV(%Spread) are robust by size and volatility subsamples. Further,
they are robust net of controlling for size and volatility.
6.6 Monthly General Volatility Over Volume Class Proxies
Now we turn to the general VoV class as given by the equation b ca V . The new volume over
volatility proxies that we have tested so far have fixed parameters for a, b, and c. Specifically, the
parameters for VoV(%Spread) are 8 , 2 3,a CPI b and 1 3c and the parameters for VoV(λ) are
1, 1,a b and 1 2c . Now we ask, how much additional global performance do we obtain if we
21
optimally select the a, b, and c parameters from the set of non-negative real numbers so as to maximize
pooled correlation?
Table 9 provides the answer. It shows the global performance of monthly general VoV class proxies
under the parameters that maximize pooled correlation and top monthly liquidity proxies compared to
monthly liquidity benchmarks. Performance is defined as the pooled correlation (i.e., combined time-series
and cross-sectional) over all global stock-months in the earlier or later periods.
Panel A is for monthly cost-per-dollar-volume proxies compared to monthly lambda. We see the
optimal values of a, b, c, and ρ for each time period. In the earlier period, the general VoV class pooled
correlation is 0.780 versus 0.777 for VoV(λ). In the later period, the general VoV class yields 0.425 versus
0.405 for VoV(λ). In both cases, gain in performance is very small and comes with the risk of overfitting
the data in-sample.
Panel B is for monthly percent-cost proxies compared to monthly percent effective spread. We see
the optimal values of a, b, c, and ρ for each time period. In the earlier period, the general VoV class yields
0.836 versus 0.817 for VoV(%Spread) and in the later period, the general VoV class yields 0.849 versus
0.826 for VoV(%Spread). Again, both periods yield very small gains in performance and come with the risk
of overfitting.
In summary, there is very little to gain by going with the general VoV class and the overfitted
performance in-sample may not hold up out-of-sample. Therefore, our recommendation is to stay with the
VoV(%Spread) and VoV(λ) proxies which are grounded by the transaction cost invariance hypothesis, rather
than the general VoV class of proxies.
7. US Performance Results
7.1 Cost-Per-Dollar-Volume Proxies
Table 10 shows the US performance of monthly (daily) cost-per-dollar-volume proxies compared
to the monthly (daily) cost-per-dollar-volume benchmark, lambda. It reports the same five performance
dimensions and provides a summary statistic across all monthly (daily) correlation results.
22
We find that monthly (daily) VoV(λ) has the highest correlation in two (four) of the four contests.
Monthly FHT Impact has the highest correlation in one case and monthly Close-High-Low Impact has the
highest correlation in another case, but monthly VoV(λ) is insignificantly different in one of these cases.
Panel E summarizes that monthly VoV(λ) has an average correlation of 0.584, which is a small gain over
the 0.569 correlation of the second place finisher, but is a large gain over the 0.433 correlation of monthly
Amihud. Daily VoV(λ) has an average correlation of 0.546, which is a large gain over 0.401 of the second
place finisher.
Panel F reports that the average error is more than an order of magnitude larger than the median of
monthly (daily) lambda. Thus, none of the monthly (daily) liquidity proxies does a good job of capturing
the level of monthly (daily) lambda.
In summary, US performance of monthly and daily VoV() mirrors their strong global performance.
Specifically, the US performance of monthly VoV() strongly dominates monthly Amihud and the US
performance of daily VoV() strongly dominates all daily cost-per-dollar-volume proxies. None of the
monthly (daily) cost-per-dollar-volume proxies capture the level of monthly (daily) lambda.
7.2 Percent-Cost Proxies
There is an important difference in the structure of US stock data versus global stock data. US stock
trade data (including closing trade price, daily volume, daily high, and daily low) is available from 1925 to
the present, but US stock quote data (including closing bid and closing ask) is only continuously available
across the whole US market from 1993 to the present.13 In other words, there is a period of 68 years in
which US trade data is available to construct percent-cost proxies, but US quote data is partially available
at best. This essentially rules out the use of Closing Percent Quoted Spread for studies that cover pre-1993
years. By contrast, a large sample of global stock quote data is available for most exchanges from 1996 to
the present.
13 Specifically, CRSP provides closing bid and ask prices for NYSE/AMEX stocks form 1926-1941 and 1993-present. It provides the same for NASDAQ Global Market and Global Select Market stocks form 1982-present and for NASDAQ Capital Market stocks from 1992-present.
23
This corresponds roughly to the time period that global stock trade data is available across a wide
cross-section of exchanges worldwide. Therefore, Closing Percent Quoted Spread works as a percent-cost
proxy for global samples and it is candidate proxy for US studies that cover 1993 to the present, but we
need to find alternatives to it for US studies that cover pre-1993 years.
Table 11 shows the US performance of monthly (daily) percent-cost proxies compared to the
monthly (daily) percent-cost benchmark, percent effective spread. It reports the same five performance
dimensions and provides a summary statistic across all monthly (daily) correlation results.
We find that monthly (daily) VoV(%Spread) has the highest correlation in two (four) of the four
contests and is statistically significantly higher than any other monthly (daily) percent-cost proxy in two
(three) of those cases. Closing Percent Quoted Spread has the highest correlation in the remaining two
(zero) contests. Panel E shows that monthly Closing Percent Quoted Spread has an average correlation of
0.719, which is a tiny gain over the 0.705 correlation of monthly VoV(%Spread) and which in turn in a
large gain over the 0.593 correlation of the third place finisher. Daily VoV(%Spread) has an average
correlation of 0.584, which is a moderate gain over the 0.516 correlation of daily Closing Percent Quoted
Spread and which is a large gain over the 0.206 correlation of the third place finisher. Panel F reports that
monthly (daily) VoV(%Spread) has the lowest average RMSE and is statistically significantly lower than
any other monthly (daily) percent-cost proxy.
In summary, for US studies that cover 1993-present, monthly (daily) VoV(%Spread) and monthly
(daily) Closing Percent Quoted Spread are very closely matched. Both perform well and either one could
be used with confidence. For US studies that cover pre-1993 years, monthly (daily) VoV(%Spread) strongly
dominates any other monthly (daily) percent-cost proxy. Monthly (daily) VoV(%Spread) will provide large
performance gains over the alternatives.
8. Do VoV Liquidity Proxies Affect Research Inferences? The Case Of The Quant Meltdown
Do our new VoV liquidity proxies matter? That is, do they yield different research inferences than
the best previous liquidity proxies from the prior literature? To find out, we perform a case study that
24
compares daily VoV() and daily Amihud when analyzing the pattern of illiquidity before, during, and after
the August 2007 “Quant Meltdown” event.
Khandani and Lo (2007, 2011) note that over the four-day period of August 6-9, 2007 some of the
most successful equity hedge funds in history reported record losses.14 Specifically, these loss
announcements were made by quantitatively-managed, market-neutral hedge funds, giving rise to the
popular term for the event: the “Quant Meltdown.” Investors reacted to this news by withdrawing large
amounts of capital from these funds, which caused them to liquidate large amounts of their invested
portfolios. These forced liquidations pushed down the price of the securities being sold, which in turn
imposed loses on other similarly-invested hedge funds. Which in turn lead investors to pull money out of
those hedge funds as well, which forced them to liquidate some of their holdings. Khandani and Lo propose
the “Unwind Hypothesis” to explain this sequence of events around the Quant Meltdown announcements.
They find that some of the downward market impacts happened before the announcements, because some
of the quantitatively-managed hedge funds had quietly began to unwind their positions in late July and early
August as they realized the magnitude of the losses that they were facing.
For their empirical analysis, Khandani and Lo use the high-frequency Trade and Quote (TAQ)
dataset. They use this high-frequency data to compute a cost-per-dollar-volume “price impact” measure for
the S&P 1500 stock portfolio. They find that “the pattern of price impact displayed in Fig. 9 documents a
substantial drop in liquidity in the days leading up to August 6, 2007.” They further note that after the quant
meltdown announcements, the resulting large-scale liquidations caused further price impacts as the “market
makers burned by the turn of events in the week of August 6, reduced their market making capital in the
14 Khandani and Lo provide the following examples of major loss announcements. The Wall Street Journal reported on August 10, 2007, that ‘‘After the close of trading, Renaissance Technologies Corp., a hedge-fund company with one of the best records in recent years, told investors that a key fund has lost 8.7% so far in August and is down 7.4% in 2007. Another big fund company, Highbridge Capital Management, told investors its Highbridge Statistical Opportunities Fund was down 18% as of the 8th of the month, and was down 16% for the year. The $1.8 billion publicly traded Highbridge Statistical Market Neutral Fund was down 5.2% for the month as of Wednesday … Tykhe Capital, LLC – a New York- based quantitative, or computer-driven, hedge-fund firm that manages about $1.8 billion – has suffered losses of about 20% in its largest hedge fund so far this month…’’ (see Zuckerman, Hagerty, and Gauthier-Villars, 2007), and on August14, The Wall Street Journal reported that the Goldman Sachs Global Equity Opportunities Fund ‘‘…lost more than 30% of its value last week…’’ (Sender, Kelly, and Zuckerman, 2007).
25
following days and in turn caused the price impacts to rise substantially starting on August 10 and remain
high for the following week.” They conclude that their evidence supports the Unwind Hypothesis.
We examine the same three periods that were before, during, and after the quant meltdown
announcements. Specifically, we follow the exact dates mentioned by Khandani and Lo (see above) as
when they found higher price impact. The three periods are: before (July 30 – August 3), during (August 6
– August 9), and after (August 10 – August 17).
We analyze the S&P 1500 stocks. For each stock, we compute their cost-per-dollar-volume
illiquidity using both the daily VoV() and daily Amihud proxies. Following Khandani and Lo, we compute
the portfolio’s overall illiquidity on an equally-weighted basis.
Figure 6 shows the percent change in S&P 1500 illiquidity from July 2 to August 31, 2007 relative
to the average value over the base period. The base period are the two weeks of July 2 – July 13, 2007,
which is well in advance of the before period. The solid red circles are based on daily VoV() and open
blue squares are based on daily Amihud. The black vertical lines demark the before, during, and after
periods. First, it is immediately clear that daily Amihud is much more volatile than daily VoV(). For
example, during the base period the standard deviation of daily Amihud is 5.7 times larger than the standard
deviation of daily VoV(). Secondly, starting from the base period daily VoV() rises in the before period,
rises again in the during period, and rises to a peak in the after period. By contrast, daily Amihud declines
slightly in the before period, continues to be negative in the during period, and rises moderately in the after
period. That is, daily VoV() matches the S&P 1500 illiquidity pattern described by Khandani and Lo,
whereas daily Amihud does not match.
Table 12 reports the percent change in S&P 1500 illiquidity before, during, and after the quant
meltdown relative to the average value over the base period. Compared to the base period, VoV() rises
52.0% in the before period, rises again to 88.4% in the during period, and rises to a peak 118.9% in the
after period. All three values are significant at either the 5% or 1% levels. By contrast, daily Amihud
declines slightly to -5.8% in the before period, continues to be negative at -2.1% in the during period, and
26
rises moderately to 24.7% in the after period. None of the three values are significant. The insignificance
in the after period is driven both by the lower level of daily Amihud and by the higher volatility of the
Amihud standard error.
In summary, the two proxies yield different research inferences. The daily VoV() proxy yields the
inference that S&P 500 illiquidity is significantly higher than the base in all three periods. By contrast, the
daily Amihud proxy yields the inference that S&P 500 liquidity in not significantly different than the base
in all three periods. We conclude that our new VoV liquidity proxies do matter.
9. Conclusion
We examine a general class of volatility over volume liquidity proxies as computed from low
frequency (daily) data. We start from the Kyle and Obizhaeva (2016) hypothesis of transaction cost
invariance to identify a new volatility over volume liquidity proxy “VoV(%Spread)” for percent spread cost
and a new volatility over volume liquidity proxy “VoV(λ)” for the slope of the transaction cost function “λ”.
We test the monthly and daily versions of these new and existing liquidity proxies against liquidity
benchmarks as estimated from high frequency (intraday) data on both a global and US basis. We find that
both the monthly and daily versions of VoV(λ) dominate the equivalent versions of Amihud and other cost-
per-dollar-volume proxies on both a global and US basis. We also find that both the monthly and daily
versions of VoV(%Spread) dominate the equivalent versions of other percent-cost proxies for US studies
that cover pre-1993 years. In a case study, we find that our new VoV liquidity proxies yield different
research inferences than the best previous liquidity proxies from the prior literature. The success of our
invariance-based liquidity proxies across exchanges and over time supports the prediction of Kyle and
Obizhaeva of a specific functional form for transaction costs across exchanges and over time.
References
Abdi, G. and A. Ranaldo, 2017. A simple estimation of bid-ask spreads from daily close, high, and low
prices. Forthcoming in the Review of Financial Studies.
Amihud, Y., 2002. Illiquidity and stock returns: Cross section and time-series effects. Journal of Financial
Markets 5, 31-56.
27
Asness, C., Moskowitz, T., Pedersen, L., 2013. Value and momentum everywhere. Journal of Finance 68,
929-985.
Attig, N., Gadhoum, Y., Lang, L., 2003. Bid-Ask Spread, Asymmetric Information and Ultimate
Ownership. University of Quebec working paper.
Bailey, W., Karolyi, A., Salva, C., 2006. The economic consequences of increased disclosure: evidence
from international cross-listings. Journal of Financial Economics 81, 175-213.
Beber, A., Pagano, M., 2013. Short-Selling Bans Around the World: Evidence from the 2007-09 Crisis.
Journal of Finance 68, 343-381.
Bekaert, G., Harvey, C., Lundblad, C., 2007. Liquidity and expected returns: lessons from
emerging markets. Review of Financial Studies 20, 1783-1831.
Bekaert, G., Harvey, C., Lundblad, C., Siegel, S., 2014. Political risk spreads. Journal of International
Business Studies 45, 471-493.
Bhattacharya, U., Daouk, H., Jorgenson, B., Kehr, C., 2000. When an event is not an event: the curious
case of an emerging market. Journal of Financial Economics 55, 69-101.
Boubakri, N., Saad, M., and Samet, A., 2017. Commonality in Liquidity: The Culture Channel.
Unpublished working paper. American University of Sharjah.
Chan, J. Jain,R., Xia,Y., 2008. Market segmentation, liquidity spillover, and closed-end country fund
discounts. Journal of Financial Markets 11, 377-399.
Chung, K., Zhang, H., 2014. A simple approximation of intraday spreads using daily data. Journal of
Financial Markets 17, 94-120.
Clark, A., 2011, Revamping Liquidity Measures: Improving Investibility in Emerging and Frontier Market
Indices and their Related ETFs. Journal of Index Investing 2, 37-43.
Corwin, S., Schultz, P. 2012. A Simple Way to Estimate Bid-Ask Spreads from Daily High and Low Prices.
Journal of Finance, 67, 719-759.
DeNicolo, G., Ivaschenko, I., 2009. Global liquidity, risk premiums, and growth opportunities.
Unpublished working paper. International Monetary Fund.
28
Erten, I., Okay, N., 2012. Deciphering liquidity risk on the Istanbul Stock Exchange. Bogazici University
working paper.
Fama, E., MacBeth, J., 1973. Risk, return, and equilibrium: empirical tests. Journal of Political Economy
81, 607–636.
Fong, K., Holden, C., Trzcinka, C., 2017. What are the best liquidity proxies for global research?
Forthcoming in the Review of Finance.
Gersl, A. Komarkov, Z., 2009. Liquidity risk and bank’s bidding behavior: Evidence from the global
financial crisis. Czech Journal of Economics and Finance 59, 577-592.
Gomez-Puig, M., 2006. Size matters for liquidity: Evidence from EMU sovereign yields spreads.
Economic Letters 90, 156-162.
Goyenko, R., Holden, C., Trzcinka C., 2009. Do liquidity measures measure liquidity? Journal of Financial
Economics 92, 153-181.
Griffin, J., Hirschey, N., Kelly, P., 2011. How important is the financial media in global markets? Review
of Financial Studies 24, 3941-3992.
Griffin, J., Kelly, P., Nardari, F., 2010. Do Market Efficiency Measures Yield Correct Inferences? A
Comparison of Developed and Emerging Markets. Review of Financial Studies 23, 3225-3277.
Hanselaar, R., Stulz, R., and van Dijk, M., 2016. Do firms issue more equity when markets become more
liquid? Unpublished working paper. The Ohio State University.
Hasbrouck, J., 2004. Liquidity in the futures pits: Inferring market dynamics from incomplete data. Journal
of Financial and Quantitative Analysis 39, 305–326.
Hasbrouck, J., 2009. Trading costs and returns for US equities: the evidence from daily data. Journal of
Finance 64, 1445–1477.
Hearn, B. 2014. The political institutional and firm governance determinants of liquidity: Evidence from
North Africa and the Arab Spring, Journal of International Financial Markets, Institutions & Money
31, 127-158.
29
Hearn, B., Piesse, J., Strange, R., 2010. Market Liquidity and Stock Size Premia in Emerging
Financial Markets: The Implications for Foreign Investment. International Business Review 19,
489-501.
Henkel, S., 2008. Is global illiquidity contagious? contagion and cross-market commonality in liquidity.
Unpublished working paper. Indiana University.
Henkel, S., Jain, P., and Lundblad, C., 2008. Liquidity dynamics and stock market automation.
Unpublished working paper. Indiana University.
Holden, C., 2009. New low-frequency liquidity measures. Journal of Financial Markets 12, 778-813.
Ince, O., Porter, R., 2006. Individual Equity Return Data From Thomson Datastream: Handle With
Care! Journal of Financial Research 29, 463-479.
Jain, P., 2005. Financial market design and the equity premium: Electronic versus floor trading. Journal of
Finance 60, 2955-2985.
Karolyi, A., Lee, K. and Van Dijk, M. 2012. Understanding commonality in liquidity around the world.
Journal of Financial Economics 105, 82-112.
Karnaukh, Nina, Angelo Renaldo and Paul Soderlind, 2015. Understanding FX Liquidity, Review of
Financial Studies, 28, 3073-3108.
Khandani, A. and A. Lo, 2007. What happened to the quants in August 2007? Journal of Investment
Management 5, 5-54.
Khandani, A. and A. Lo, 2011. What happened to the quants in August 2007? Evidence from factors and
transactions data. Journal of Financial Markets 14, 1-46.
Kyle, A. and A. Obizhaeva, 2016. Market Microstructure Invariance: Empirical Hypotheses. Econometrica
84, 1345-1404.
LaFond, R., Lang, M., and Skaife, H., 2007. Earnings smoothing, governance and liquidity:
international evidence. Unpublished working paper. Massachusetts Institute of Technology.
Lang, M., Lins, K., and Maffett, M., 2012. Transparency, Liquidity, and Valuation: International Evidence
on When Transparency Matters Most. Journal of Accounting Research 50, 729-774.
30
Lee, C., Ready, M., 1991. Inferring trade direction from intraday data. Journal of Finance 46, 733–746.
Lee, H., Tseng, Y., Yang, C., 2014. Commonality in liquidity, liquidity distribution, and financial crisis:
Evidence from country ETFs. Pacific-Basin Finance Journal 29, 35-58.
Lee, K., 2011. The world price of liquidity risk. Journal of Financial Economics 99, 136-191.
Lee, K., Sapriza, H., and Wu, Y., 2016. Sovereign debt ratings and stock liquidity around the world, Journal
of Banking and Finance 73, 99-112.
Lesmond, D., 2005. Liquidity of emerging markets. Journal of Financial Economics 77, 411-452.
Lesmond, D., Ogden J., Trzcinka C., 1999. A new estimate of transaction costs. Review of
Financial Studies 12, 1113-1141.
Levine, R., Schmukler, S., 2006. Internationalization and stock market liquidity. Review of Finance 10,
153-187.
Liang, S., Wei, K.C., 2012. Liquidity risk and stock returns around the world. Journal of Banking and
Finance 36, 3274-3288.
Marshall, B., Nguyen, N., and Visaltanachoti, N., 2013. Liquidity Measurement in Frontier Markets,
Journal of International Financial Markets, Institutions and Money, 27, 1-12.
Massa, M., Mataigne, V., Vermaelen, T., and Xu, M., 2017. Choices in Equity Finance: A Global
Perspective, Unpublished working paper. Centre for Economic Policy Research.
Noss, J., Pedace, L., Tobek, O., Linton, O. and Crowley-Reidy, L., 2017. The October 2016 Sterling Flash
Episode: When Liquidity Disappeared from One of the World's Most Liquid Markets. Bank of
England Working Paper No. 687
Orlov, V., 2016, Currency momentum, carry trade, and market illiquidity. Journal of Banking and Finance
67, 1-11.
Parkinson, M., 1980. The extreme value method for estimating the variance of the rate of return, Journal of
Business 53, 61-65.
Roll, R., 1984. A simple implicit measure of the effective bid-ask spread in an efficient market. Journal of
Finance 39, 1127-1139.
31
Sender, H., Kelly, K., and Zuckerman, G., 2007. Goldman wagers on cash infusion to show resolve. The
Wall Street Journal (Eastern edition) August 14, A.1.
Stahel, C., 2005. Is there a global liquidity factor? Unpublished working paper. George Mason
University.
Swinscow, T., 1997. Statistics at Square One, 9th ed. BMJ Publishing Group, London.
Zuckerman, G., Hagerty, J., and Gauthier-Villars, D., 2007. Impact of mortgage crisis spreads; Dow
tumbles 2.8% as fallout intensifies; moves by central banks. The Wall Street Journal (Eastern
edition) August 10, A.1.
32
Appendix A: Liquidity Measure Derivations
VoV(%Spread) Derivation
Transaction cost invariance supports alternative functional forms. One of those cases is a constant
percent quoted spread cost as given by their equation (19) for the percent quoted spread
1
30 ,jt
jt jtjt
sW
P
(15)
where jts is the dollar quoted spread of asset j on date t , jtP is the price of asset j on date t , 0 is a
constant, and jt is bet volatility, which is the standard deviation of returns generated by bet price impact.
jtW is bet activity the of asset j on date t , which is defined to be the product of price ,jtP bet volume
,jtV and bet volatility jt , as given by
.jt jt jt jtW P V (16)
Bet volume jtV is assumed to be related to trading volume jtV as follows
2.jt jt
jt
V V
(17)
where jt is a volume multiplier. Bet volatility jt is assumed to be related to returns volatility jt as
follows
jt jt jt (18)
where jt is the scale factor.
Substituting (16), (17), and (18) into (15), we obtain
1
3
0
1 2 1 2 13 3 3 3 3
0
2
2 ,
jtjt jt jt jt jt jt
jt jt
jt jt jt jt
sP V
P
V
(19)
33
where the second equality follows by just rearranging terms. Equation (19) is our VoV(%Spread) proxy and
can be written in volatility over volume form as
%
bjt
cjt
VoV Spreads a
P V
, (20)
where 1 2 1
3 3 302 ,jt jta
2
,3
b and 1
3c .
VoV(λ) Derivation
Alternatively, transaction cost invariance supports a linear expression for price impact jt jtC Q of
asset j on date t . Substituting 0 0 into Kyle and Obizhaeva equation (18), we obtain
1
2
,jtjt jt I jt
jt
QC Q
V (21)
where I is a constant and jtQ is unsigned number of shares in a bet. Substituting equations (17) and (18)
into (21), we obtain
1
2
.2
jtjt jt I jt jt
jtjt
QC Q
V
(22)
Define jtD as the sign of the net order imbalance in asset j on date t . Multiplying both sides of equation
(22) by jtD and multiplying the RHS numerator and denominator by 1
2jtP , we obtain
1
21
1 122 2
.
2
I jt jtjt jt jt jt jt jt
jt jtjt
D C Q D P Q
P V
(23)
34
Equation (23) is essentially a regression of signed price impact returns on the signed square root of dollar
volume. The RHS coefficient of this regression, which is shown in curly brackets, is our VoV(λ) price
impact proxy. It can be written in the volatility over volume form as
b
cVoV
a
V
, (24)
where we can map our notation to the Kyle and Obizhaeva notation as follows jt and jt jtV P V and
where ,2
I jt jta
1,b and
1
2c .
Appendix B: Low-Frequency Proxies from the Existing Literature.
The Gibbs sampler is an estimation process with three steps for each iterative sweep. First, given the
sample of prices P from all days in the time interval, starting values for the buy/sell indicator ,Q a prior
for the effective half-spread c , and a prior for the variance of public information innovations 2u ,
estimate c using a Bayesian regression that is restricted to the positive domain. Second, given , Q,P the
prior for 2u , and the updated estimate of c , estimate the residuals and make a new draw of 2
u from an
inverted gamma distribution. Third, given ,P the updated estimate of c , and the new draw of 2u , make
new draws of and Q V . You run 1,000 sweeps of the sampler, discard the first 200 as burn-in, and then
take the mean of the c values in the remaining 800 sweeps as the final estimate of c .
1 1+2 ,
2
zFHT N
where is the return standard deviation, z is the frequency of zero returns, and
1N is the inverse function of the cumulative normal distribution.
2 1;
1
t
t
eHigh Low Average
e
where 2
,3 2 2 3 2 2
t t tt
t is the sum over two
days of the squared daily log(high/low), and t is the squared log(High/Low) where the High (Low)
value is over two days.
35
411
max ,0 ,N
t t t tN tCHL c c
where tc is the close log-price on day t and the mid-range
/ 2t t tl h is the average of the daily high and low log-prices on day t.
/ 2
t t
t t
Closing Ask Closing BidClosing Percent Quoted Spread Average
Closing Ask Closing Bid
Listed below are five cost-per-dollar-volume proxies for month i (or day i) based on the “Extended
Amihud” class of proxies as defined in Goyenko, Holden, and Trzcinka (2009), Section 5.2:
Gibbs Impacti = Gibbsi / (Average Daily US Dollar Value of Local Volume) i.
FHT Impacti = FHTi / (Average Daily US Dollar Value of Local Volume) i.
High-Low Impacti = High-Lowi / (Average Daily US Dollar Value of Local Volume) i.
Close-High-Low Impacti = Close-High-Lowi / (Average Daily US Dollar Value of Local Volume) i.
Closing Percent Quoted Spread Impacti = Closing Percent Quoted Spreadi / (Average Daily US Dollar
Value of Local Volume) i.
t
t
rAmihud Average
Dollar Volume
, where the average is computed over positive volume days
only and where tr is the stock return on day t and tDollar Volume is the US Dollar value of volume on
day t.
36
Figure 1. Monthly VoV(%Spread) adjusted for inflation vs. unadjusted compared to close-high-low in the
US over time.
Figure 2. Global average cross-sectional correlations between monthly cost-per-dollar-volume proxies
and monthly lambda over time.
37
Figure 3. The global values of monthly VoV(λ) and monthly Amihud compared to monthly lambda by the
quantile of monthly lambda.
Figure 4. Global average cross-sectional correlations between monthly percent-cost proxies and monthly
percent effective spread over time.
38
Figure 5. The global values of monthly VoV(%Spread) and monthly closing percent quoted spread
compared to monthly percent effective spread by quantile of monthly percent effective spread.
Figure 6. Using VoV() and Amihud to compute the percent change in S&P 1500 illiquidity before,
during, and after the August 6-9, 2007 “quant meltdown.”
‐100.0%
‐50.0%
0.0%
50.0%
100.0%
150.0%
200.0%
7/2/07 7/9/07 7/16/07 7/23/07 7/30/07 8/6/07 8/13/07 8/20/07 8/27/07 9/3/07
Percent Chan
ge in S&P 1500 Illiquidity
(Base Period: July 2 ‐July 13, 2007)
VoV(λ)Amihud
After = Aug 10 ‐ Aug 17
During = Aug 6 ‐ Aug 9
Before = July 30‐ Aug 3
39
Table 1.
Numerical Illustration
Day Return
Dollar
Volume
Amihud
ratio
1 0.050 $300,000 0.00000017
2 ‐0.030 $250,000 0.00000012
3 0.020 $200,000 0.00000010
4 ‐0.010 $210,000 0.00000005
5a 0.040 $5,000 0.00000800
5b 0.040 $250,000 0.00000016
Std Dev of
Return
Average
Volume
VoV(λ)
* 1,000,000
Amihud
* 1,000,000
Days 1‐4 & 5a 0.034 $193,000 76.52 1.69
Days 1‐4 & 5b 0.034 $242,000 68.33 0.12
Ratio 1.1 14.2
40
Table 2
Descriptive Statistics for the Global and US Samples
Percent‐
Cost
Benchmark
Gibbs FHT
High‐
Low
Close‐
High‐Low
Closing %
Quo Sprd
VoV
(%Spread)
% Effective
Spread
Stock‐
Months Stocks
Panel A: Mean of the Global Monthly Percent‐Cost Proxies and Benchmarks
1996‐2007 0.024 0.014 0.013 0.015 0.021 0.020 0.017 1,500,611 24,419
2008‐2014 0.021 0.011 0.013 0.015 0.018 0.019 0.015 85,897 1,807
Panel B: Mean of the Global Daily Percent‐Cost Proxies and Benchmarks
1996‐2007 0.011 0.014 0.017 0.015 0.015 28,042,581 23,308
2008‐2014 0.012 0.014 0.016 0.018 0.014 1,687,131 1,840
Panel C: Mean of the US Percent‐Cost Proxies and Benchmarks 1996‐2014
Monthly 0.028 0.010 0.017 0.019 0.023 0.016 0.020 60,537 962
Daily 0.013 0.015 0.013 0.011 0.011 793,057 467
Cost‐Per‐$‐
Volume
Benchmark
Gibbs
Impact
FHT
Impact
High‐
Low
Impact
Close‐
High‐Low
Impact
Closing %
Quo Sprd
Impact Amihud VoV() Lambda
Panel D: Median of the Global Monthly Cost‐Per‐Dollar‐Volume Proxies and Benchmarks
1996‐2007 0.060 0.015 0.028 0.036 0.032 0.099 47.253 26.013
2008‐2014 0.050 0.009 0.026 0.031 0.030 0.084 41.152 9.190
Panel E: Median of the Global Daily Cost‐Per‐Dollar‐Volume Proxies and Benchmarks
1996‐2007 0.007 0.001 0.007 0.016 22.966 19.310
2008‐2014 0.004 0.000 0.004 0.008 17.434 6.246
Panel F: Median of the US Cost‐Per‐Dollar‐Volume Proxies and Benchmarks 1996‐2014
Monthly 0.030 0.002 0.017 0.020 0.016 0.049 36.763 8.581
Daily 0.002 0.000 0.001 0.004 12.721 2.406
Percent‐Cost Proxies
Cost‐Per‐Dollar‐Volume Proxies
The monthly (daily) high‐frequency liquidity benchmarks, percent effective spread and lambda, are
calculated from every trade and corresponding BBO quote in the SIRCA Thomson Reuters Tick
History database for a sample stock‐month (stock‐day). The monthly (daily) percent‐cost proxies
and cost‐per‐dollar‐volume proxies are calculated from daily stock price data for a sample stock‐
month (stock‐day). The earlier (later) global sample spans 42 exchanges around the world from
1996‐2007 (2008‐2014). All stock‐months are required to have at least five positive‐volume days and
eleven non‐zero return days. The medians of all of the cost‐per‐dollar‐volume proxies and
benchmark have been multiplied by 1,000.
41
Table 3
Gibbs
Impact
FHT
Impact
High‐Low
Impact
Close‐
High‐Low
Impact
Closing %
Quo Sprd
Impact Amihud VoV()
1996‐2007 0.671 0.605 0.644 0.661 0.685 0.588 0.755
2008‐2014 0.424 0.377 0.426 0.417 0.406 0.404 0.471
1996‐2007 0.490 0.330 0.481 0.486 0.467 0.468 0.490
2008‐2014 0.538 0.389 0.533 0.532 0.522 0.525 0.551
1996‐2007 0.781 0.805 0.760 0.764 0.797 0.854 0.937
2008‐2014 0.232 0.240 0.272 0.293 0.265 0.305 0.510
1996‐2007 0.363 0.218 0.348 0.348 0.365 0.304 0.406
2008‐2014 0.322 0.175 0.321 0.319 0.319 0.262 0.375
Panel E: Average Correlation of the Eight Contests Above
Summary Stat. 0.478 0.392 0.473 0.478 0.478 0.464 0.562
1996‐2007 102 102 104 103 112 95 74
2008‐2014 103 104 104 104 102 100 93
The Global Performance of Monthly Cost‐Per‐Dollar‐Volume Proxies Compared to Monthly
Lambda
Panel D: Individual Stock Time‐Series Correlation of First Differences of Monthly Cost‐Per‐Dollar‐
Volume Proxies Compared to the Monthly Lambda
Panel F: Average Root Mean Squared Error (RMSE) of Monthly Cost‐Per‐Dollar‐Volume Proxies
Compared to Monthly Lambda / Median of Monthly Lambda
The high‐frequency cost‐per‐dollar‐volume benchmark, monthly lambda, is calculated from
every trade and corresponding BBO quote in the SIRCA Thomson Reuters Tick History database
for a sample stock‐month. The monthly cost‐per‐dollar‐volume proxies are calculated from daily
stock price data for a sample stock‐month. The earlier (later) sample spans 42 exchanges around
the world from 1996‐2007 (2008‐2014). All stock‐months are required to have at least five positive‐
volume days and eleven non‐zero return days. A solid box means the highest correlation or the
lowest average root mean squared error (RMSE) in the row. Dashed boxes mean correlations that
are statistically indistinguishable from the highest correlation or average RMSEs that are
statistically indistinguishable from the lowest average RMSE in the row at the 5% level. Bold‐
faced numbers are statistically different from zero or proxies have predictive power that is
significant at the 5% level.
Panel A: Average Cross‐Sectional Correlation of Monthly Cost‐Per‐Dollar‐Volume Proxies
Compared to Monthly Lambda
Panel B: Average Spearman's Cross‐Sectional Correlation of Monthly Cost‐Per‐Dollar‐Volume
Proxies Compared to Monthly Lambda
Panel C: Portfolio Time‐Series Correlation of First Differences of Monthly Cost‐Per‐Dollar‐
Volume Proxies Compared to Monthly Lambda
42
Table 4
High‐Low
Impact
Close‐
High‐Low
Impact
Closing %
Quo Sprd
Impact Amihud VoV()
1996‐2007 0.481 0.421 0.528 0.530 0.638
2008‐2014 0.347 0.304 0.364 0.372 0.429
1996‐2007 0.259 0.117 0.339 0.366 0.440
2008‐2014 0.360 0.165 0.467 0.430 0.541
1996‐2007 0.761 0.797 0.721 0.775 0.827
2008‐2014 0.834 0.811 0.809 0.826 0.851
1996‐2007 0.057 0.071 0.110 0.116 0.258
2008‐2014 0.089 0.062 0.136 0.133 0.291
Panel E: Average Correlation of the Eight Contests Above
Summary Stat. 0.399 0.344 0.434 0.444 0.534
1996‐2007 51 51 46 51 67
2008‐2014 113 113 108 113 103
Panel D: Individual Stock Time‐Series Correlation of First Differences of Daily Cost‐Per‐Dollar‐
Volume Proxies Compared to the Daily Lambda
Panel F: Average Root Mean Squared Error (RMSE) of Daily Cost‐Per‐Dollar‐Volume Proxies
Compared to Daily Lambda / Median of Daily Lambda
The high‐frequency cost‐per‐dollar‐volume benchmark, daily lambda, is calculated from every
trade and corresponding BBO quote in the SIRCA Thomson Reuters Tick History database for a
sample stock‐day. The daily cost‐per‐dollar‐volume proxies are calculated from daily stock price
data for a sample stock‐day. The earlier (later) sample spans 42 exchanges around the world
from 1996‐2007 (2008‐2014). All stock‐days are contained within stock‐months with at least five
positive‐volume days and eleven non‐zero return days. A solid box means the highest
correlation or the lowest average root mean squared error (RMSE) in the row. Dashed boxes
mean correlations that are statistically indistinguishable from the highest correlation or average
RMSEs that are statistically indistinguishable from the lowest average RMSE in the row at the 5%
level. Bold‐faced numbers are statistically different from zero or proxies have predictive power
that is significant at the 5% level.
The Global Performance of Daily Cost‐Per‐Dollar‐Volume Proxies Compared to Daily Lambda
Panel A: Average Cross‐Sectional Correlation of Daily Cost‐Per‐Dollar‐Volume Proxies
Compared to Daily Lambda
Panel B: Average Spearman's Cross‐Sectional Correlation of Daily Cost‐Per‐Dollar‐Volume
Proxies Compared to Daily Lambda
Panel C: Portfolio Time‐Series Correlation of First Differences of Daily Cost‐Per‐Dollar‐Volume
Proxies Compared to the Daily Lambda
43
Table 5
Gibbs FHT High‐Low
Close‐
High‐Low
Closing %
Quo Sprd VoV(%Spread)
1996‐2007 0.679 0.688 0.727 0.749 0.864 0.820
2008‐2014 0.654 0.710 0.749 0.687 0.857 0.828
1996‐2007 0.548 0.672 0.567 0.613 0.899 0.849
2008‐2014 0.520 0.675 0.586 0.637 0.922 0.911
1996‐2007 0.640 0.725 0.855 0.873 0.892 0.923
2008‐2014 0.796 0.802 0.954 0.858 0.952 0.947
1996‐2007 0.209 0.150 0.237 0.245 0.468 0.392
2008‐2014 0.195 0.120 0.246 0.248 0.419 0.382
Panel E: Average Correlation of the Eight Contests Above
Summary Stat. 0.530 0.568 0.615 0.614 0.784 0.757
1996‐2007 0.019 0.019 0.017 0.015 0.016 0.018
2008‐2014 0.016 0.015 0.013 0.015 0.013 0.014
The Global Performance of Monthly Percent‐Cost Proxies Compared to Monthly Percent
Effective Spread
Panel D: Individual Stock Time‐Series Correlation of First Differences of Monthly Percent‐Cost
Proxies Compared to Monthly Percent Effective Spread
Panel F: Average Root Mean Squared Error (RMSE) of Monthly Percent‐Cost Proxies Compared to
Monthly Percent Effective Spread
The high‐frequency percent‐cost benchmark, monthly percent effective spread, is calculated
from every trade and corresponding BBO quote in the SIRCA Thomson Reuters Tick History
database for a sample stock‐month. The monthly percent‐cost proxies are calculated from daily
stock price data for a sample stock‐month. The earlier (later) sample spans 42 exchanges around
the world from 1996‐2007 (2008‐2014). All stock‐months are required to have at least five positive‐
volume days and eleven non‐zero return days. A solid box means the highest correlation or the
lowest average root mean squared error (RMSE) in the row. Dashed boxes mean correlations that
are statistically indistinguishable from the highest correlation or average RMSEs that are
statistically indistinguishable from the lowest average RMSE in the row at the 5% level. Bold‐
faced numbers are statistically different from zero or proxies have predictive power that is
significant at the 5% level.
Panel A: Average Cross‐Sectional Correlation of Monthly Percent‐Cost Proxies Compared to
Monthly Percent Effective Spread
Panel B: Average Spearman's Cross‐Sectional Correlation of Monthly Percent‐Cost Proxies
Compared to Monthly Percent Effective Spread
Panel C: Portfolio Time‐Series Correlation of First Differences of Monthly Percent‐Cost Proxies
Compared to Monthly Percent Effective Spread
44
Table 6
High‐Low
Close‐
High‐Low
Closing %
Quo Sprd VoV(%Spread)
1996‐2007 0.395 0.401 0.765 0.660
2008‐2014 0.529 0.476 0.790 0.716
1996‐2007 0.214 0.196 0.800 0.737
2008‐2014 0.272 0.235 0.824 0.853
1996‐2007 0.673 0.589 0.917 0.952
2008‐2014 0.359 0.316 0.834 0.760
1996‐2007 ‐0.013 0.088 0.233 0.271
2008‐2014 ‐0.013 0.091 0.222 0.266
Panel E: Average Correlation of the Eight Contests Above
Summary Stat. 0.302 0.299 0.673 0.652
1996‐2007 0.020 0.024 0.018 0.016
2008‐2014 0.019 0.023 0.017 0.026
Panel D: Individual Stock Time‐Series Correlation of First Differences of Daily Percent‐Cost
Proxies Compared to Daily Percent Effective Spread
Panel F: Average Root Mean Squared Error (RMSE) of Daily Percent‐Cost Proxies Compared to
Daily Percent Effective Spread
The high‐frequency percent‐cost benchmark, daily percent effective spread, is calculated from
every trade and corresponding BBO quote in the SIRCA Thomson Reuters Tick History database
for a sample stock‐day. The daily percent‐cost proxies are calculated from daily stock price data
for a sample stock‐day. The earlier (later) sample spans 42 exchanges around the world from
1996‐2007 (2008‐2014). All stock days are contained within stock‐months with at least five
positive‐volume days and eleven non‐zero return days. A solid box means the highest
correlation or the lowest average root mean squared error (RMSE) in the row. Dashed boxes
mean correlations that are statistically indistinguishable from the highest correlation or average
RMSEs that are statistically indistinguishable from the lowest average RMSE in the row at the 5%
level. Bold‐faced numbers are statistically different from zero or proxies have predictive power
that is significant at the 5% level.
The Global Performance of Daily Percent‐Cost Proxies Compared to Daily Percent Effective
Spread
Panel A: Average Cross‐Sectional Correlation of Daily Percent‐Cost Proxies Compared to Daily
Percent Effective Spread
Panel B: Average Spearman's Cross‐Sectional Correlation of Daily Percent‐Cost Proxies
Compared to Daily Percent Effective Spread
Panel C: Portfolio Time‐Series Correlation of First Differences of Daily Percent‐Cost Proxies
Compared to Daily Percent Effective Spread
45
Table 7
Gibbs
Impact
FHT
Impact
High‐Low
Impact
Close‐
High‐Low
Impact
Closing %
Quo Sprd
Impact Amihud VoV()
Largest quintile 0.244 0.224 0.235 0.222 0.236 0.215 0.344
Quintile 2 0.349 0.333 0.344 0.334 0.331 0.285 0.373
Quintile 3 0.290 0.249 0.277 0.281 0.275 0.235 0.326
Quintile 4 0.291 0.265 0.279 0.291 0.289 0.227 0.313
Smallest quintile 0.569 0.503 0.552 0.560 0.572 0.505 0.625
Smallest quintile 0.395 0.358 0.373 0.371 0.358 0.315 0.458
Quintile 2 0.488 0.412 0.477 0.475 0.472 0.392 0.523
Quintile 3 0.557 0.441 0.543 0.544 0.539 0.450 0.592
Quintile 4 0.581 0.468 0.565 0.566 0.573 0.526 0.649
Largest quintile 0.587 0.518 0.569 0.578 0.589 0.519 0.643
Largest quintile 0.364 0.135 0.375 0.361 0.362 0.371 0.418
Quintile 2 0.432 0.221 0.401 0.406 0.417 0.386 0.487
Quintile 3 0.367 0.233 0.340 0.340 0.369 0.301 0.414
Quintile 4 0.334 0.268 0.308 0.312 0.340 0.231 0.372
Smallest quintile 0.308 0.250 0.294 0.301 0.319 0.200 0.340
Smallest quintile 0.319 0.144 0.309 0.307 0.326 0.294 0.372
Quintile 2 0.428 0.218 0.409 0.403 0.414 0.388 0.476
Quintile 3 0.391 0.226 0.382 0.377 0.399 0.340 0.429
Quintile 4 0.349 0.260 0.336 0.342 0.356 0.251 0.388
Largest quintile 0.297 0.252 0.276 0.291 0.301 0.194 0.337
0.635 0.593 0.591 0.616 0.636 0.528 0.739
0.278 0.066 0.254 0.268 0.266 0.231 0.273
Panel E: Pooled Partial Correlations after Controlling for Log Size and Volatility
Panel F: Pooled Partial Spearman's Correlations after Controlling for Log Size and Volatility
Panel C: Individual Stock Time‐Series Correlations of First Differences by Size Quintiles
Panel D: Individual Stock Time‐Series Correlations of First Differences by Volatility Quintiles
The Global Performance of Monthly Cost‐Per‐Dollar‐Volume Proxies with Monthly Lambda by
Size and Volatility Quintiles and After Controlling for Size and Volatility
Panel A: Average Cross‐sectional Correlations by Size Quintiles
Panel B: Average Cross‐sectional Correlations by Volatility Quintiles
The monthly high‐frequency liquidity benchmark, lambda, is calculated from every trade and
corresponding BBO quote in the SIRCA Thomson Reuters Tick History database for a sample
stock‐month. The monthly cost‐per‐dollar‐volume proxies are calculated from daily stock price
data for a sample stock‐month. The full global sample spans 42 exchanges around the world
from 1996‐2014. All stock‐months are required to have at least five positive‐volume days and
eleven non‐zero return days.
46
Table 8
Gibbs FHT High‐Low
Close‐
High‐Low
Closing %
Quo Sprd
VoV
(%Spread)
Largest quintile 0.200 0.435 0.266 0.263 0.635 0.552
Quintile 2 0.176 0.474 0.229 0.279 0.680 0.540
Quintile 3 0.265 0.520 0.361 0.383 0.724 0.564
Quintile 4 0.405 0.553 0.509 0.482 0.776 0.594
Smallest quintile 0.662 0.638 0.700 0.720 0.828 0.768
Smallest quintile 0.168 0.471 0.339 0.371 0.681 0.606
Quintile 2 0.159 0.502 0.381 0.394 0.712 0.658
Quintile 3 0.193 0.496 0.428 0.431 0.746 0.675
Quintile 4 0.230 0.482 0.481 0.466 0.768 0.702
Largest quintile 0.594 0.625 0.690 0.709 0.847 0.781
Largest quintile 0.187 0.082 0.207 0.197 0.373 0.315
Quintile 2 0.203 0.110 0.212 0.216 0.454 0.386
Quintile 3 0.208 0.154 0.240 0.243 0.506 0.419
Quintile 4 0.212 0.203 0.265 0.282 0.543 0.446
Smallest quintile 0.245 0.255 0.285 0.307 0.510 0.454
Smallest quintile 0.212 0.099 0.219 0.222 0.387 0.333
Quintile 2 0.198 0.106 0.207 0.210 0.435 0.369
Quintile 3 0.185 0.130 0.215 0.216 0.474 0.387
Quintile 4 0.201 0.189 0.253 0.267 0.515 0.423
Largest quintile 0.263 0.266 0.325 0.344 0.546 0.473
0.376 0.461 0.512 0.514 0.758 0.657
0.082 0.461 0.149 0.223 0.824 0.620
Panel D: Individual Stock Time‐Series Correlations of First Differences by Volatility Quintiles
Panel F: Pooled Partial Spearman's Correlations after Controlling for Log Size and Volatility
Panel E: Pooled Partial Correlations after Controlling for Log Size and Volatility
The monthly high‐frequency liquidity benchmark, percent effective spread, is calculated from
every trade and corresponding BBO quote in the SIRCA Thomson Reuters Tick History database
for a sample stock‐month. The monthly percent‐cost proxies are calculated from daily stock price
data for a sample stock‐month. The full global sample spans 42 exchanges around the world from
1996‐2014. All stock‐months are required to have at least five positive‐volume days and eleven
non‐zero return days.
The Global Performance of Monthly Percent‐Cost Proxies with Monthly Percent Effective Spread
by Size and Volatility Quintiles and After Controlling for Size and Volatility
Panel A: Average Cross‐sectional Correlations by Size Quintiles
Panel B: Average Cross‐sectional Correlations by Volatility Quintiles
Panel C: Individual Stock Time‐Series Correlations of First Differences by Size Quintiles
47
Table 9
a b c VoV() Amihud
1996‐2007 0.829 0.512 1.220 0.780 0.777 0.678
(0.001) (0.001) (0.004) (0.001) (0.001) (0.001)
2008‐2014 0.526 0.485 0.222 0.425 0.405 0.308
(0.014) (0.003) (0.010) (0.003) (0.003) (0.003)
a b c VoV(%Spread)
Closing %
Quo Sprd
1996‐2007 0.640 0.238 2.400 0.836 0.817 0.863
(0.001) (0.000) (0.004) (0.000) (0.000) (0.000)
2008‐2014 0.723 0.229 2.700 0.849 0.826 0.849
(0.003) (0.001) (0.024) (0.002) (0.002) (0.002)
The Global Performance of Monthly General Volatility Over Volume Class Proxies Under Optimal
Parameters And Top Monthly Liquidity Proxies Compared to Monthly Liquidity Benchmarks
The high‐frequency benchmarks, monthly percent effective spread and monthly lambda, are
calculated from every trade and corresponding BBO quote in the SIRCA Thomson Reuters Tick
History database for a sample stock‐month. The monthly general volatility over volume class
proxies and top liquidity proxies are calculated from daily stock price data for a sample stock‐
month. The earlier (later) sample spans 42 exchanges around the world from 1996‐2007 (2008‐
2014). All stock‐months are required to have at least five positive‐volume days and eleven non‐
zero return days. The general volatility over volume model specification is a*( ^b)/(V^c) . The
model parameters are estimated by maximizing pooled correlation with the liquidity benchmark
using non‐linear least squares over all stock‐months. Standard errors are reported in the
parenthesis.
General Volatility Over Volume Class Proxy Under Top Percent‐Cost Proxies
Panel B: Pooled Correlation of a Monthly General Volatility Over Volume Class Proxy and Top
Monthly Percent‐Cost Proxies Compared to Monthly Percent Effective Spread
General Volatility Over Volume Class Proxy Under
Parameters That Maximize Pooled Correlation
Top Cost‐Per‐Dollar‐
Volume Proxies
Panel A: Pooled Correlation of a Monthly General Volatility Over Volume Class Proxy and Top
Monthly Cost‐Per‐Dollar‐Volume Proxies Compared to Monthly Lambda
48
Table 10
Gibbs
Impact
FHT
Impact
High‐Low
Impact
Close‐
High‐Low
Impact
Closing %
Quo Sprd
Impact Amihud VoV()
US Monthly 0.579 0.435 0.603 0.606 0.580 0.489 0.663
US Daily 0.476 0.424 0.536 0.556 0.694
US Monthly 0.770 0.774 0.778 0.773 0.760 0.763 0.788
US Daily 0.445 0.224 0.600 0.563 0.656
US Monthly 0.528 0.557 0.517 0.490 0.536 0.204 0.551
US Daily 0.250 0.269 0.352 0.364 0.530
US Monthly 0.354 0.201 0.379 0.384 0.378 0.276 0.333
US Daily 0.064 0.076 0.098 0.123 0.304
Panel E: Average Correlation of the Contests Above
US Monthly 0.558 0.492 0.569 0.563 0.564 0.433 0.584
US Daily 0.309 0.248 0.396 0.401 0.546
US Monthly 18.1 18.3 18.3 18.2 18.1 17.7 23.3
US Daily 13 13 13.1 13 26.6
Panel F: Average Root Mean Squared Error (RMSE) of Monthly (Daily) Cost‐Per‐Dollar‐Volume
Proxies Compared to Monthly (Daily) Lambda / Median of Monthly (Daily) Lambda
The US Performance of Monthly (Daily) Cost‐Per‐Dollar‐Volume Proxies Compared to Monthly
(Daily) Lambda from 1996‐2014
The high‐frequency percent‐cost benchmark, monthly (daily) lambda, is calculated from every
trade and corresponding BBO quote in the NYSE Trade and Quote database for a sample stock‐
month (stock‐day). The monthly (daily) cost‐per‐dollar‐volume proxies are calculated from CRSP
daily stock price data for a sample stock‐month (stock‐day). The US sample spans the NYSE,
AMEX, and NASDAQ exchanges from 1996‐2014. All stock‐months (stock‐days) are required to
have (are contained within stock‐months with) at least five positive‐volume days and eleven non‐
zero return days. A solid box means the highest correlation or the lowest average root mean
squared error (RMSE) in the row. Dashed boxes mean correlations that are statistically
indistinguishable from the highest correlation or average RMSEs that are statistically
indistinguishable from the lowest average RMSE in the row at the 5% level. Bold‐faced numbers
are statistically different from zero or proxies have predictive power that is significant at the 5%
level.
Panel A: Average Cross‐Sectional Correlation of Monthly (Daily) Cost‐Per‐Dollar‐Volume Proxies
Compared to Monthly (Daily) Lambda
Panel B: Average Spearman's Cross‐Sectional Correlation of Monthly (Daily) Cost‐Per‐Dollar‐
Volume Proxies Compared to Monthly (Daily) Lambda
Panel C: Portfolio Time‐Series Correlation of First Differences of Monthly (Daily) Cost‐Per‐Dollar‐
Volume Proxies Compared to Monthly (Daily) Lambda
Panel D: Individual Stock Time‐Series Correlation of First Differences of Monthly (Daily) Cost‐Per‐
Dollar‐Volume Proxies Compared to the Monthly (Daily) Lambda
49
Table 11
Gibbs FHT High‐Low
Close‐
High‐Low
Closing %
Quo Sprd VoV(%Spread)
US Monthly 0.607 0.590 0.660 0.711 0.861 0.871
US Daily 0.358 0.361 0.742 0.767
US Monthly 0.630 0.604 0.666 0.720 0.909 0.938
US Daily 0.327 0.252 0.786 0.897
US Monthly 0.661 0.659 0.727 0.740 0.822 0.749
US Daily 0.066 0.148 0.396 0.424
US Monthly 0.181 0.055 0.179 0.202 0.282 0.263
US Daily ‐0.001 0.063 0.141 0.249
Panel E: Average Correlation of the Contests Above
US Monthly 0.520 0.477 0.558 0.593 0.719 0.705
US Daily 0.188 0.206 0.516 0.584
US Monthly 0.020 0.018 0.015 0.014 0.015 0.012
US Daily 0.018 0.024 0.014 0.012
Panel F: Average Root Mean Squared Error (RMSE) of Monthly (Daily) Percent‐Cost Proxies
Compared to Monthly (Daily) Percent Effective Spread
The US Performance of Monthly (Daily) Percent‐Cost Proxies Compared to Monthly (Daily)
Percent Effective Spread from 1996‐2014
The high‐frequency percent‐cost benchmark, monthly (daily) percent effective spread, is
calculated from every trade and corresponding BBO quote in the NYSE Trade and Quote database
for a sample stock‐month (stock‐day). The monthly (daily) percent‐cost proxies are calculated
from CRSP daily stock price data for a sample stock‐month (stock‐day). The US sample spans the
NYSE, AMEX, and NASDAQ exchanges from 1996‐2014. All stock‐months (stock‐days) are required
to have (are contained within stock‐months with) at least five positive‐volume days and eleven
non‐zero return days. A solid box means the highest correlation or the lowest average root mean
squared error (RMSE) in the row. Dashed boxes mean correlations that are statistically
indistinguishable from the highest correlation or average RMSEs that are statistically
indistinguishable from the lowest average RMSE in the row at the 5% level. Bold‐faced numbers
are statistically different from zero or proxies have predictive power that is significant at the 5%
level.
Panel A: Average Cross‐Sectional Correlation of Monthly (Daily) Percent‐Cost Proxies Compared
to Monthly (Daily) Percent Effective Spread
Panel B: Average Spearman's Cross‐Sectional Correlation of Monthly (Daily) Percent‐Cost Proxies
Compared to Monthly (Daily) Percent Effective Spread
Panel C: Portfolio Time‐Series Correlation of First Differences of Monthly (Daily) Percent‐Cost
Proxies Compared to Monthly (Daily) Percent Effective Spread
Panel D: Individual Stock Time‐Series Correlation of First Differences of Monthly (Daily) Percent‐
Cost Proxies Compared to Monthly (Daily) Percent Effective Spread
50
Table 12
VoV() Amihud
Before (July 30 ‐ Aug 3) 52.0% ‐5.8%
2.99** ‐0.06
During (Aug 6 ‐ Aug 9) 88.4% ‐2.1%
5.69*** ‐0.02
After (Aug 10 ‐ Aug 17) 118.9% 24.7%
6.25*** 0.23
Do VoV() and Amihud yield different inferences about the change in S&P 1500 illiquidity
before, during, and after the August 6‐9, 2007 "quant meltdown"?
Both VoV() and Amihud are used to measure the percent change in S&P 1500 illiquidity before
(July 30 ‐ August 3), during (August 6 ‐ August 9), and after (August 10 ‐ August 17) the August 6‐9,
2007 "quant meltdown." The base period is July 2 ‐ July 13, 2007. CRSP daily stock data provides
the inputs to both cost‐per‐dollar‐volume proxies. The t‐statistic is shown below each estimate.
*, **, and *** means significant at the 10%, 5%, and 1% levels, respectively.
Online Appendix for Are Volatility Over Volume Liquidity
Proxies Useful For Global Or US Research?
Table IA.1The Performance of Monthly Cost-Per-Dollar-Volume Proxies Compared to Monthly λ by Exchange
The high-frequency cost-per-dollar-volume benchmark, monthly lambda, is calculated from every trade and corresponding BBOquote in the SIRCA Thomson Reuters Tick History database for a sample stock-month. The monthly cost-per-dollar-volumeproxies are calculated from daily stock price data for a sample stock-month. The earlier (later) sample spans 42 exchangesaround the world from 1996-2007 (2008-2014). All stock-months are required to have at leastfive positive-volume days and eleven non-zero return days. A solid box means the highest correlation or the lowest averageroot mean squared error (RMSE) in the row. Dashed boxes mean correlations that are statistically indistinguishable from thehighest correlation or average RMSEs that are statistically indistinguishable from the lowest average RMSE in the row at the5% level. Bold-faced numbers are statistically different from zero or proxies have predictive power that is significant at the 5%level.
Average Cross-Sectional Correlation FD of Portfolio Time-Series Correlation
HLImpact
CHLImpact
Closing %QS
ImpactAmihud VoV(λ)
HLImpact
CHLImpact
Closing %QS
ImpactAmihud VoV(λ)
Argentina Buenos Ar. 0.729 0.757 0.758 0.701 0.834 0.429 0.518 0.706 0.569 0.732
Australia Australian 0.232 0.233 0.244 0.153 0.296 0.160 0.156 0.154 0.017 0.183
Austria Vienna 0.670 0.676 0.672 0.673 0.842 0.579 0.604 0.507 0.445 0.712
Belgium Brussels 0.642 0.648 0.648 0.606 0.747 0.380 0.405 0.418 0.312 0.686
Brazil Sao Paulo 0.572 0.566 0.549 0.514 0.602 0.193 0.200 0.166 0.079 0.183
Canada Toronto 0.754 0.775 0.721 0.680 0.858 0.554 0.574 0.509 0.308 0.688
Chile Santiago 0.362 0.291 0.417 0.377 0.405 0.368 0.242 0.373 0.588 0.509
China Hong Kong 0.543 0.551 0.552 0.424 0.645 0.370 0.435 0.307 0.299 0.431
China Shanghai 0.755 0.740 0.697 0.730 0.773 0.688 0.727 0.522 0.724 0.837
China Shenzhen 0.809 0.793 0.667 0.794 0.856 0.763 0.790 0.684 0.694 0.843
Denmark Copenhag. 0.413 0.420 0.440 0.388 0.487 0.331 0.311 0.334 0.348 0.482
France Paris 0.689 0.678 0.686 0.637 0.781 0.465 0.436 0.499 0.458 0.662
Finland Helsinki 0.351 0.354 0.371 0.336 0.402 0.398 0.350 0.329 0.350 0.465
Germany Frankfurt 0.498 0.507 0.534 0.459 0.638 0.755 0.720 0.673 0.472 0.752
Greece Athens 0.661 0.650 0.664 0.623 0.709 0.185 0.099 0.298 0.125 0.359
India Bombay 0.548 0.557 0.572 0.474 0.643 0.427 0.442 0.470 0.503 0.771
Indonesia Jakarta 0.386 0.391 0.396 0.289 0.430 0.319 0.325 0.327 0.262 0.386
Ireland Irish 0.505 0.526 0.525 0.509 0.596 0.176 0.237 0.178 0.100 0.263
Israel Tel Aviv 0.708 0.702 0.719 0.679 0.787 0.702 0.653 0.604 0.658 0.725
Italy Milan 0.538 0.536 0.536 0.485 0.619 0.593 0.508 0.419 0.424 0.179
Japan Tokyo 0.741 0.758 0.731 0.731 0.869 0.631 0.736 0.641 0.843 0.906
Malaysia Kuala Lum. 0.707 0.727 0.724 0.591 0.794 0.585 0.686 0.657 0.037 0.834
Mexico Mexican 0.755 0.762 0.750 0.635 0.817 0.506 0.562 0.652 0.299 0.703
Netherlands AEX 0.715 0.724 0.716 0.670 0.842 0.444 0.484 0.483 0.427 0.680
New Zeland New Zea. 0.491 0.505 0.511 0.444 0.544 0.338 0.343 0.356 0.096 0.328
Norway Oslo 0.424 0.434 0.432 0.393 0.469 0.537 0.532 0.500 0.446 0.538
Philippines Phillipine 0.397 0.406 0.398 0.334 0.460 0.148 0.156 0.143 0.255 0.254
Poland Warsaw 0.553 0.560 0.558 0.477 0.636 -0.252 -0.305 -0.363 -0.073 0.025
Portugal Lisbon 0.756 0.756 0.748 0.712 0.853 0.481 0.482 0.572 0.192 0.626
Singapore Singapore 0.686 0.691 0.688 0.557 0.754 0.516 0.562 0.545 0.446 0.724
South Africa Johannes. 0.624 0.630 0.648 0.415 0.699 0.491 0.517 0.567 0.179 0.593
South Korea Korea 0.802 0.778 0.765 0.715 0.896 0.676 0.707 0.735 0.736 0.812
Spain Barcelona 0.791 0.776 0.762 0.771 0.833 0.522 0.481 0.648 0.579 0.681
Sweden Stockholm 0.481 0.489 0.482 0.432 0.540 0.552 0.485 0.510 0.516 0.580
Switzerland SWX Swiss 0.513 0.522 0.525 0.466 0.592 0.533 0.568 0.559 0.419 0.717
Taiwan Taiwan 0.868 0.854 0.817 0.828 0.925 0.787 0.676 0.660 0.628 0.816
Thailand Thailand 0.327 0.336 0.319 0.262 0.382 0.159 0.255 0.216 0.417 0.422
Turkey Istanbul 0.746 0.743 0.694 0.687 0.775 0.761 0.852 0.731 0.600 0.892
UK London 0.737 0.719 0.753 0.613 0.829 0.508 0.563 0.673 0.403 0.700
US New York 0.753 0.746 0.686 0.719 0.810 0.496 0.498 0.396 0.535 0.474
US American 0.564 0.558 0.539 0.508 0.586 0.390 0.251 0.328 -0.364 0.335
US NASDAQ 0.659 0.674 0.635 0.512 0.725 0.633 0.668 0.614 0.232 0.625
US All US 0.603 0.606 0.580 0.488 0.663 0.517 0.490 0.536 0.204 0.551
Developed 0.411 0.424 0.409 0.339 0.489 0.324 0.353 0.318 0.146 0.561
Developing 0.556 0.564 0.573 0.512 0.654 0.769 0.775 0.777 0.805 0.931
Global 0.564 0.571 0.582 0.520 0.650 0.733 0.741 0.780 0.791 0.906
2
Table IA.2The Performance of Monthly Cost-Per-Dollar-Volume Proxies Compared to Monthly λ by Exchange
The high-frequency cost-per-dollar-volume benchmark, monthly lambda, is calculated from every trade and corresponding BBOquote in the SIRCA Thomson Reuters Tick History database for a sample stock-month. The monthly cost-per-dollar-volumeproxies are calculated from daily stock price data for a sample stock-month. The earlier (later) sample spans 42 exchangesaround the world from 1996-2007 (2008-2014). All stock-months are required to have at leastfive positive-volume days and eleven non-zero return days. A solid box means the highest correlation or the lowest averageroot mean squared error (RMSE) in the row. Dashed boxes mean correlations that are statistically indistinguishable from thehighest correlation or average RMSEs that are statistically indistinguishable from the lowest average RMSE in the row at the5% level. Bold-faced numbers are statistically different from zero or proxies have predictive power that is significant at the 5%level.
FD of Individual Stock Time-Series Correlation Average Root Mean Squared Error
HLImpact
CHLImpact
Closing %QS
ImpactAmihud VoV(λ)
HLImpact
CHLImpact
Closing %QS
ImpactAmihud VoV(λ)
Argentina Buenos Ar. 0.309 0.364 0.387 0.341 0.407 3.19 3.18 4.31 3.14 2.74
Australia Australian 0.175 0.185 0.203 0.130 0.257 11 11 8.99 13 50
Austria Vienna 0.145 0.198 0.137 0.211 0.473 4.63 4.61 4.39 4.48 9.35
Belgium Brussels 0.296 0.303 0.309 0.265 0.372 4.70 4.68 4.66 4.56 12
Brazil Sao Paulo 0.316 0.271 0.301 0.182 0.255 46 46 46 45 43
Canada Toronto 0.440 0.448 0.444 0.360 0.494 7.11 7.07 3.55 6.76 15
Chile Santiago 0.097 0.080 0.151 0.036 0.184 3.42 3.42 3.38 3.42 3.46
China Hong Kong 0.192 0.197 0.190 0.133 0.336 10 10 10 10 13
China Shanghai 0.624 0.629 0.585 0.613 0.677 0.96 0.96 0.96 0.96 0.37
China Shenzhen 0.631 0.633 0.579 0.592 0.693 1.06 1.06 1.06 1.06 0.34
Denmark Copenhag. 0.177 0.173 0.188 0.139 0.214 64 64 62 62 70
France Paris 0.349 0.363 0.341 0.318 0.451 6.68 6.62 6.85 6.70 15
Finland Helsinki 0.056 0.059 0.073 0.049 0.082 3.73 3.73 3.71 3.96 34
Germany Frankfurt 0.175 0.186 0.193 0.148 0.265 4.96 4.95 4.82 5.26 8.91
Greece Athens 0.480 0.460 0.502 0.370 0.326 4.63 4.57 5.16 4.31 15
India Bombay 0.315 0.324 0.347 0.242 0.367 4.49 4.48 4.56 4.19 3.16
Indonesia Jakarta 0.115 0.110 0.127 0.074 0.138 509 509 541 499 472
Ireland Irish 0.113 0.132 0.158 0.087 0.206 18 18 18 17 49
Israel Tel Aviv 0.427 0.431 0.469 0.327 0.424 3.99 3.99 3.91 3.91 2.30
Italy Milan 0.265 0.246 0.232 0.165 0.183 6.81 6.78 6.68 6.66 29
Japan Tokyo 0.392 0.420 0.440 0.432 0.501 3.00 3.00 2.67 3.00 2.33
Malaysia Kuala Lum. 0.435 0.468 0.472 0.256 0.530 2.56 2.55 2.53 2.45 2.28
Mexico Mexican 0.258 0.275 0.310 0.208 0.336 9.43 9.42 9.37 8.60 5.05
Netherlands AEX 0.279 0.283 0.270 0.256 0.393 10 10 9.97 9.74 18
New Zeland New Zea. 0.193 0.213 0.241 0.119 0.227 8.23 8.23 8.22 8.25 21
Norway Oslo 0.164 0.182 0.182 0.103 0.200 1830 1830 1830 1800 2270
Philippines Phillipine 0.077 0.092 0.087 0.084 0.142 203 203 192 201 208
Poland Warsaw 0.374 0.381 0.391 0.228 0.402 18 17 15 18 19
Portugal Lisbon 0.388 0.372 0.352 0.259 0.470 8.47 8.39 8.26 8.10 24
Singapore Singapore 0.255 0.270 0.271 0.211 0.344 5.74 5.72 5.76 6.21 19
South Africa Johannes. 0.250 0.247 0.289 0.119 0.315 15 15 15 16 16
South Korea Korea 0.572 0.568 0.578 0.435 0.582 2.56 2.34 2.36 3.53 225
Spain Barcelona 0.409 0.381 0.408 0.312 0.328 5.13 5.13 5.11 5.08 12
Sweden Stockholm 0.203 0.195 0.200 0.186 0.255 3.52 3.51 4.36 3.34 3.90
Switzerland SWX Swiss 0.291 0.303 0.312 0.235 0.394 3.16 3.15 3.18 3.14 8.38
Taiwan Taiwan 0.491 0.295 0.517 0.594 0.597 3.02 3.04 2.64 2.72 48
Thailand Thailand 0.014 0.017 0.004 0.027 0.050 46 46 20 45 42
Turkey Istanbul 0.662 0.619 0.701 0.544 0.500 1.91 1.91 1.91 1.90 2.56
UK London 0.273 0.258 0.286 0.232 0.317 15 15 15 15 7.92
US New York 0.281 0.269 0.262 0.263 0.220 6.67 6.66 6.61 6.63 12
US American 0.274 0.304 0.288 0.162 0.267 9.01 8.96 8.86 8.66 9.50
US NASDAQ 0.428 0.435 0.430 0.300 0.380 9.69 9.65 9.61 9.41 14
US All US 0.379 0.384 0.378 0.278 0.333 18 18 18 18 23
Developed 0.303 0.314 0.324 0.277 0.368 9.03 9.02 6.92 8.88 9.68
Developing 0.392 0.382 0.402 0.325 0.443 136 135 135 127 101
Global 0.347 0.347 0.363 0.301 0.405 82 82 88 77 62
3
Table IA.3The Performance of Daily Cost-Per-Dollar-Volume Proxies Compared to Daily λ by Exchange
The high-frequency cost-per-dollar-volume benchmark, daily lambda, is calculated from every trade and corresponding BBOquote in the SIRCA Thomson Reuters Tick History database for a sample stock-day. The daily cost-per-dollar-volume proxiesare calculated from daily stock price data for a sample stock-day. The earlier (later) sample spans 42 exchanges around theworld from 1996-2007 (2008-2014). All stock-months are required to have at leastfive positive-volume days and eleven non-zero return days. A solid box means the highest correlation or the lowest averageroot mean squared error (RMSE) in the row. Dashed boxes mean correlations that are statistically indistinguishable from thehighest correlation or average RMSEs that are statistically indistinguishable from the lowest average RMSE in the row at the5% level. Bold-faced numbers are statistically different from zero or proxies have predictive power that is significant at the 5%level.
Average Cross-Sectional Correlation FD of Portfolio Time-Series Correlation
HLImpact
CHLImpact
Closing %QS
ImpactAmihud VoV(λ)
HLImpact
CHLImpact
Closing %QS
ImpactAmihud VoV(λ)
Argentina Buenos Ar. 0.447 0.374 0.616 0.590 0.691 0.238 0.244 0.424 0.374 0.552
Australia Australian 0.187 0.179 0.229 0.260 0.350 -0.021 0.010 -0.014 0.019 0.099
Austria Vienna 0.504 0.419 0.583 0.609 0.779 0.154 0.218 0.130 0.295 0.480
Belgium Brussels 0.487 0.465 0.580 0.590 0.765 0.170 0.193 0.317 0.315 0.483
Brazil Sao Paulo 0.313 0.252 0.348 0.345 0.426 0.240 0.166 0.243 0.196 0.312
Canada Toronto 0.610 0.496 0.651 0.658 0.818 0.383 0.375 0.426 0.508 0.695
Chile Santiago 0.220 0.158 0.302 0.294 0.310 0.196 0.217 0.442 0.404 0.363
China Hong Kong 0.336 0.301 0.422 0.467 0.608 0.194 0.199 0.322 0.226 0.387
China Shanghai 0.429 0.309 0.467 0.498 0.658 0.034 0.026 0.060 0.003 0.255
China Shenzhen 0.472 0.337 0.518 0.561 0.764 0.052 0.035 0.085 -0.013 0.372
Denmark Copenhag. 0.311 0.240 0.375 0.376 0.448 0.062 0.099 0.101 0.090 0.268
France Paris 0.559 0.480 0.660 0.659 0.825 0.328 0.330 0.424 0.446 0.643
Finland Helsinki 0.217 0.183 0.259 0.267 0.367 0.190 0.176 0.247 0.228 0.350
Germany Frankfurt 0.406 0.361 0.569 0.573 0.678 0.446 0.384 0.519 0.543 0.677
Greece Athens 0.570 0.415 0.600 0.575 0.750 0.372 0.355 0.223 0.492 0.597
India Bombay 0.451 0.376 0.525 0.515 0.650 0.273 0.204 0.189 0.268 0.380
Indonesia Jakarta 0.250 0.207 0.285 0.275 0.370 0.161 0.172 0.222 0.203 0.255
Ireland Irish 0.280 0.215 0.343 0.338 0.418 0.062 0.114 0.103 0.157 0.269
Israel Tel Aviv 0.546 0.386 0.680 0.700 0.823 0.322 0.305 0.522 0.545 0.680
Italy Milan 0.403 0.362 0.459 0.451 0.624 0.235 0.353 0.399 0.361 0.601
Japan Tokyo 0.528 0.454 0.628 0.631 0.816 0.150 0.217 0.375 0.366 0.576
Malaysia Kuala Lum. 0.471 0.410 0.575 0.546 0.680 0.314 0.312 0.499 0.282 0.476
Mexico Mexican 0.532 0.430 0.628 0.626 0.718 0.351 0.335 0.444 0.397 0.505
Netherlands AEX 0.545 0.473 0.613 0.650 0.788 0.224 0.208 0.259 0.273 0.479
New Zeland New Zea. 0.154 0.139 0.265 0.259 0.295 0.161 0.206 0.245 0.269 0.334
Norway Oslo 0.290 0.256 0.365 0.365 0.446 0.356 0.345 0.408 0.397 0.509
Philippines Phillipine 0.299 0.227 0.370 0.358 0.459 0.140 0.093 0.155 0.106 0.130
Poland Warsaw 0.506 0.417 0.593 0.594 0.766 0.116 0.128 0.078 0.145 0.458
Portugal Lisbon 0.517 0.413 0.593 0.589 0.762 0.342 0.343 0.430 0.443 0.664
Singapore Singapore 0.452 0.354 0.527 0.481 0.628 0.366 0.290 0.462 0.334 0.483
South Africa Johannes. 0.444 0.365 0.550 0.537 0.636 0.280 0.325 0.344 0.337 0.472
South Korea Korea 0.596 0.490 0.627 0.621 0.808 0.304 0.241 0.403 0.267 0.532
Spain Barcelona 0.566 0.451 0.611 0.622 0.800 0.367 0.348 0.492 0.475 0.637
Sweden Stockholm 0.356 0.318 0.426 0.429 0.531 0.311 0.324 0.371 0.360 0.470
Switzerland SWX Swiss 0.366 0.348 0.440 0.460 0.577 0.131 0.162 0.165 0.187 0.305
Taiwan Taiwan 0.590 0.432 0.643 0.651 0.802 0.193 0.237 0.331 0.195 0.532
Thailand Thailand 0.184 0.154 0.204 0.204 0.347 0.093 -0.021 0.124 0.144 0.195
Turkey Istanbul 0.584 0.412 0.660 0.578 0.698 0.212 0.257 0.335 0.442 0.592
UK London 0.440 0.365 0.556 0.604 0.711 0.344 0.315 0.366 0.483 0.714
US New York 0.524 0.421 0.569 0.606 0.737 0.092 0.079 0.221 0.138 0.349
US American 0.393 0.305 0.474 0.459 0.624 0.219 0.201 0.281 0.327 0.533
US NASDAQ 0.474 0.417 0.528 0.549 0.690 0.344 0.347 0.419 0.388 0.592
US All US 0.476 0.424 0.536 0.556 0.694 0.251 0.273 0.347 0.367 0.528
Developed 0.263 0.248 0.305 0.329 0.450 -0.198 -0.103 -0.156 -0.083 -0.110
Developing 0.435 0.381 0.472 0.476 0.572 0.566 0.543 0.521 0.561 0.595
Global 0.432 0.378 0.469 0.473 0.562 0.826 0.809 0.798 0.821 0.779
4
Table IA.4The Performance of Daily Cost-Per-Dollar-Volume Proxies Compared to Daily λ by Exchange
The high-frequency cost-per-dollar-volume benchmark, daily lambda, is calculated from every trade and corresponding BBOquote in the SIRCA Thomson Reuters Tick History database for a sample stock-day. The daily cost-per-dollar-volume proxiesare calculated from daily stock price data for a sample stock-day. The earlier (later) sample spans 42 exchanges around theworld from 1996-2007 (2008-2014). All stock-months are required to have at leastfive positive-volume days and eleven non-zero return days. A solid box means the highest correlation or the lowest averageroot mean squared error (RMSE) in the row. Dashed boxes mean correlations that are statistically indistinguishable from thehighest correlation or average RMSEs that are statistically indistinguishable from the lowest average RMSE in the row at the5% level. Bold-faced numbers are statistically different from zero or proxies have predictive power that is significant at the 5%level.
FD of Individual Stock Time-Series Correlation Average Root Mean Squared Error
HLImpact
CHLImpact
Closing %QS
ImpactAmihud VoV(λ)
HLImpact
CHLImpact
Closing %QS
ImpactAmihud VoV(λ)
Argentina Buenos Ar. 0.142 0.145 0.257 0.228 0.328 5.67 5.67 10 5.66 5.94
Australia Australian 0.014 0.042 0.058 0.077 0.170 23 23 29 23 221
Austria Vienna -0.003 0.082 0.087 0.091 0.340 5.00 4.95 5.57 4.94 19
Belgium Brussels 0.082 0.112 0.159 0.155 0.360 9.26 9.25 9.23 9.23 34
Brazil Sao Paulo 0.143 0.129 0.182 0.201 0.330 12 12 12 12 15
Canada Toronto 0.139 0.146 0.223 0.258 0.465 13 13 23 13 46
Chile Santiago 0.014 0.014 0.017 0.022 0.083 2.62 2.67 3.79 2.68 2.84
China Hong Kong -0.034 0.037 0.027 0.089 0.231 43 43 43 43 73
China Shanghai 0.051 0.034 0.070 0.070 0.274 2.06 2.06 2.06 2.05 5.78
China Shenzhen 0.062 0.040 0.080 0.068 0.292 2.53 2.53 2.54 2.53 4.17
Denmark Copenhag. 0.036 0.075 0.091 0.075 0.146 428 436 436 437 1320
France Paris 0.092 0.103 0.157 0.182 0.397 15 15 15 15 44
Finland Helsinki 0.001 0.015 0.029 0.044 0.118 3.29 3.31 3.31 3.31 23
Germany Frankfurt 0.034 0.071 0.118 0.141 0.251 8.03 8.05 7.98 7.98 33
Greece Athens 0.167 0.138 0.223 0.154 0.342 12 12 13 12 30
India Bombay 0.097 0.110 0.146 0.132 0.254 6.12 6.14 5.74 6.12 5.95
Indonesia Jakarta 0.034 0.019 0.045 0.061 0.114 3.58 3.59 3.84 3.59 3.48
Ireland Irish 0.079 0.086 0.124 0.150 0.218 15 15 16 15 100
Israel Tel Aviv 0.151 0.130 0.268 0.194 0.428 7.46 7.45 14 7.44 7.73
Italy Milan 0.040 0.067 0.085 0.058 0.234 28 28 28 28 103
Japan Tokyo 0.026 0.058 0.087 0.095 0.297 9.06 9.05 12 9.05 7.58
Malaysia Kuala Lum. 0.085 0.108 0.167 0.167 0.269 6.66 6.65 7.02 6.63 12
Mexico Mexican 0.087 0.099 0.194 0.177 0.285 9.57 8.83 8.79 8.82 8.16
Netherlands AEX 0.041 0.105 0.097 0.158 0.325 13 13 13 13 53
New Zeland New Zea. 0.059 0.073 0.118 0.130 0.135 46 46 46 46 265
Norway Oslo 0.047 0.070 0.083 0.116 0.186 166 174 174 174 299
Philippines Phillipine 0.032 0.030 0.093 0.117 0.161 176 177 182 177 185
Poland Warsaw 0.138 0.134 0.204 0.198 0.430 4.58 4.58 4.29 4.57 4.92
Portugal Lisbon 0.042 0.085 0.133 0.133 0.356 6.33 6.33 6.31 6.33 10
Singapore Singapore 0.037 0.049 0.073 0.082 0.170 13 13 16 13 40
South Africa Johannes. 0.093 0.082 0.145 0.155 0.256 26 26 25 25 40
South Korea Korea 0.079 0.064 0.113 0.076 0.268 7.39 7.36 7.53 7.27 159
Spain Barcelona 0.098 0.112 0.137 0.114 0.340 17 17 17 17 47
Sweden Stockholm 0.048 0.065 0.095 0.115 0.213 654 660 886 660 925
Switzerland SWX Swiss 0.070 0.099 0.125 0.154 0.328 5.41 5.41 5.40 5.39 39
Taiwan Taiwan 0.048 0.015 0.112 0.092 0.067 11 11 20 11 26
Thailand Thailand -0.022 0.004 -0.014 0.018 0.087 85 85 101 85 81
Turkey Istanbul 0.263 0.191 0.399 0.251 0.280 3.72 3.73 3.73 3.72 3.56
UK London 0.078 0.110 0.116 0.186 0.298 46 46 45 46 36
US New York 0.055 0.041 0.075 0.090 0.196 6.36 6.36 6.27 6.35 11
US American 0.094 0.062 0.102 0.123 0.317 4.58 4.58 4.63 4.57 7.75
US NASDAQ 0.061 0.094 0.107 0.151 0.346 9.00 8.99 9.03 8.98 20
US All US 0.064 0.076 0.098 0.132 0.304 13 13 13 13 27
Developed 0.047 0.074 0.104 0.122 0.281 14 14 14 14 20
Developing 0.068 0.069 0.119 0.112 0.239 165 165 156 165 167
Global 0.058 0.072 0.112 0.117 0.260 110 111 105 111 113
5
Table IA.5The Performance of Monthly Percent-Cost Proxies Compared to Monthly Percent Effective Spread by
Exchange
The high-frequency percent-cost benchmark, monthly percent effective spread, is calculated from every trade and correspondingBBO quote in the SIRCA Thomson Reuters Tick History database for a sample stock-month. The monthly percent-cost proxiesare calculated from daily stock price data for a sample stock-month. The earlier (later) sample spans 42 exchanges around theworld from 1996-2007 (2008-2014). All stock-months are required to have at leastfive positive-volume days and eleven non-zero return days. A solid box means the highest correlation or the lowest averageroot mean squared error (RMSE) in the row. Dashed boxes mean correlations that are statistically indistinguishable from thehighest correlation or average RMSEs that are statistically indistinguishable from the lowest average RMSE in the row at the5% level. Bold-faced numbers are statistically different from zero or proxies have predictive power that is signiffcant at the 5%level.
Average Cross-Sectional Correlation FD of Portfolio Time-Series Correlation
Country Exchange HL CHLClosing %
QSVoV
(% Spread)HL CHL
Closing %QS
VoV(% Spread)
Argentina Buenos Ar. 0.487 0.547 0.891 0.846 0.538 0.626 0.861 0.804
Australia Australian 0.814 0.836 0.929 0.900 0.732 0.796 0.899 0.853
Austria Vienna 0.576 0.645 0.854 0.863 0.471 0.498 0.805 0.728
Belgium Brussels 0.548 0.663 0.837 0.826 0.524 0.618 0.831 0.775
Brazil Sao Paulo 0.578 0.644 0.827 0.759 0.475 0.436 0.348 0.314
Canada Toronto 0.787 0.837 0.855 0.914 0.768 0.833 0.790 0.787
Chile Santiago 0.357 0.330 0.808 0.681 0.439 0.178 0.552 0.578
China Hong Kong 0.653 0.580 0.855 0.792 0.761 0.659 0.919 0.877
China Shanghai 0.214 0.141 0.716 0.536 0.384 0.409 0.736 0.674
China Shenzhen 0.209 0.128 0.657 0.490 0.471 0.482 0.533 0.697
Denmark Copenhag. 0.682 0.671 0.849 0.821 0.661 0.721 0.554 0.678
France Paris 0.618 0.677 0.836 0.811 0.682 0.668 0.833 0.711
Finland Helsinki 0.752 0.690 0.920 0.867 0.844 0.485 0.887 0.811
Germany Frankfurt 0.703 0.709 0.884 0.857 0.605 0.687 0.828 0.768
Greece Athens 0.612 0.681 0.801 0.732 0.576 0.588 0.759 0.715
India Bombay 0.591 0.637 0.764 0.743 0.716 0.746 0.781 0.764
Indonesia Jakarta 0.750 0.743 0.856 0.747 0.885 0.889 0.884 0.834
Ireland Irish 0.789 0.775 0.895 0.866 0.594 0.661 0.785 0.670
Israel Tel Aviv 0.651 0.558 0.811 0.833 0.346 0.079 0.583 0.721
Italy Milan 0.499 0.595 0.843 0.845 0.684 0.131 0.523 0.310
Japan Tokyo 0.404 0.481 0.915 0.860 0.771 0.771 0.929 0.893
Malaysia Kuala Lum. 0.683 0.709 0.887 0.810 0.723 0.768 0.835 0.828
Mexico Mexican 0.556 0.634 0.794 0.791 0.559 0.520 0.690 0.683
Netherlands AEX 0.692 0.772 0.905 0.904 0.656 0.766 0.840 0.753
New Zeland New Zea. 0.649 0.679 0.771 0.711 0.612 0.671 0.697 0.628
Norway Oslo 0.587 0.631 0.752 0.697 0.559 0.586 0.796 0.724
Philippines Phillipine 0.693 0.697 0.812 0.708 0.713 0.733 0.829 0.783
Poland Warsaw 0.502 0.570 0.724 0.698 0.366 0.428 0.712 0.591
Portugal Lisbon 0.805 0.838 0.922 0.856 0.816 0.781 0.886 0.792
Singapore Singapore 0.880 0.881 0.944 0.868 0.903 0.924 0.961 0.893
South Africa Johannes. 0.762 0.732 0.868 0.815 0.734 0.741 0.766 0.751
South Korea Korea 0.415 0.453 0.852 0.856 0.678 0.718 0.888 0.789
Spain Barcelona 0.624 0.514 0.870 0.831 0.560 0.467 0.414 0.728
Sweden Stockholm 0.780 0.816 0.905 0.885 0.787 0.815 0.846 0.877
Switzerland SWX Swiss 0.699 0.733 0.844 0.834 0.769 0.810 0.620 0.834
Taiwan Taiwan 0.318 0.226 0.869 0.876 0.350 0.111 0.839 0.857
Thailand Thailand 0.560 0.539 0.858 0.744 0.785 0.787 0.908 0.861
Turkey Istanbul 0.573 0.331 0.909 0.505 0.624 0.651 0.926 0.803
UK London 0.594 0.525 0.882 0.843 0.416 0.501 0.819 0.813
US New York 0.429 0.469 0.749 0.771 0.255 0.286 0.219 0.461
US American 0.636 0.700 0.830 0.830 0.533 0.624 0.453 0.579
US NASDAQ 0.609 0.666 0.885 0.871 0.659 0.663 0.849 0.721
US All US 0.660 0.711 0.861 0.871 0.727 0.740 0.822 0.749
Developed 0.733 0.737 0.848 0.835 0.901 0.837 0.879 0.915
Developing 0.723 0.705 0.866 0.813 0.851 0.802 0.904 0.933
Global 0.735 0.726 0.862 0.823 0.905 0.861 0.914 0.933
6
Table IA.6The Performance of Monthly Percent-Cost Proxies Compared to Monthly Percent Effective Spread by
Exchange
The high-frequency percent-cost benchmark, monthly percent effective spread, is calculated from every trade and correspondingBBO quote in the SIRCA Thomson Reuters Tick History database for a sample stock-month. The monthly percent-cost proxiesare calculated from daily stock price data for a sample stock-month. The earlier (later) sample spans 42 exchanges around theworld from 1996-2007 (2008-2014). All stock-months are required to have at leastfive positive-volume days and eleven non-zero return days. A solid box means the highest correlation or the lowest averageroot mean squared error (RMSE) in the row. Dashed boxes mean correlations that are statistically indistinguishable from thehighest correlation or average RMSEs that are statistically indistinguishable from the lowest average RMSE in the row at the5% level. Bold-faced numbers are statistically different from zero or proxies have predictive power that is significant at the 5%level.
FD of Individual Stock Time-Series Correlation Average Root Mean Squared Error
Country Exchange HL CHLClosing %
QSVoV
(% Spread)HL CHL
Closing %QS
VoV(% Spread)
Argentina Buenos Ar. 0.246 0.256 0.498 0.415 0.012 0.010 0.007 0.015
Australia Australian 0.324 0.340 0.573 0.487 0.016 0.014 0.015 0.013
Austria Vienna 0.221 0.264 0.341 0.415 0.012 0.011 0.009 0.011
Belgium Brussels 0.255 0.309 0.406 0.334 0.008 0.007 0.005 0.007
Brazil Sao Paulo 0.131 0.140 0.331 0.308 0.020 0.018 0.015 0.016
Canada Toronto 0.273 0.284 0.419 0.424 0.017 0.015 0.015 0.014
Chile Santiago 0.154 0.116 0.245 0.231 0.015 0.028 0.013 0.011
China Hong Kong 0.239 0.217 0.487 0.346 0.014 0.015 0.012 0.011
China Shanghai 0.162 0.166 0.425 0.404 0.007 0.007 0.002 0.003
China Shenzhen 0.202 0.212 0.376 0.386 0.007 0.007 0.003 0.003
Denmark Copenhag. 0.221 0.245 0.409 0.340 0.013 0.013 0.011 0.010
France Paris 0.258 0.296 0.383 0.323 0.014 0.013 0.010 0.012
Finland Helsinki 0.263 0.245 0.493 0.294 0.008 0.010 0.007 0.008
Germany Frankfurt 0.198 0.204 0.434 0.355 0.018 0.016 0.023 0.021
Greece Athens 0.201 0.247 0.433 0.191 0.019 0.017 0.015 0.025
India Bombay 0.276 0.313 0.413 0.398 0.038 0.034 0.032 0.045
Indonesia Jakarta 0.519 0.527 0.686 0.534 0.011 0.011 0.016 0.028
Ireland Irish 0.264 0.206 0.423 0.292 0.012 0.013 0.013 0.011
Israel Tel Aviv 0.237 0.185 0.483 0.322 0.027 0.028 0.027 0.018
Italy Milan 0.250 0.217 0.340 0.252 0.007 0.008 0.011 0.007
Japan Tokyo 0.237 0.287 0.549 0.486 0.006 0.007 0.004 0.006
Malaysia Kuala Lum. 0.301 0.344 0.608 0.513 0.011 0.011 0.013 0.018
Mexico Mexican 0.145 0.169 0.364 0.315 0.013 0.012 0.018 0.011
Netherlands AEX 0.302 0.331 0.453 0.412 0.010 0.009 0.007 0.007
New Zeland New Zea. 0.159 0.218 0.325 0.171 0.012 0.010 0.008 0.009
Norway Oslo 0.154 0.174 0.326 0.191 0.016 0.014 0.013 0.013
Philippines Phillipine 0.261 0.279 0.516 0.402 0.012 0.011 0.014 0.020
Poland Warsaw 0.174 0.187 0.358 0.233 0.034 0.033 0.036 0.045
Portugal Lisbon 0.244 0.252 0.361 0.312 0.009 0.008 0.005 0.009
Singapore Singapore 0.416 0.421 0.668 0.566 0.008 0.008 0.010 0.011
South Africa Johannes. 0.224 0.230 0.392 0.225 0.020 0.019 0.015 0.017
South Korea Korea 0.169 0.195 0.414 0.368 0.010 0.010 0.006 0.007
Spain Barcelona 0.322 0.278 0.457 0.306 0.006 0.007 0.005 0.004
Sweden Stockholm 0.215 0.227 0.432 0.267 0.015 0.012 0.011 0.012
Switzerland SWX Swiss 0.238 0.282 0.343 0.358 0.009 0.008 0.007 0.007
Taiwan Taiwan 0.172 -0.082 0.407 0.434 0.005 0.007 0.002 0.006
Thailand Thailand 0.266 0.265 0.620 0.474 0.009 0.009 0.006 0.018
Turkey Istanbul 0.204 0.150 0.601 0.367 0.005 0.005 0.002 0.005
UK London 0.118 0.123 0.301 0.252 0.018 0.018 0.016 0.013
US New York 0.123 0.144 0.143 0.224 0.008 0.008 0.008 0.005
US American 0.200 0.239 0.280 0.264 0.022 0.020 0.037 0.018
US NASDAQ 0.244 0.272 0.421 0.317 0.016 0.015 0.012 0.012
US All US 0.212 0.240 0.344 0.290 0.015 0.014 0.015 0.012
Developed 0.237 0.260 0.455 0.387 0.013 0.013 0.014 0.012
Developing 0.242 0.232 0.479 0.398 0.017 0.017 0.015 0.020
Global 0.239 0.246 0.467 0.392 0.015 0.015 0.015 0.016
7
Table IA.7The Performance of Daily Percent-Cost Proxies Compared to Daily Percent Effective Spread by Exchange
The high-frequency percent-cost benchmark, daily percent effective spread, is calculated from every trade and correspondingBBO quote in the SIRCA Thomson Reuters Tick History database for a sample stock-day. The daily percent-cost proxiesare calculated from daily stock price data for a sample stock-day. The earlier (later) sample spans 42 exchanges around theworld from 1996-2007 (2008-2014). All stock-months are required to have at least five positive volume days and eleven non-zeroreturn days. A solid box means the highest correlation or the lowest average root mean squared error (RMSE) in the row.Dashed boxes mean correlations that are statistically indistinguishable from the highest correlation or average RMSEs that arestatistically indistinguishable from the lowest average RMSE in the row at the 5% level. Bold-faced numbers are statisticallydifferent from zero or proxies have predictive power that is significant at the 5% level.
Average Cross-Sectional Correlation FD of Portfolio Time-Series Correlation
Country Exchange HL CHLClosing %
QSVoV
(% Spread)HL CHL
Closing %QS
VoV(% Spread)
Argentina Buenos Ar. 0.138 0.194 0.645 0.655 0.050 0.095 0.430 0.428
Australia Australian 0.526 0.492 0.781 0.741 0.147 0.236 0.577 0.441
Austria Vienna 0.290 0.324 0.751 0.772 0.044 0.171 0.299 0.487
Belgium Brussels 0.280 0.350 0.730 0.759 0.108 0.196 0.497 0.491
Brazil Sao Paulo 0.296 0.298 0.683 0.674 0.111 0.162 0.371 0.338
Canada Toronto 0.476 0.463 0.741 0.790 0.051 0.270 0.299 0.377
Chile Santiago 0.078 0.130 0.636 0.549 0.064 0.129 0.292 0.287
China Hong Kong 0.329 0.231 0.739 0.730 0.092 0.062 0.571 0.556
China Shanghai 0.082 0.042 0.473 0.629 -0.004 0.036 0.346 0.471
China Shenzhen 0.076 0.043 0.504 0.578 0.015 0.117 0.361 0.535
Denmark Copenhag. 0.326 0.293 0.713 0.658 0.587 0.590 0.833 0.631
France Paris 0.373 0.403 0.781 0.783 0.105 0.275 0.507 0.517
Finland Helsinki 0.471 0.375 0.805 0.755 0.113 0.169 0.465 0.471
Germany Frankfurt 0.495 0.439 0.767 0.710 0.632 0.530 0.715 0.476
Greece Athens 0.305 0.331 0.665 0.692 0.151 0.199 0.421 0.418
India Bombay 0.262 0.365 0.626 0.680 0.007 0.306 0.346 0.457
Indonesia Jakarta 0.472 0.411 0.808 0.672 0.145 0.319 0.638 0.523
Ireland Irish 0.407 0.389 0.791 0.779 0.190 0.291 0.628 0.588
Israel Tel Aviv 0.252 0.234 0.736 0.688 0.150 0.158 0.506 0.468
Italy Milan 0.222 0.260 0.660 0.803 -0.005 0.188 0.237 0.519
Japan Tokyo 0.158 0.193 0.694 0.731 -0.089 0.153 0.387 0.496
Malaysia Kuala Lum. 0.412 0.410 0.767 0.694 0.151 0.267 0.611 0.572
Mexico Mexican 0.191 0.242 0.689 0.660 0.080 0.144 0.432 0.329
Netherlands AEX 0.400 0.423 0.798 0.826 0.071 0.188 0.516 0.474
New Zeland New Zea. 0.357 0.336 0.634 0.627 0.085 0.171 0.315 0.335
Norway Oslo 0.316 0.343 0.714 0.674 0.123 0.227 0.494 0.440
Philippines Phillipine 0.374 0.342 0.747 0.608 0.063 0.212 0.541 0.420
Poland Warsaw 0.316 0.343 0.680 0.672 0.085 0.087 0.537 0.109
Portugal Lisbon 0.488 0.449 0.838 0.788 0.226 0.205 0.567 0.480
Singapore Singapore 0.635 0.518 0.844 0.786 0.385 0.308 0.670 0.592
South Africa Johannes. 0.477 0.429 0.772 0.731 0.211 0.321 0.549 0.479
South Korea Korea 0.174 0.187 0.666 0.822 -0.075 0.192 0.505 0.653
Spain Barcelona 0.335 0.213 0.695 0.819 0.024 0.114 0.118 0.489
Sweden Stockholm 0.452 0.472 0.810 0.768 0.170 0.298 0.571 0.418
Switzerland SWX Swiss 0.388 0.414 0.736 0.754 0.046 0.236 0.407 0.533
Taiwan Taiwan 0.118 0.081 0.606 0.789 -0.103 0.100 0.345 0.442
Thailand Thailand 0.243 0.242 0.707 0.618 0.040 0.107 0.495 0.482
Turkey Istanbul 0.281 0.105 0.828 0.393 -0.184 0.230 0.286 0.631
UK London 0.335 0.204 0.852 0.749 0.018 0.082 0.420 0.444
US New York 0.174 0.145 0.619 0.769 0.021 -0.006 0.053 0.064
US American 0.281 0.259 0.604 0.638 0.041 0.104 0.134 0.196
US NASDAQ 0.302 0.334 0.786 0.773 0.064 0.151 0.374 0.346
US All US 0.358 0.361 0.742 0.767 0.069 0.142 0.393 0.421
Developed 0.451 0.411 0.778 0.704 0.465 0.370 0.682 0.627
Developing 0.410 0.403 0.773 0.726 0.500 0.444 0.904 0.892
Global 0.445 0.429 0.774 0.718 0.598 0.516 0.905 0.901
8
Table IA.8The Performance of Daily Percent-Cost Proxies Compared to Daily Percent Effective Spread by Exchange
The high-frequency percent-cost benchmark, daily percent effective spread, is calculated from every trade and correspondingBBO quote in the SIRCA Thomson Reuters Tick History database for a sample stock-day. The daily percent-cost proxiesare calculated from daily stock price data for a sample stock-day. The earlier (later) sample spans 42 exchanges around theworld from 1996-2007 (2008-2014). All stock-months are required to have at least five positive volume days and eleven non-zeroreturn days. A solid box means the highest correlation or the lowest average root mean squared error (RMSE) in the row.Dashed boxes mean correlations that are statistically indistinguishable from the highest correlation or average RMSEs that arestatistically indistinguishable from the lowest average RMSE in the row at the 5% level. Bold-faced numbers are statisticallydifferent from zero or proxies have predictive power that is significant at the 5% level.
FD of Individual Stock Time-Series Correlation Average Root Mean Squared Error
Country Exchange HL CHLClosing %
QSVoV
(% Spread)HL CHL
Closing %QS
VoV(% Spread)
Argentina Buenos Ar. 0.011 0.062 0.277 0.231 0.016 0.019 0.011 0.013
Australia Australian -0.026 0.130 0.243 0.224 0.024 0.029 0.022 0.019
Austria Vienna 0.000 0.118 0.273 0.311 0.010 0.014 0.008 0.007
Belgium Brussels 0.014 0.123 0.275 0.299 0.010 0.013 0.008 0.006
Brazil Sao Paulo -0.015 0.043 0.157 0.170 0.025 0.028 0.020 0.020
Canada Toronto -0.014 0.101 0.205 0.320 0.022 0.026 0.019 0.017
Chile Santiago 0.029 0.026 0.287 0.135 0.016 0.017 0.016 0.013
China Hong Kong 0.001 0.075 0.281 0.281 0.023 0.028 0.016 0.016
China Shanghai 0.008 0.006 0.143 0.368 0.010 0.014 0.002 0.004
China Shenzhen 0.002 0.016 0.140 0.359 0.010 0.014 0.002 0.004
Denmark Copenhag. -0.030 0.118 0.268 0.249 0.018 0.020 0.014 0.015
France Paris -0.008 0.163 0.296 0.352 0.014 0.017 0.010 0.010
Finland Helsinki -0.017 0.127 0.244 0.274 0.013 0.017 0.010 0.011
Germany Frankfurt -0.005 0.068 0.205 0.161 0.032 0.037 0.037 0.060
Greece Athens -0.007 0.112 0.285 0.324 0.021 0.025 0.018 0.025
India Bombay -0.030 0.172 0.247 0.297 0.046 0.047 0.044 0.045
Indonesia Jakarta 0.011 0.114 0.367 0.227 0.019 0.027 0.013 0.028
Ireland Irish -0.025 0.149 0.302 0.258 0.016 0.020 0.013 0.012
Israel Tel Aviv -0.025 0.077 0.308 0.304 0.032 0.033 0.020 0.025
Italy Milan -0.012 0.094 0.219 0.350 0.010 0.014 0.013 0.006
Japan Tokyo -0.016 0.087 0.221 0.330 0.009 0.015 0.006 0.006
Malaysia Kuala Lum. -0.014 0.125 0.338 0.309 0.018 0.023 0.015 0.016
Mexico Mexican 0.011 0.064 0.250 0.154 0.016 0.018 0.018 0.014
Netherlands AEX -0.019 0.114 0.250 0.318 0.012 0.015 0.009 0.008
New Zeland New Zea. -0.042 0.111 0.213 0.175 0.012 0.014 0.011 0.010
Norway Oslo -0.022 0.114 0.261 0.229 0.019 0.022 0.015 0.014
Philippines Phillipine -0.011 0.109 0.406 0.203 0.019 0.024 0.014 0.018
Poland Warsaw -0.014 0.104 0.267 0.308 0.025 0.029 0.024 0.031
Portugal Lisbon -0.013 0.105 0.217 0.263 0.012 0.016 0.008 0.012
Singapore Singapore -0.012 0.090 0.224 0.257 0.015 0.020 0.012 0.013
South Africa Johannes. -0.023 0.121 0.303 0.226 0.024 0.026 0.020 0.019
South Korea Korea -0.024 0.081 0.183 0.428 0.014 0.020 0.007 0.009
Spain Barcelona 0.011 0.086 0.178 0.380 0.009 0.013 0.008 0.004
Sweden Stockholm -0.021 0.105 0.220 0.249 0.020 0.024 0.016 0.016
Switzerland SWX Swiss -0.027 0.141 0.248 0.319 0.012 0.015 0.010 0.008
Taiwan Taiwan -0.032 -0.010 0.152 0.041 0.008 0.015 0.003 0.007
Thailand Thailand -0.009 0.082 0.260 0.245 0.015 0.021 0.007 0.017
Turkey Istanbul -0.028 0.058 0.170 0.378 0.009 0.014 0.002 0.007
UK London 0.007 0.025 0.215 0.124 0.021 0.024 0.018 0.016
US New York -0.006 0.023 0.036 0.123 0.009 0.014 0.008 0.003
US American 0.001 0.048 0.104 0.191 0.024 0.029 0.024 0.019
US NASDAQ -0.000 0.083 0.192 0.315 0.019 0.025 0.013 0.012
US All US -0.001 0.063 0.141 0.249 0.018 0.024 0.014 0.012
Developed -0.013 0.092 0.224 0.258 0.018 0.022 0.016 0.019
Developing -0.013 0.084 0.244 0.285 0.021 0.025 0.017 0.019
Global -0.013 0.088 0.233 0.271 0.020 0.024 0.018 0.020
9