The Cost of Liquidity:

Results from a Natural Experiment1

Eugene Kandel
School of Business Administration
and Department of Economics
Hebrew University and CEPR,
mskandel@mscc.huji.ac.il
and
Isabel Tkatch2
Department of Finance
J. Mack Robinson College of Business
Georgia State University
itkatch@gsu.edu
This Version: November 2009

We thank Reza Mahani and seminar participants at the University of Arizona

for helpful comments. We also thank the Tel Aviv Stock Exchange for providing
us with the data and we gratefully acknowledge nancial support from the Krueger
Center for Financial Research.
2
Corresponding author.

Abstract
Tel Aviv Stock Exchange is the only market open on Sundays and trades
ETNs on local and US indices. We use this feature to estimate the causal
relation between the inventory cost and the bid-ask spread. We compare
spreads on the most liquid indices on Sundays and other days using dierencesin-dierences approach to estimate the inventory component of the spread
without making structural assumptions. We show that on Sundays the bid
ask spread in the US ETNs more than doubles relative to the rest of the
week, indicating that inventory cost is an economically signicant transaction cost. The same eect is found during the weekday morning hours when
the Israeli market is open, but the European markets are still closed - the
inventory component can reach as high as 80% of the spread during some
time intervals. No such eects are observed in the Israeli indices.
JEL classication: G10, G12, G15.
Key words: Inventory, Liquidity, Limit order book, ETF.

Introduction

We use a unique feature of the Tel Aviv Stock Exchange (TASE) to estimate the causal relation between the cost of inventory and the bid-ask
spread. TASE is the only market in the world that is open on Sundays,
and which trades Exchange Traded Notes (ETNs) on domestic and foreign
indices throughout the week.1 The ETN market is organized as an electronic limit order book, where the issuers of ETNs implicitly act as market
makers providing additional liquidity. All ETN liquidity providers, including issuers, can hedge their exposure to ETNs on Israeli indices throughout
the week, however, they cannot hedge their exposure to ETNs on foreign
indices on Sunday and holidays.2 This implies that the cost of carrying inventory of foreign ETNs on Sundays is much higher than during the rest of
the week, while there is no corresponding dierence for the domestic ETNs.
We use this natural experiment to estimate the eect of inventory costs on
the spread in ETN markets. We nd that on Sundays and holidays the
bid ask spread in foreign ETNs more than doubles relative to the rest of
the week, indicating that inventory holding cost imposes a large transaction
cost. The same eect is found during the weekday morning hours, when the
Israeli market is open but the European markets are still closed. No eects
are found in the Israeli ETNs.
Our methodological approach is completely dierent from those used in
previous studies of the eect of inventory costs on the spread, which fall into
two broad categories: studies that estimate spread components, and those
1

Exchange Traded Notes (ETNs) are similar to Exchange Traded Funds (ETFs), except
that they are structured as bonds, and the issuer is obligated to follow the index rather
than to make the best eort. The dierence is due to the Israeli regulatory environment
and as far as we can tell has no bearing on the issue we study.
2
Hereafter we will use the terms US / foreign ETNs and Israeli / domestic ETNs to
shorten for ETNs on US indices and ETNs on Israeli Indices.

that relate spreads to the actual inventory of the dealer. Empirical papers in
the rst category (e.g. Glosten and Harris 1988, Hasbrouck 1988, Stoll 1989,
George, Kaul and Nimalendran 1991, Huang and Stoll 1997, Madhavan,
Richardson and Roomans 1997, Bollen, Smith and Whaley 2004) impose
strong assumptions on the order arrival and the fundamental price processes
to separate the variation in the bid-ask spread associated with inventory
control considerations from that associated with adverse selection and from
the order processing costs (see Hasbrouck 2002). The common approach is
to model the price process as the sum of transitory and permanent changes:
transitory price changes are attributed to inventory control, while permanent
ones are attributed to information.3 Most studies nd that inventory costs
constitute a small proportion of the spread.
The second category includes papers that use data of market makers on
various exchanges (e.g. Hasbrouck and Soanos 1993, Madhavan and Smidt
1993, Hansh, Naik, and Vishwanathan 1998, and Reiss and Werner 1998),
and show that dealers hold large inventories, which deviate from their target
levels over long periods of time. The inventory positions are endogenous, and
so is the spread, which requires instruments to determine causality. Some try
to address this problem (e.g. Reiss and Werner 1998), but the identication
relies on similarly strong structural assumptions about the price and order
arrival processes.
Our approach does not impose any structural assumptions, and does
not make use of the dealersendogenous inventory levels. Instead, we focus
3

Huang and Stoll (1997) is an interesting example of the eect of the structural assumption on the estimated eects. Their two-way decomposition of the spread implies
that inventory and adverse selection together account for 11% of the spread. Yet, the
result is dierent in a three-way decomposition, which that requires additional structural
assumptions needed to separate the adverse selection component from the inventory component. Now, the inventory component is estimated as 29% of the spread and the adverse
selection is 10%.

on changes in the inventory holding cost due to an exogenous shock which

prevents hedging positions on Sundays, as opposed to any other day of
the week. This allows us to establish a causal relation between inventory
management costs and the bid-ask spread and to estimate the magnitude
of this eect in a clean way. To the best of our knowledge, this is the
rst study in which the inventory eect is estimated without making any
structural assumptions about the order ow and price processes. We also
nd a larger inventory cost component of the bid ask spread than previously
estimated.
Our ndings suggest that in order to reduce the cost of liquidity, it is
important to allow the suppliers of liquidity to hedge their positions and
control their inventory exposure. Derivative markets, as well as trading
venues for designated market makers, are important for ensuring the low
cost of liquidity in the main (spot) security markets.4
The paper is organized as follows. Section 2 presents the data and the
empirical hypotheses. Section 3 presents the results and Section 4 concludes.

Hypotheses and Data

In this section we describe the specic features of the Tel Aviv Stock Exchange (TASE) that allow us to estimate the eect of the inventory cost
4
Our study is tangentially related to the literature on the relation between derivatives
and spot markets. Facilitating e cient risk sharing is the main argument supporting new
derivatives markets, yet, the benets of an increased speculative activity associated with
them are controversial (for example, see the theoretical discussion in Stein, 1987, and the
empirical analyses in Bessembinder and Seguin, 1992). In our data, we cannot determine
which traders provide liquidity. They could be better diversied (Stoll, 1978), informed
(Bloomeld et al., 2005), speculators (Grossman and Miller, 1988) or patient traders
(Foucault et al., 2005). Even though we can not test how any one of the trader types
aects liquidity, we are able to show that when liquidity providers can hedge against
inventory risks the costs of liquidity provision decreases, depth increases, spreads are
narrow and the traded volume is high.

on the spread and establish a causal relation; we postulate the empirical

hypotheses and describe the data.

2.1

TASE Features and Empirical Hypotheses

TASE equity market operates quite similarly to Paris Euronext and Milan
Exchange. Trading starts with a call auction, followed by a continuous phase
and culminates with a closing phase. The day starts with an empty order
book at 8:30 AM, when traders start submitting limit and market orders.
At a random time between 9:45AM and 9:50AM all the submitted orders
are crossed, using an auction mechanism with time and price priority rules,
and the continuous trading phase begins. The continuous phase is a limit
order book, with designated market makers in some securities. All traders
observe the best three prices and quantities on each side. Traders may post
either market or limit orders, and those are executed according to price and
time priority rules. The closing phase begins at 4:55 PM; during our sample
period it was a simple crossing of tradersmarket orders that were executed
at the closing price. All unexecuted orders are cancelled at the end of the
trading day and the next day starts, again, with an empty book. Equities,
bonds, index options, futures contracts and ETNs are all traded on the same
TASE platform, with some minor dierences in hours, minimum order size
and tick size.
Since their inception, ETNs gained popularity in Israel as an inexpensive
tool for retail and institutional investors to get exposure to local and foreign
indices. We focus on three institutions that dominate the market with a
large number of oerings of ETNs on Israeli and US indices.5 The issuers
serve as informal market makers and supply liquidity for their ETNs along
5

We chose ETNs on the major US indices, since they are much more liquid in Israel
than ETNs on other foreign indices.

with other market participants.

We utilize a specic feature in the TASE trading schedule that is driven
by the Jewish calendar, according to which Israeli markets and institutions
are closed on Fridays, but open on Sundays. This means that between Monday and Thursday market makers and other liquidity suppliers can fully
hedge their open positions in foreign indices in the European futures markets. They do not use the US markets due to time dierences - US markets
open shortly before the Israeli markets close. European markets are closed
on Sunday, thus liquidity providers in US ETNs must carry the risk until the
next day, making the inventory cost on Sundays much higher than during
the rest of the week. Since the Israeli stock market is open on Sunday, no
such problem arises for Israeli ETNs.6
These features, combined with the implications of Grossman and Miller
(1988) and with the standard inventory models (e.g. Stoll 1978, Amihud
and Mendelson 1980, Ho and Stoll 1983, and OHara and Oldeld 1986)
allow us to postulate the following empirical hypotheses. Since carrying
inventory on Sundays is more expensive, liquidity providers will demand
a higher premium. The same intuition applies to intraday spreads, since
European futures markets open two hours later than the Israeli market,
which leads to the rst hypothesis.
H1: Daily (Hourly) average bid ask spreads in ETNs on foreign indices,
during the days (hours) when European futures markets are closed are higher
than the spreads during the days (hours) when these markets are open. There
is no similar prediction for ETNs on Israeli indices.
Stoll (1989) suggests that the trading volume is inversely related to the
inventory holding period and, therefore, aects the dealers ability to return
6

Bollen, Smith and Whaley (2004) show that the availability of options reduces inventory costs on NASDAQ, since dealers may use them to hedg their inventory position.

to his preferred inventory level. We expect the eect of trading volume to

be more pronounced for US ETNs on Sundays.
H2: The e ect of trading volume on the daily average bid ask spread
during the days when European futures markets are closed is higher than
the same e ect during the days when these markets are open. There is no
similar prediction for ETNs on Israeli indices.
Chordia, Roll and Subrahmanyam (2002) nd that higher order imbalances tend to increase spreads. When liquidity providers can hedge their
exposure this eect should be rather small, but it should manifest itself
mostly when hedging is not available.
H3: The e ect of order imbalances on the daily average bid ask spread
during the days when European futures markets are closed is higher than
the same e ect during the days when these markets are open. There is no
similar prediction for ETNs on Israeli indices.
Finally, we use additional data for the year of 2008 (before and after
Lehman Brothers led for bankruptcy) and compare the spreads to those
in 2006. The additional data presents one more natural experiment which
allows us to test the eect of an exogenous shock to price volatility on the
cost of inventory and, through the inventory channel, on the bid-ask spread.
Our design allows for separation between the eect of volatility on the spread
through the free option channel suggested by Copeland and Galai (1983),
and the eect of volatility on the spread through the inventory channel
suggested by Stoll (1978) and OHara and Oldleld (1986).7 In periods of
high volatility we expect spreads to be wider due to an increase in the cost
7

Note that in Stoll(1978) and OHara and Oldeld (1986) price volatility aects the
spread because it increases the probability of loss due to an adverse price change after the
trade in which inventory was acquired (inventory eect). In Copeland and Galai (1983)
price volatility aects the spread because it increases the probability of loss due to an
adverse price change before the trade (free option / picking o eect).

of inventory.
H4: The inventory component of the spread for US ETNs on Sundays
(described in Hypothesis H1) increases in the volatility of the price process.
When price volatility is low (in 2006) we expect the inventory component to
be small; when price volatility is high (in the second period of 2008, after
Lehman Brothers led for bankruptcy) we expect the inventory component to
be large.

2.2

Sample and Descriptive Statistics

We have chosen a sample period of one year: January 1 to December 31, 2006.
During this period there were no drastic events aecting the TASE market,
and there were no signicant regulatory changes. Our methodology relies on
dierences between days of the week thus, which requires a relatively long
sample period. We chose 14 most liquid ETNs and obtain quotes, orders
and transactions data on all the above mentioned ETNs from TASE. We
restrict our sample to the three largest issuers, each of whom has ETNs on
both Israeli and US indices. In Table 1 we present summary statistics for
the sample of ETNs, and it is evident that there is a similar pattern for all
issuers: ETNs on the Israeli indices have higher daily volumes and lower
spreads.
As stated above, we are interested in comparisons of the percentage
quoted bid-ask spread across two regimes: on days/hours when the European
futures exchanges are closed (Sundays and European holidays, as well as
the morning hours of regular weekdays) versus other trading days/hours.
Unlike the US ETNs, the Israeli ETNs are not predicted to exhibit signicant
dierences under the two regimes.8 In Table 2 we present summary statistics
8

We will shorten the description of ETNs and use the terms US/foreign ETNs and
Israeli/domestic ETNs.

for Israeli and US ETNs and the two regimes. The bid ask spread for
Israeli ETNs is practically unaected by Sundays (0.11% versus 0.13%),
while the spread for US ETNs is more than doubled (0.40% versus 0.94%).
The spreads for US ETNs are higher on any day of the week, which may
result from FX exposure risk, lower liquidity and higher cost of hedging
abroad. It is clear that the Sunday eect on Israeli ETNs is trivial while the
eect on US ETNs is large.
It is evident that other variables are also aected by the type of the
ETN and by holidays. The average daily trading volume, which proxies
for liquidity, is higher for Israeli ETNs and it decreases much more for US
ETNs on Sundays. We also dene a measure of order imbalance as the net
buy-side order volume scaled by the total order volume for the day. This
measure ranges from minus one to one, and assumes the extreme values if
all the orders are either sell side (-1) or buy-side orders (+1). Since we are
interested in the imbalance for orders that demand liquidity, we consider only
market orders in this calculation.9 Order imbalance is on average negative
and, on average, it is more extreme for Israeli ETNs. For US ETNs on
Sundays, we observe the lowest average imbalance.
Lastly, we show that there is intense competition among liquidity providers
on TASE. The daily order volume is about 1,000 times higher than the
traded volume, and practically all of that volume is limit orders (supply of
liquidity). Market orders, which constitute about 35% of the order ow for
the most liquid stocks traded on TASE, constitute only 2% of the ETNs
daily order ow. This is what we expect to see in a market with high frequency computerized trading, in which traders make money as voluntary
9

We will use the term market order to describe both market and marketable limit
orders, since market orders per se are almost never used on TASE. A marketable limit
order is a dened as limit buy (sell) order priced at or above the best ask (bid).

suppliers of liquidity. Hendershott et al. (2009) show that this phenomena contributed to the decrease in spreads on the NYSE. Even though the
trading and order volumes for US ETNs Sundays are low relative to other
trading day, we still nd 94% limit order in the order ow, which suggests
an intense competition among liquidity suppliers.

Results

We start by modeling the daily average bid-ask spread. Then, we investigate a more detailed intraday version of the data, in which the dependent
variable is the average bid-ask spread measured for ve 90 minutes time
intervals during the day. We use two dierent estimation approaches: a
regression model with clustered (robust) standard errors and a mixed linear
model approach.10 The disadvantage of mixed linear models is the assumption of normality,11 but it has one major advantage. We are able to explicitly model a specic variance-covariance structure, which contributed to
our understanding of inventories and spreads beyond the control for possible
correlations and heteroskedasticity.

3.1

The Daily Average Bid-Ask Spread

The daily panel data regression is designed for a simple dierences-in-dierences

analysis: we are interested in the change in spreads for Israeli ETNs versus
the change in spreads for US ETNs due to an exogenous shock: whether the
trading day is Sunday (Sundays / holidays) or not. The exogenous shock to
the dealers ability to hedge against inventory risk, and therefore the DID
10
For discussion of clustered standard errors see Petersen (2009). For discussion of
mixed linear model see Hsiao (2003).
11
The normality assumption should not be a problem for our dependent variable since
we model the average spread.

estimate, captures the eect of inventory on the bid-ask spread.

Let us dene the notations. The time-weighted average percentagespread is denoted by yst , which is observed for ETN s on day t (s = 1; :::; 14,
t = 1; :::; 248).12 We estimate the model
yst =

ij

+ xst

ij

+ "st ;

where i takes the value of one (or U SIndexs = 1) if ETN s is on one of the US
indices, and zero if the ETN is on an Israeli index. Similarly, j takes the value
of one (or Sundayt = 1) if the European markets are open on date t (Sunday or holiday), and zero otherwise. The superscript ij identies four sets of
parameters, one for each combination of the binary variables U SIndexs and
Sundayt . We estimate four intercepts (
slope coe cients (

00 ;

01 ;

10 ;

11 ),

00 ;

01 ;

10 ;

11 )

and four vectors of

to allow for variation in the eects of the

explanatory variables on the spreads, for the Israeli and US ETNs, on Sundays and on other trading days. The vector of explanatory variables is the
same for all ETNs and regimes ij, xst = [LogV olumst ; Imbalancest ; D1st ; D2st ].
We denote by LogV olumst the log of daily trading volume for ETN s on
day t. Imbalancest is the absolute value of net buy-side market order ow
scaled by the total market order volume for for ETN s on day t. We also
use dealer / issuer xed eects (D1st and D2st ) and cluster the standard
errors by ETN and date.
Estimation results are presented in Table 3. Model 1 uses only four
intercepts, which are the conditional means of the spread for various categories: weekdays versus Sundays, and Israeli versus US ETNs.13 At the
12
We have spread data for 248 days, but the orders and transactions data is missing for
the week of January 22 - January 26. Therefore, we include 248 days in all analyses of the
spread, but only 243 day in the analyses involving trading volume and order imbalances.
13
Note that the specication of an intercept for every one of the four combinations
of U SIndexs , and Sundayt is standard for Generalized Linear Models (GLM), and it is

10

bottom of the table we present the estimates of dierences and dierencesin-dierences across various categories. First, the spreads of US ETNs are
always signicantly higher than those of Israeli ETNs, which is not surprising as the former involve higher processing and hedging costs. The dierence
between the spread on Israeli ETNs on weekdays and Sundays is not statistically signicant (0.113% versus 0.127%) as there are no dierences in the
inventory costs. The spread of US ETNs is more than doubled on Sundays
relative to other trading days (0.404% versus 0.943%) and the dierence
is highly statistically signicant. The DID estimate is large (0.526%) and
highly signicant. This implies that the inventory component of the bid-ask
spread is over 55% , which is higher than what was traditionally found in
the empirical studies.
Model 2 controls for other explanatory variables: trading volume, order
imbalances and dealer eects, which implies that the intercepts are no longer
simple conditional means. Yet, the implied inventory component still range
between 51% and 59% across dealers. The coe cients on the explanatory
variables are in line with our expectations: the trading volume has a negative
and signicant eect on the spread, while the eect of order imbalances is
positive but insignicant. We test the DID contrasts for those two variables
and show that trading volume and order imbalances have stronger eects
when traders cannot hedge inventory risk: the contrast of trading volume is
negative and statistically signicant while the order imbalances contrast is
positive but insignicant.
When presented with the predictions of theoretical models of inventory
equivalent to the specication yst = + 1 U SIndexs + 2 Sundayt + 3 U SIndexs
Sundayt +"st , in which 00 = , 01 = + 1 , 10 = + 2 and 11 = + 1 + 2 + 3 . We
choose to structure the equation that way to simplify the interpretation of the parameters,
interaction terms and DID tests.

11

control, dealers tend to dismiss them as an oversimplication (see Hasbrouck,

2007). We suspect that they do not constantly adjust the spread in response
to inventory changes since they are able to hedge those risks. When hedging
is not possible spreads become more sensitive to inventory changes and thus
to order imbalances.
The estimation approach used above is standard for panel data, but it
does allow for an explicit estimation of a more complex variance-covariance
structure. As the variance of spreads may be of interest to regulators and
other investors, we estimate a similar model of daily spreads using the mixed
linear eects approach with repeated observations. this way we account for
possible correlation and obtain estimates of the variance covariance structure
under various regimes. Specically, we estimate the model
yst =
in which the xed part
s

and

t,

ij

ij

+ xst

+ xst

ij

ij

+ "st ;

remains the same as before. We add

ETN and date random eects to control for correlated standard

errors instead of clustering. We also control for possible heteroskedasticity

by estimating four dierent variances, one for each combination of U SIndexs
and Sundayt .
The results presented in Table 3 are similar to those presented before
for the xed eects and the DID tests. To check whether heteroskedasticity
adds explanatory power to the model we compare the likelihood of our model
to that of restricted models, which assume either two variances (determined
by categories of U SIndexs ) or a single variance parameter. Both likelihood
ratio tests are signicant, indicating that the chosen structures contributes
to the explanatory power of our model.14 The heteroskedasticity result
14

Estimation results of the restricted models are not presented in the paper. They are
available upon request.

12

has an appealing intuitive explanation and it extends our understanding

of the eect of inventory risk. If a trader demands liquidity in US ETNs
and arrives to the market on Sunday, not only the expected spread is wide
relative to other trading days (0.952% versus 0.412%) but there is also more
uncertainty about the spread, as implied by the higher variance (0.570 versus
0.094). We expect to see higher price volatility for US ETNs on Sunday, not
only because the average spread is wide but because the range of spread
realizations is also wide.
Lastly, the variance result remains in Model 2, where we control for
trading volume and order imbalances, which rules them out as explanations
of this phenomena. We suspect that properties of the limit order book,
beyond the best bid and ask prices, are driving the heteroskedasticity result:
the distance to the next price level and depth. When we observe small
variances in cases such as Israeli ETNs on Sundays, we nd that the depth
at the best quote is high and the distance between the best and the next
quote is small. When we observe large variances in cases such as US ETNs
on Sundays, we nd that depth at the best quote is low and the distance
between the best and the next quote is large. This implies that an order
that will not have any eect on the spread for Israeli ETNs will increase the
spread for US ETNs, which leads to higher variance. Since we only model
the inside spread rather than the whole book, it is important to allow for
heteroskedasticity to control for such dierences.
The daily data clearly indicates that we cannot reject the causal relation between the exogenous shock to inventory cost and the spread that is
statistically and economically signicant. We now proceed to the intraday
analyses.

13

3.2

The Intraday Spread

We calculate the time weighted percentage bid-ask spread for every 90

minutes interval during the continuous trading phase, and obtain ve intraday spread observations rather than one daily average for every ETN
and trading day in our sample. To account for the additional complexity of
the data, we extend the specication of the mixed model described above
and estimate
ystk =

ijk

+ xstk

ij

+ "stk :

As before, i takes the value of one if ETN s is on one of the US indices,

and zero if the ETN is on an Israeli index; j takes the value of one if the
European markets are open on date t and zero otherwise. The new index
k (or the variable Daytime) takes the values of one through ve according
to the time interval
8
1 between 09:45 - 11:15
>
>
>
>
< 2 between 11:15 - 12:45
3 between 12:45 - 14:15 :
Daytime =
>
>
4 between 14:15 - 15:45
>
>
:
5 between 15:45 - 17:15
The superscript ij identies four sets of slope coe cients (

00 ;

01 ;

10 ;

11 ),

one for each combination of the binary variables U SIndexs and Sundayt .
As for the intercepts, we now use the superscript ijk to identify 20 parameters (

001 ; :::;

005 ;

011 ; :::;

115 ),

one for each combination of U SIndexs ,

Sundayt and Daytime (3-way interaction). The explanatory variables are

the same as before calculated for every 90 minutes interval rather than
on a daily basis. As before, we add

and

t,

ETN and date random

eects to control for correlated standard errors. We control for possible

heteroskedasticity by estimating 20 dierent variances, one for each combination of U SIndexs , Sundayt and Daytime, and four autocorrelations of
the intraday spreads, one for each combination of U SIndexs and Sundayt .
14

Results are presented in Table 5. First, we observe a clear U-shaped

intraday pattern for Israeli ETN spreads and an inverted J-shaped pattern
for US ETNs. Second, US ETN spreads are much wider, especially in the
rst hour of the day when the European futures markets are closed due
to time dierences between Europe and Israel. Third, as before, Sunday
spreads for Israeli ETNs are no dierent than spreads during the rest of the
week, but the spreads for US ETNs are much wider on Sundays. The DID
estimates, for the second through fth time intervals, range between 0.470%
and 0.619% and. In relative terms, it implies an inventory component of 60%
to 78%. In the rst time interval we get the lowest estimate (0.466% which
is 31% of the spread), but in this case the inventory eect is underestimated
since hedging is not available for US ETNs on the rst time interval of
any trading day. This is also supported by the positive DID estimates of
the rst and second time intervals, which are positive and signicant. This
implies that the expectation to be able to hedge in the next time interval
decreases the spreads for US ETNs on weekdays, but not to their levels in
the second time interval when hedging is actually possible. Again, these
ndings indicate that the eect of the inventory cost can be larger than
previously documented.
Controlling for dealer eects, trading volume and order imbalances during the time interval does not change our results. Both volume and imbalances have the correct signs and the eect of volume is even signicant, but
the variables dont contribute much to the explanatory power of the model.
In fact Model 2 implies an even higher inventory component than Model 1.
In Table 5, we extend the number of groups from four that we used in the
daily model to twenty: four for every time interval. We allow for intraday
variation in the variance of the spread and assume that the intraday spreads

15

may be autocorrelated (follow an ARH(1) process).15 Similar to the daily

model in Table 4, we nd that the inventory cost aects not only the average
spread, but also its volatility. First, the volatility of the spread for US ETNs
is higher than that of Israeli ETNs. Second, there is another dierence in
the intraday patterns between the local and US ETNs: while the volatility is
much higher towards the end of the day in the Israeli ETNs on all days, the
highest volatility for the US ETNs is during the rst time interval, and it is
much higher on Sundays. Third, the dierences in spread volatility between
Sundays and the rest of the week for the Israeli ETNs are negligible, while
they are very large for the US ETNs. As we showed for the daily spreads,
the increased cost of inventory control has a strong independent eect on
the variances of intraday spreads.

3.3

The Eect of Price Volatility

In this section we introduce one more exogenous shock to test the volatility
hypothesis. The model is similar to that described for average daily bid
ask spreads, but now we simultaneously estimate the intercepts for three
time periods: 2006, 2008 before Lehman Brothers led for bankruptcy (January, 1 - September, 15) and 2008 after Lehman Brothers les for bankruptcy. Estimation results are presented in Table 6. Model 1 uses only
twelve intercepts (four for every period), which are the conditional means
of the spread for various categories: weekdays versus Sundays, and Israeli
versus US ETNs. At the bottom of the table we present the estimates of
dierences-in-dierences (DID) and dierences-in-dierences-in-dierences
(DIDID) across categories.
15

ARH(1) is a heterogenous AR(1) process. In our case, we get 5 variance parameters

(one for every daytime category) and one autocorrelation parameter. Likelihood ratio
tests indicate that the ARH(1) specication signicantly increases the explanatory power
of the model.

16

First, it is clear that spreads for all ETNs on all trading days are relatively low in 2006 and they increase to the highest level in the second period
of 2008. This eect may be attributed to the high levels of uncertainty and
price volatility during the nancial crisis, but we are interested in the causal
eect of volatility on the cost of inventory and therefore on the spread rather
than a general relation between price volatility and spreads. We start with
the rst three DID tests, which are similar to those we presented in Tables
3 and 4 and imply a signicant inventory cost in every one of the three time
intervals. Yet, the test of a causal volatility eect comes from a comparison
of the inventory component over time. Since price volatility increases from
a relatively low level in 2006 to a higher level in the rst period of 2008
and its highest level in the second period of 2008, we expect the inventory
component to increase as well. Comparing the three inventory components,
we get a signicant increases as a result of the rise in price volatility (0.7410.526=0.215 and 1.129-0.741=0.388), which supports our fourth hypothesis.
In this test we establish a causal relation between the price volatility and the
cost of inventory, which leads to a statistically and economically signicant
increase in the bid ask spread.

Conclusions

In this paper we use the fact that Tel Aviv Stock Exchange is open on
Sundays to estimate the causal relation between the inventory cost and the
bid-ask spread. We compare spreads on the most liquid ETNs on the major Israeli and US indices, on Sunday and other days, using dierences-indierences approach. This allows us to avoid making structural assumptions
to estimate the inventory component of the spread. We show that on Sundays and other European holidays the bid ask spread in the US ETNs more
17

than doubles relative to the rest of the week, indicating that inventory cost
is an economically signicant transaction cost. The same eect is found
during the weekday morning hours when the Israeli market is open, but the
European markets are still closed - the inventory component can reach as
high as 80% of the spread during some intraday time intervals. No such
eect is observed in the Israeli indices. Finally, we show that the eect of
the trading volume on the spread is much more pronounced in foreign indices on Sunday than on other days or on the Israeli indices. The eect of
buying/selling pressure on the spread is positive but insignicant. The eect
of price volatility on the inventory component of the spread is economically
and statistically signicant.
This study shows that inventory control consideration can have large
eects on the trading costs. Market designers and regulators should make
sure that liquidity providers can hedge their inventory risk to reduce the
costs of liquidity provision.

References
[1] Amihud, Y., Mendelson, H., 1980. Dealership markets: market-making
with inventory. Journal of Financial Economics 8, 31-53.
[2] Bessembinder, H., Seguin, P., 1992. Futures Trading Activity and Stock
Price Volatility. The Journal of Finance 47, 2015-2034.
[3] Bloomeld, R., OHara, M., Saar, G., 2005. The make or takedecision
in an electronic market: evidence on the evolution of liquidity. Journal
of Financial Economics 75, 165-199.
[4] Bollen, N., Smith, T., Whaley, R., 2004. Modeling the bid/ask spread:
measuring the inventory-holding premium. Journal of Financial Economics 72, 97-141.
[5] Copeland, T., Galai, D., 1983. Information Eects on the Bid-Ask
Spread, Journal of Finance, 38, 1457-69.
[6] Chordia, T., Roll, R., Subrahmanyam, A., 2002. Order imbalance, liquidity, and market returns. Journal of Financial Economics 65, 111-130.
18

[7] Foucault T., Kadan, O., Kandel, E. 2005. Limit order book as a market
for liquidity. Review of Financial Studies 18, 1171-1218.
[8] George, T., Kaul, G., Nimalendran, M., 1991. Estimation of the bid-ask
spread and its components: a new approach. The Review of Financial
Studies 4, 623-656.
[9] Glosten, L., Harris, L., 1988. Estimating the components of the bid-ask
spread. Journal of Financial Economics 21, 123-142.
[10] Grossman, S. J., Miller, M. H., 1988. Liquidity and market structure.
The Journal of Finance 43, 617-633.
[11] Hansch, O., Naik, N., Viswanathan, S., 1998. Do inventories matter in
dealership markets: evidence from the London Stock Exchange. Journal
of Finance 53, 1623-1656.
[12] Hasbrouck, J., 1988. Trades, quotes, inventories and information. Journal of Financial Economics 22, 229-252.
[13] Hasbrouck, J., 2002. Stalking the e cient pricein market microstructure specications: an overview. Journal of Financial Markets 5, 329
339.
[14] Hasbrouck, J., 2007. Empirical Market Microstructure: The Institutions, Economics, and Econometrics of Securities Trading. Oxford University Press, US.
[15] Hasbrouck, J., Soanos, G., 1993. The trades of market makers: an
empirical analysis of NYSE specialists. The Journal of Finance 48, 15651593.
[16] Hsiao, C., 2003. Analysis of Panel Data. Cambrige University Press,
New York.
[17] Hendershott, T., Jones, C. M., Menkveld, A. J., 2009. Does algorithmic
trading improve liquidity? Unpublished working paper.
[18] Ho, T., Stoll, H., 1983. The dynamics of dealer markets under competition. Journal of Finance 38, 1053-1074.
[19] Huang, R. D., Stoll, H. R., 1997. The components of the bid-ask spread:
a general approach. Review of Financial Studies 10, 995-1034.
[20] Kaniel, R., Liu, H., 2006. So What Orders Do Informed Traders Use?
Journal of Business 79, 1867-1913.
[21] Madhavan, A., Richardson, M., Roomans, M., 1997. Why do security
prices change? A transaction-level analysis of NYSE stocks. Review of
Financial Studies 10, 1035-1064.
19

[22] Madhavan, A., Smidt, S., 1993. An analysis of changes in specialist

inventories and quotations. The Journal of Finance 48, 1595-1628.
[23] OHara, M., Oldeld, G. S., 1986. The microeconomics of market making. The Journal of Financial and Quantitative Analysis 21, 361-376.
[24] Petersen, O. P., 2009. Estimating standard errors in nance panel data
sets: comparing approaches. The Review of Financial Studies 22, 435480.
[25] Reiss. P. C., Werner, I. M., 1998. Does risk sharing motivate interdealer
trading? Journal of Finance 53, 16571703.
[26] Stein, J., 1987. Informational externalities and welfare-reducing speculation. Journal of Political Economy 95, 1123-1145.
[27] Stoll, H., 1978. The supply of dealer services in securities markets.
Journal of Finance 33, 1133-1151.
[28] Stoll, H. R., 1989, Inferring the components of the bid-ask spread: Theory and empirical tests. Journal of Finance 44, 115-134.

20

Table 1: Sample of ETNs

Summary statistics for the sample of 14 most liquid ETNs traded on TASE. The sample period is from January 1, 2006 to December
31, 2006 with the exception of the last two ETNs, which started trading on February 9 and February 2 respectively. We report the
issuer, index and index country, average bid-ask spread and the mean daily NIS volume for every ETN. There are 248 trading days in
2006 but transactions and orders data is missing for 5 trading days (January 22 to January 26). Any calculation involving trading and
order volume will exclude the observations for those days.

Avg. Price*

Avg. Bid-Ask

Daily Trading

Issuer (Dealer)

Index

Country

Trading Days

(NIS)**

Spread

Volume (NIS)**

Excellence

TA100

Israel

248

85.3

0.10%

8,783,466
21,121,451

Excellence

TA25

Israel

248

84.4

0.08%

Excellence

SP500

US

248

58.4

0.22%

3,857,287

Excellence

NASDQ100

US

248

73.1

0.21%

5,013,806

50.9

0.52%

618,422
1,152,598

Excellence

DJ30

US

248

Excellence

Russell2000

US

248

32.7

0.50%

CLAL Finance

TA100

Israel

248

8.6

0.18%

3,075,061

8.4

0.09%

10,999,631
596,337

CLAL Finance

Number of

TA25

Israel

248

CLAL Finance

SP500

US

248

58.4

0.74%

CLAL Finance

NASDQ100

US

248

16.2

0.44%

562,095

TACHLIT

TA100

Israel

248

85.9

0.16%

3,637,179
7,967,820

TACHLIT

TA25

Israel

248

8.4

0.09%

TACHLIT

SP500

US

224

58.3

0.67%

661,398

TACHLIT

NASDQ100

US

219

36.5

0.89%

611,715

The tick size is {0.001 if price < NIS 10; 0.01 if NIS 10 price< NIS 100; 0.1 if NIS 100 price< NIS 1000; 1 if price > NIS 1,000}

** In 2006 the exchange rate was about NIS 4.5 to $US 1.

21

Table 2: Descriptive Statistics

Descriptive statistics for the sample of 14 most liquid ETNs traded on TASE. The sample period is from January 1, 2006 to December
31, 2006. We present the median, mean and standard deviation for the spread, order imbalance measure, trading volume and order
volume. All distribution statistics are presented by Index country (Israel / US) and for two regimes (Sunday / Weekday).

Percentage Spread
Israel

# of Obs.

Median

Mean

STD

1,176

0.07%

0.11%

0.23%

Weekday
Sunday

312

0.07%

0.13%

0.25%

1,526

0.31%

0.40%

0.38%

405

0.68%

0.94%

0.82%

1,152

-0.102

-0.077

0.405

306

-0.074

-0.108

0.401

1,502

-0.026

-0.053

0.658

399

0.000

-0.032

0.597

Daily Trading Volume (1,000 NIS)

Israel
Weekday
1,152

US

Weekday
Sunday

Order Imbalance Measure

Israel
Weekday
Sunday
US

Weekday
Sunday

Sunday
US

Weekday
Sunday

Daily Order Volume (1,000 NIS)

Israel
Weekday
Sunday
US

Weekday
Sunday

5,298

9,794

13,378

306

4,549

7,270

9,577

1,502

564

1,993

3,898

399

123

392

623

1,152

6,922,351

9,693,085

11,099,233

306

6,433,447

9,577,979

12,083,183

1,502

537,151

1,010,417

1,277,976

399

5,985

12,694

25,103

Daily Limit Order Volume / Total Order Volume

Israel
Weekday
1,152
Sunday
US

Weekday
Sunday

99.71%

99.34%

1.44%

306

99.73%

99.32%

1.64%

1,502

99.56%

98.51%

3.93%

399

94.04%

86.94%

18.45%

0.12%

0.46%

1.24%

306

0.10%

0.49%

1.45%

1,502

0.15%

0.90%

3.19%

399

4.50%

11.23%

17.68%

49.56%

53.08%

30.35%

306

46.19%

50.95%

32.31%

1,502

60.60%

65.34%

118.04%

399

96.60%

95.88%

67.62%

Daily Trading Volume / Daily Order Volume

Israel
Weekday
1,152
Sunday
US

Weekday
Sunday

Daily Trading Volume / Market Order Volume

Israel
Weekday
1,152
Sunday
US

Weekday
Sunday

22

Table 3: Daily Bid-Ask Spreads I

Table 3 reports the results of a panel data regression model for the average daily bid-ask spread. The model is estimated using least
squares approach and the standard errors are clustered by ETN and date. Model 1 uses only four intercepts, which are the
conditional means of the spread for various categories: weekdays versus Sundays, and Israeli versus US ETNs. Model 2 uses four
intercepts and controls for other explanatory variables: trading volume, order imbalances and dealer effects. The section Hypothesis
Tests reports estimates and tests of the DID contrasts.
Model 1
Effects

Model 2

Notation

Estimate

P_Value

Estimate

P_Value

USindex = 0

Sunday = 0

00

0.113

<.01

0.361

<.01

USindex = 0

Sunday = 1

01

0.127

<.01

0.406

0.01

USindex = 1

Sunday = 0

10

0.404

<.01

0.776

0.03

USindex = 1

Sunday = 1

11

0.943

<.01

2.011

<.01

Sunday = 0

V
00

-0.017

0.02

01

-0.019

0.02

10

-0.037

0.12

USindex = 0
USindex = 0

Log Volume

Sunday = 1

USindex = 1

Log Volume

Sunday = 0

USindex = 1

Log Volume

Sunday = 1

USindex = 0
USindex = 0

Sunday = 1

Usindex = 1

Sunday = 0

V
11

-0.106

<.01

Imbalance

00

0.025

0.35

Imbalance

01

0.017

0.65

Imbalance

10

0.045

0.37

Imbalance

0.061

0.56

Log Volume

Sunday = 0

USindex = 1

Sunday = 1

11

USindex = 0

Dealer = D1

0.017

0.54

USindex = 0

Dealer = D2

0.010

0.61

Usindex = 1

Dealer = D1

0.024

0.81

USindex = 1

Dealer = D2

0.311

<.01

Hypothesis Tests

Estimate

P_Value

Estimate

P_Value

1. H0: (01 00) = 0

0.013

0.56

0.044

0.73

2. H0: (11 10) = 0

0.540

<.01

1.235

0.02

0.526

<.01

1.190

0.02

-0.067

0.05

0.024

0.83

3. H0: (11 10) (01 00) = 0

V

4. H0: (

11

10)

01

00)

=0

5. H0: ( 11 10) ( 01 00) = 0

Fit Statistics
2

Adjusted R

0.52

0.62

Number of Observations

3,419

3,359

23

Table 4: Daily Bid-Ask Spreads II

Table 4 reports the estimation results of as a mixed linear model of the average daily bid-ask spread. Model 1 uses only four
intercepts, which are the conditional means of the spread for various categories: weekdays versus Sundays, and Israeli versus US
ETNs. Model 2 uses four intercepts and controls for other explanatory variables: trading volume, order imbalances and dealer
effects. The section Hypothesis Tests reports estimates and test of the DID contrasts. The section Variance-Covariance reports the
random effects and estimates of the variances for four categories: weekdays versus Sundays, and Israeli versus US ETNs.
Model 1
Effects

Model 2

Notation

Estimate

P_Value

Estimate

P_Value

USindex = 0

Sunday = 0

00

0.113

0.07

0.112

0.26

USindex = 0

Sunday = 1

01

0.127

0.05

-0.089

0.61

USindex = 1

Sunday = 0

10

0.412

<.01

0.612

<.01

USindex = 1

Sunday = 1

11

0.952

<.01

1.905

<.01

USindex = 0

Sunday = 0

00

-0.002

0.78

01

0.012

0.25

10

-0.023

<.01

USindex = 0

Log Volume

Sunday = 1

USindex = 1

Log Volume

Sunday = 0

Log Volume

USindex = 1

Sunday = 1

Log Volume

11

-0.097

<.01

USindex = 0

Sunday = 0

Imbalance

00

-0.005

0.83

Imbalance

01

0.005

0.91

Imbalance

10

-0.013

0.61

Imbalance

0.018

0.85

USindex = 0

Sunday = 1

USindex = 1

Sunday = 0

USindex = 1

Sunday = 1

11

USindex = 0

Dealer = D1

0.041

0.49

USindex = 0

Dealer = D2

0.036

0.55

Usindex = 1

Dealer = D1

0.065

0.24

USindex = 1

Dealer = D2

0.384

<.01

Hypothesis Tests

Estimate

P_Value

Estimate

P_Value

1. H0: (01 00) = 0

0.013

0.58

-0.201

0.25

2. H0: (11 10) = 0

0.540

<.01

1.293

<.01

0.527

<.01

1.494

<.01

3. H0: (11 10) (01 00) = 0

V

4. H0: (

11

10)

01

00)

=0

5. H0: ( 11 10) ( 01 00) = 0

Variance-Covariance
ETN RE
Date RE
USindex = 0

Sunday = 0

USindex = 0

Sunday = 1

USindex = 1

Sunday = 0

USindex = 1

Sunday = 1

-0.087

<.01

0.020

0.86

Notation

Estimate

P_Value

Estimate

P_Value

2
ETN
2
Date

0.023

<.01

0.003

0.01

0.017

<.01

0.018

<.01

2
00
2
01
2
10
2
11

0.036

<.01

0.037

<.01

0.039

<.01

0.039

<.01

0.094

<.01

0.093

<.01

0.570

<.01

0.469

<.01

Fit Statistics
-2 Log Likelihood
Number of Observations

24

1,414.8

1,290.7

3,419

3,359

Table 5: Intraday Spreads

Table 5 reports the estimation results of as a mixed linear model of the average bid-ask spread calculated for five 90 minutes
intervals during the day. Model 1 uses 20 intercepts, which are the conditional means of the spread for various categories: time
interval (1-5), weekdays / Sundays and Israeli / US ETN. Model 2 uses 20 intercepts and controls for other explanatory variables:
trading volume, order imbalances and dealer effects. The section Hypothesis Tests reports estimates and test of the DID contrasts.
The section Variance-Covariance reports the random effects and estimates of the variances for 20 categories: time intervals (1-5),
weekdays / Sundays and Israeli / US ETN, as well as intraday autocorrelations.
Model 1
Effects

Model 2

Notation

Estimate

P_Value

Estimate

P_Value

0.138

<.01

0.146

<.01

USindex = 0

Sunday = 0

09:45 - 11:15

USindex = 0

Sunday = 0

11:15 - 12:45

001
002

0.060

0.03

0.066

0.01

12:45 - 14:15

003

0.064

0.02

0.069

0.01

0.110

<.01

0.116

<.01

USindex = 0

Sunday = 0

USindex = 0

Sunday = 0

14:15 - 15:45

004

USindex = 0

Sunday = 0

15:45 - 17:15

005

0.156

<.01

0.157

<.01

USindex = 0

Sunday = 1

09:45 - 11:15

0.131

<.01

0.127

<.01

USindex = 0

Sunday = 1

11:15 - 12:45

011
012

0.060

0.03

0.056

0.03

0.057

0.04

0.052

0.04

USindex = 0

Sunday = 1

12:45 - 14:15

013

USindex = 0

Sunday = 1

14:15 - 15:45

014

0.141

<.01

0.138

<.01

USindex = 0

Sunday = 1

15:45 - 17:15

015

0.192

<.01

0.175

<.01

USindex = 1

Sunday = 0

09:45 - 11:15

1.066

<.01

1.074

<.01

USindex = 1

Sunday = 0

11:15 - 12:45

101
102

0.183

<.01

0.183

<.01

USindex = 1

Sunday = 0

12:45 - 14:15

103

0.166

<.01

0.165

<.01

14:15 - 15:45

104

0.235

<.01

0.234

<.01

0.271

<.01

0.266

<.01

USindex = 1

Sunday = 0

USindex = 1

Sunday = 0

15:45 - 17:15

105

USindex = 1

Sunday = 1

09:45 - 11:15

1.525

<.01

1.650

<.01

USindex = 1

Sunday = 1

11:15 - 12:45

111
112

0.802

<.01

0.909

<.01

0.713

<.01

0.812

<.01

USindex = 1

Sunday = 1

12:45 - 14:15

113

USindex = 1

Sunday = 1

14:15 - 15:45

114

0.788

<.01

0.880

<.01

USindex = 1

Sunday = 1

15:45 - 17:15

115

0.777

<.01

0.871

<.01

USindex = 0

Sunday = 0

00

-0.001

0.01

01

-0.000

0.33

10

-0.002

<.01

USindex = 0
USindex = 1
USindex = 1
USindex = 0
USindex = 0
USindex = 1

Sunday = 1
Sunday = 0
Sunday = 1
Sunday = 0
Sunday = 1
Sunday = 0

Log Volume
Log Volume
Log Volume

V
11

-0.018

<.01

Imbalance

00

0.008

0.02

Imbalance

01

0.005

0.12

Imbalance

10

-0.000

0.99

Imbalance

0.019

0.63

Log Volume

USindex = 1

Sunday = 1

USindex = 0

Dealer = D1

0.006

0.86

USindex = 0

Dealer = D2

0.011

0.75

Usindex = 1

Dealer = D1

0.139

<.01

Dealer = D2

-0.051

0.10

USindex = 1

11

25

Table 5: Intraday Spreads (Continued)

Model 1
Hypothesis Tests

Model 2

Estimate

P_Value

Estimate

P_Value

1. H0: (111 101) (011 001) = 0

0.466

<.01

0.594

<.01

2. H0: (112 102) (012 002) = 0

0.619

<.01

0.736

<.01

3. H0: (113 103) (013 003) = 0

0.553

<.01

0.664

<.01

4. H0: (114 104) (014 004) = 0

0.522

<.01

0.623

<.01

5. H0: (115 105) (015 005) = 0

0.4670

<.01

0.587

<.01

<.01

0.142

0.03

6. H0: (1) (5) simultaneously

0.153

7. H0: [(112 102) (012 002)] - [(111 101) (011 001)] = 0

V

8. H0: (

11

10)

01

00)

0.02

=0

9. H0: ( 11 10) ( 01 00) = 0

Variance-Covariance
ETN RE
Date RE

0.03
-0.017

<.01

0.022

0.58

Notation

Estimate

P_Value

Estimate

P_Value

2
ETN
2
Date

0.004

<.01

0.001

0.01

0.001

<.01

0.001

<.01

0.010

<.01

0.010

<.01

0.002

<.01

0.002

<.01

0.043

<.01

0.044

<.01

0.243

<.01

0.248

<.01

USindex = 0

Sunday = 0

11:15 - 12:45

USindex = 0

Sunday = 0

12:45 - 14:15

USindex = 0

Sunday = 0

14:15 - 15:45

USindex = 0

Sunday = 0

15:45 - 17:15

2
001
2
002
2
003
2
004
2
005

0.135

<.01

0.129

<.01

0.202

<.01

0.202

<.01

0.008

<.01

0.008

<.01

0.001

<.01

0.001

<.01

0.001

<.01

0.001

<.01

0.379

<.01

0.386

<.01

0.124

<.01

0.097

<.01

USindex = 0

Sunday = 0

09:45 - 11:15

USindex = 0

Sunday = 0

ARH(1)

00

USindex = 0

Sunday = 1

09:45 - 11:15

USindex = 0

Sunday = 1

11:15 - 12:45

011
2
012

USindex = 0

Sunday = 1

12:45 - 14:15

USindex = 0

Sunday = 1

14:15 - 15:45

USindex = 0

Sunday = 1

15:45 - 17:15

2
013
2
014
2
015

USindex = 0

Sunday = 1

ARH(1)

10

0.490

<.01

0.494

<.01

USindex = 1

Sunday = 0

09:45 - 11:15

1.518

<.01

1.534

<.01

USindex = 1

Sunday = 0

11:15 - 12:45

101
2
102

0.035

<.01

0.035

<.01

12:45 - 14:15

0.011

<.01

0.011

<.01

0.459

<.01

0.464

<.01

0.118

<.01

0.110

<.01

USindex = 1

Sunday = 0

14:15 - 15:45

USindex = 1

Sunday = 0

15:45 - 17:15

2
103
2
104
2
105

USindex = 1

Sunday = 0

ARH(1)

10

0.343

<.01

0.345

<.01

USindex = 1

Sunday = 1

09:45 - 11:15

2.468

<.01

2.351

<.01

USindex = 1

Sunday = 1

11:15 - 12:45

111
2
112

0.662

<.01

0.606

<.01

0.435

<.01

0.392

<.01

0.952

<.01

0.909

<.01

0.692

<.01

0.631

<.01

0.722

<.01

0.694

<.01

USindex = 1

Sunday = 0

USindex = 1

Sunday = 1

12:45 - 14:15

USindex = 1

Sunday = 1

14:15 - 15:45

USindex = 1

Sunday = 1

15:45 - 17:15

2
113
2
114
2
115

ARH(1)

11

USindex = 1

Sunday = 1

Fit Statistics
-2 Log Likelihood

2,795.0

2,618.0

Number of Observations

16,699

16,401

26

Table 6: Volatility and the Daily Bid-Ask Spreads

Table 6 reports the estimation results of as a mixed linear model of the average daily bid-ask spread. Model 1 uses only twelve
intercepts, four intercepts for each one of the three time periods: the year of 2006, the first part of 2008 (before Lehman Brothers
filed for bankruptcy in September 15, 2008) and the second part of 2008 (after Lehman Brothers filed for bankruptcy). We estimate
the conditional means of the spread, in every time period, for various categories: weekdays versus Sundays, and Israeli versus US
ETNs. Model 2 uses twelve intercepts and controls for other explanatory variables: trading volume, order imbalances and dealer
effects. The section Hypothesis Tests reports estimates and test of the DID and DIDID contrasts. The section Variance-Covariance
reports the random effects and estimates of the variances for twelve categories: weekdays versus Sundays, and Israeli versus US
ETNs in every one of the three time periods described above.
Model 1
Effects

Model 2

Notation

Estimate

P_Value

Estimate

P_Value

USindex = 0

Sunday = 0

2006

006

0.113

0.07

0.089

0.43

USindex = 0

Sunday = 1

2006

016

0.127

0.05

-0.170

0.27

USindex = 1

Sunday = 0

2006

106

0.413

<.01

0.565

<.01

USindex = 1

Sunday = 1

2006

116

0.953

<.01

1.774

<.01

USindex = 0

Sunday = 0

2008 BLB

00B

0.231

<.01

0.208

0.02

USindex = 0

Sunday = 1

2008 BLB

01B

0.186

<.01

-0.117

0.41

USindex = 1

Sunday = 0

2008 BLB

10B

0.417

<.01

0.569

<.01

USindex = 1

Sunday = 1

2008 BLB

11B

1.113

<.01

1.837

<.01

USindex = 0

Sunday = 0

2008 ALB

00A

0.282

0.10

0.259

0.15

USindex = 0

Sunday = 1

2008 ALB

01A

0.294

0.09

-0.015

0.94

USindex = 1

Sunday = 0

2008 ALB

10A

0.707

<.01

0.857

<.01

1.847

<.01

2.606

<.01

00

-0.002

0.69

01

0.016

0.06

10

-0.016

<.01

USindex = 1

Sunday = 1

2008 ALB

11A

USindex = 0

Sunday = 0

Log Volume

USindex = 0

Sunday = 1

USindex = 1

Log Volume

Sunday = 0

Log Volume

V
V

USindex = 1

Sunday = 1

Log Volume

11

-0.080

<.01

USindex = 0

Sunday = 0

Imbalance

00

-0.047

0.14

Imbalance

01

-0.031

0.59

Imbalance

10

0.004

0.85

Imbalance

0.043

0.47

USindex = 0

Sunday = 1

USindex = 1

Sunday = 0

USindex = 1

Sunday = 1

11

USindex = 0

Dealer = D1

0.154

0.00

USindex = 0

Dealer = D2

0.014

0.78

Usindex = 1

Dealer = D1

0.264

<.01

USindex = 1

Dealer = D2

-0.066

0.16

Hypothesis Tests

Estimate

P_Value

Estimate

P_Value

1. H0: (116 106) (016 006) = 0

0.527

<.01

1.469

<.01

2. H0: (11B 10B) (01B 00B) = 0

0.741

<.01

1.593

<.01

3. H0: (11A 10A) (01A 00A) = 0

1.129

<.01

2.022

<.01

4. H0: [(11B 10B) (01B 00B)] [(116 106) (016 006)] = 0

0.214

<.01

0.125

0.04

5. H0: [(11A 10A) (01A 00A)] [(11B 10B) (01B 00B)] = 0

0.388

<.01

0.429

<.01

-0.082

<.01

0.023

0.80

4. H0: (

11

10)

(
I

01

00)

=0

5. H0: ( 11 10) ( 01 00) = 0

27

Table 6: Volatility and the Daily Bid-Ask Spreads (Continued)

Model 1
Variance-Covariance
ETN RE
ETN RE

2006

ETN RE

2008 ALB

2008 BLB

Date RE

2006

Date RE

2008 BLB

Date RE

2008 ALB

USindex = 0

Sunday = 0

2006

USindex = 0

Sunday = 1

2006

USindex = 1

Sunday = 0

2006

USindex = 1

Sunday = 1

2006

Notation

Estimate

P_Value

Estimate

P_Value

2
ETN,6
2
ETN,B
2
ETN,A

0.023

<.01

0.037

<.01

0.021

<.01

0.002

0.08

0.172

<.01

0.153

<.01

2
Date,6
2
Date,B
2
Date,A

0.017

<.01

0.018

<.01

0.046

<.01

0.046

<.01

0.018

<.01

0.018

<.01

2
006
2
016
2
106
2
116

0.036

<.01

0.037

<.01

0.039

<.01

0.039

<.01

0.094

<.01

0.093

<.01

0.570

<.01

0.466

<.01

0.115

<.01

0.115

<.01

0.051

<.01

0.050

<.01

0.182

<.01

0.183

<.01

0.432

<.01

0.411

<.01

2
00A
2
01A
2
10A
2
11A

0.026

<.01

0.026

<.01

0.026

<.01

0.026

<.01

0.509

<.01

0.517

<.01

0.964

<.01

0.891

<.01

USindex = 1

Sunday = 1

2008 BLB

2
00B
2
01B
2
10B
2
11B

USindex = 0

Sunday = 0

2008 ALB

USindex = 0

Sunday = 1

2008 ALB

USindex = 0

Sunday = 0

2008 BLB

USindex = 0

Sunday = 1

2008 BLB

USindex = 1

Sunday = 0

2008 BLB

USindex = 1

Sunday = 0

2008 ALB

USindex = 1

Sunday = 1

2008 ALB

Model 2

Fit Statistics
-2 Log Likelihood
Number of Observations

28

5,724.7

5,130.1

6,849

6,789