-
Price Clustering Discrepancies in Limit Order Flows
A B S T R A C T
This article provides evidence that price clustering can be
explained, in part, by asymmetry in sell and buy limit order flows.
Consistent with extant literature (e.g. Kraus and Stoll, 1972;
Holthausen, Leftwich and Mayers, 1987 & 1990; and Gemmill,
1996), we report that buy orders are relatively more informative
than sell orders. We report that, on average, limit sell orders
cluster on round increments significantly more than that of limit
buy orders. Empirical literature shows that the price impact of buy
trades exceeds that of sell trades. We provide evidence that
limit-buy orders are not always more informative than limit-sell
orders, as measured by price clustering, and the informativeness is
dependent upon the condition of the firm/market. Contrary to prior
transaction level research (Harris, 1991), we report that limit
order clustering does not significantly increase following extreme
market downturns. Conversely, we find that limit order clustering
either decreases, or remains unchanged, as prices decline.
1. Introduction
The purpose of this paper is to investigate how the frequency
with which security prices fall on
round increments, or price clustering, responds to the broader
market state. By using a comprehensive
intra-day limit order dataset that identifies liquidity supplier
buy and sell orders, we are able to study
clustering frequencies of limit-buy and -sell orders separately.
In aggregate, we find that executed limit-
sell orders cluster on five penny increments more frequently
than that of limit-buy orders. If some
quantity of these limit orders are submitted on round increments
in an effort to minimize cognitive
processing costs, the transactions occurring at these prices are
less representative of the security’s
fundamental value. Thus, asymmetry in price clustering between
executed limit-buy and –sell orders
will reduce the relative informativeness of sell orders.
Prior research shows that markets react differently to buy and
sell orders and that the price impact
of purchases exceeds that of sales (Kraus and Stoll, 1972;
Holthausen, Leftwich and Mayers, 1987 and
1990; and Gemmill, 1996). Saar (2001) provides a theoretical
justification for this impact differential
-
2
based on the way institutional portfolio managers optimally use
information given a constrained set
of allowable strategies. The aggregate asymmetry in clustering
frequency reported in this study
provides another plausible explanation for this price impact
disparity.
Even though asymmetrical clustering frequency may be a mechanism
through which the price
impact differential is generated, it is not obvious why
investors would expend less cognitive resources
on limit-sell orders in aggregate. Closer examination reveals
that the direction of the clustering
disparity is contingent on the ambient state of the market and
the performance of the firm’s equity.
We document that limit-buy orders cluster significantly more
during market and stock price declines,
while limit-sell orders cluster more during advances. Thus, buy
trades are not universally more
informative than sell trades. In both univariate and
multivariate analyses we find that limit-sell orders
cluster significantly more than limit-buy orders when either
firm or market prices are increasing. When
prices are declining, however, we find that the converse is
true.
In an environment where prices are steadily rising, information
relevant to a particular security is
more likely to inspire optimism about future cash flows. Thus,
the suppliers of liquidity, traders that
submit limit-sell orders in this scenario, are less likely to be
motivated by information. Harris (1991)
contends that traders use discrete price sets to lower the cost
of negotiating. Along this line of
reasoning, liquidity suppliers, without any superior
information, might rationally submit orders
clustered on round increments to increase the probability that
their orders execute. If the decision to
trade is not motivated by information, as in the case of
portfolio rebalancing, expending additional
cognitive resources to derive a more precise assessment of a
security’s value is suboptimal if doing so
would lower the potential of execution.
The foundational research on price clustering documents that
security prices, across various
financial markets, tend to land on certain number sets more
frequently than others (Wyckoff, 1963;
Osborne, 1962; Niederhoffer, 1965 and 1966; and Ball, Torous,
and Tschoegl, 1985). Empirical
research prior to decimalization demonstrates that stock prices
cluster most on whole integers,
followed by halves, then odd quarters, and lastly on odd
eighths. Major U.S. stock exchanges reduced
-
3
the minimum tick size to pennies in the early 2000s, and a
variety of papers, including Ikenberry and
Weston (2008), provide significant evidence that the reduction
in tick size from one sixteenth of a
dollar did not taper the clustering phenomenon. In the
post-decimalization regime, stocks should
trade on five penny increments 20% of the time in the absence of
price clustering. In agreement with
previous studies(e.g. Alexander and Peterson, 2007; Ikenberry
and Weston, 2008), we also show that
transaction prices fall on nickels and dimes significantly more
than the expected 20%.
Harris (1991) also provides evidence that the frequency of
prices clustering significantly increases
following an extreme market decline. By analyzing the difference
in transaction price clustering
between the week prior to the 1987 stock market crash and the
week of the crash, he finds that
clustering increases across all price levels. The author notes
that this finding is to some extent
troubling because a large body of theoretical and empirical work
has documented a positive relation
between underlying security prices and the frequency with which
those prices fall on discrete price
sets. According to this reasoning, a significant market decline
should place downward pressure on
prices leading to a decline in overall clustering
frequencies.
Contrary to Harris (1991), we find that limit-order clustering
decreases in response to both firm
and market price declines. Asymmetry in price clustering among
order flows expands with volatility.
We document a significant difference in clustering between buy
and sell limit orders during volatility
shifts, with limit-sell orders clustering more during price
advances and limit-buy orders clustering more
during price declines. These findings are both economically and
statistically significant.
This paper contributes to two important areas of the finance
literature: investor behavior and
market efficiency. We report that investors submitting buy- and
sell-limit orders have conflicting
preferences for discrete price sets, which is particularly true
during market/firm price movements.
Liquidity suppliers initiating buy-limit orders become
increasingly interested in finding equilibrium
values during price inclines, while those submitting sell-limit
orders are more willing to deal on discrete
price sets. With the roles reversing during firm and/or market
price declines. We also contribute to
prior literature by providing a meaningful explanation as to why
price clustering decreases during
-
4
market/firm declines and increases during advances. We find that
limit-sell orders play a significant
role in driving clustering frequency down during declining
markets, and up during advancing markets.
The rest of the paper is organized as follows. Section 2
discusses the data used throughout the
analysis. Section 3 reports the results from our empirical
examinations along with motivational
literature. Section 4 provides a summary and conclusions.
2. Data description
Throughout the trading day, the NASDAQ TotalView-ITCH data feed
provides a real-time view
of the order book for all NASDAQ market participants. A
tick-by-tick record of the order and trade
transaction history appearing on the NASDAQ TotalView-ITCH
dataset, subsequently referred to as
(ITCH), is made available by NASDAQ to researchers and analysts.
From January 3, 2012 to
December 31, 2012, we obtain from ITCH, intra-day trading
volume, order prices (buy and sell
stamped), and executed shares. The ITCH dataset consists only of
orders and trades occurring on the
NASDAQ exchange, however, the sample firms can be listed on any
exchange. Order messages are
attached to each order on the book. We extract all executed
orders by isolating message types 'E', 'C',
and 'P'. The message 'E' is sent on an order that is executed in
whole or in part at the initial display
price. The message 'C' is sent on an order that is executed in
whole or in part at a price different from
the initial display price. Lastly, message type 'P' is sent on
an order that matches between non-
displayable order types, i.e. "hidden liquidity". From the
Center for Research in Security Prices (CRSP)
we gather daily prices, volume, shares outstanding, and closing
bid and ask prices. Only common
stocks (i.e. share codes 10 & 11) whose share prices are
above $5 and less than $500, and exhibit a
minimum of two trades per day are retained in the sample. In
addition, we require our sample firms
to trade in at least 80% of the total number of trading days
within our one year sample period.1 We
apply these restrictions on the data to avoid infrequently
traded firms and noise introduced by the
bid/ask bounce. Our final sample consists of 1,243 NYSE/AMEX
listed stocks and 1,383 NASDAQ
listed stocks.
1 Ikenberry and Weston (2008) make similar data restrictions
when examining price clustering in U.S. stock prices
-
5
Cluster% is the variable of interest throughout the empirical
analysis. We follow previous studies
that examine limit-order price clustering (e.g. Niederhoffer,
1965; Ahn, Cai, and Cheung, 2005; Chiao
and Wang, 2009, etc.). We measure the weekly frequency with
which buy (sell) limit-order prices
execute on $0.05 increments2. For each firm, we summate the
number of times orders fall on an
increment of five cents, i.e. $.00, $.05, $.10, $.15, by week.
We then divide that total by the number
of limit orders executed within that same week time period. We
denote this ratio as Cluster%, which
captures the frequency with which limit orders fall on a
discrete price set.
3. Empirical results
The summary statistics for our sample of firms are reported in
Table 1. As expected, the
NYSE/AMEX listed stocks have a mean market value, $9.05 billion,
which is much larger than the
$3.24 billion average size for NASDAQ firms. The NYSE/AMEX
listed stocks also account for more
daily trading volume on the NASDAQ exchange than that of the
NASDAQ-listed firms. However,
the average NYSE/AMEX firm in our sample executes 277,680 shares
per day, or approximately 138
shares per trade, whereas the average NASDAQ listed stock
executes 197,105 shares per day, or
approximately 140 shares per trade. The NASDAQ-listed firms
appear to be slightly more volatile,
experience higher percentage bid-ask spreads, and have a much
lower average share price. The average
number of limit-sell orders executed daily does not seem to be
significantly different from that of the
number of limit-buy orders, for all firms in the sample.
3.1. Univariate results
We begin by partitioning the intra-day ITCH dataset into
executed limit-buy and limit-sell orders,
with the intent of investigating whether the order prices differ
significantly in the frequency with which
they fall on increments of $0.05. Chao and Wang (2009) make the
argument that limit-order data
reflects investors’ intentions more realistically than trade and
quote data. We therefore focus on limit
2 Alexander and Peterson (2007) examine price clustering in the
post-decimalization period (post 2001) and find that
prices fall on nickels and dimes more frequently than any other
price increment.
-
6
orders because it provides a unique setting in which examining
investor transaction behavior can be
more directly observed.
3.1.1. Asymmetry in price clustering between limit order
flows
Table 2 provides univariate evidence that mean clustering
percentages between limit-buy and
limit-sell orders significantly differ across several specified
price levels. We find that limit-sell orders
cluster consistently more than that of limit-buy orders. These
results are convincingly significant and
robust to exchange listings. In aggregate, limit-sell orders
falls on round numbers 0.87 percentage
points more frequently than that of limit-buy orders. These
results are not only statistically significant,
but also economically meaningful. If prices are assumed to be
uniformly distributed, the expected
clustering percentage on nickels and dimes in the post
decimalization regime is equal to 20%;
therefore, clustering in excess of 20% is considered unexpected.
The average firm in the sample
displays a weekly abnormal clustering percentage of 5.26%. This
implies that the 0.87 percentage
point difference in clustering between limit-buy and -sell
orders is no trivial amount.
In order to make more valid inferences on the difference in
clustering percentage between limit-
buy and limit-sell orders, we partition the data into price
levels. We find that price clustering exhibits
a u-shaped relation with underlying prices. Prices tend to
cluster substantially more among the most
highly priced securities, over $200. Many researchers argue that
discrete price sets are used to trade
more highly priced assets. Empirical research showing a positive
relation between clustering
percentage and nominal share prices is extensive (e.g.
Niederhoffer, 1965; Ikenberry and Weston,
2008; Alexander and Peterson, 2007). The purchase of a home is a
commonly used example in which
a highly priced commodity is generally transacted on discrete
sets, such as $1,000s or $5,000s. The
average unexpected clustering percentage among the most highly
priced firms is 13.16%, compared
to that of the lowest priced firms, 6.52%. This pattern is well
documented in prior literature (e.g.
Harris 1991); therefore, the more interesting part is what
happens in-between the extreme price levels.
Clustering percentage decreases as prices rise between $20 and
$200. This decrease is more difficult
to explain as theory implies that clustering is increasing with
the underlying security price.
-
7
Chan and Lakonishok (1993) make the argument that there are
several possible explanations as
to why one might dispose of stock, but the choice to buy a
particular stock is likely to carry positive
firm-specific news. There are several reasons for why an
investor may seek to sell stock, so the
cognitive choice to set an order on a desired round number may
not have a tremendous impact on
fundamental values. However, the decision to submit a limit-buy
order on a round increment might
have a larger rippling effect on price resolution. If buy trades
indeed result in greater price impact
(Saar, 2001), we would expect to witness similar findings within
the clustering of limit order prices.
That is, limit-buy orders should transact on finer prices that
are more indicative of the stock's intrinsic
value. Therefore, investors submitting limit-buy orders may be
more concerned with transacting on
finer prices and less concerned with limiting negotiation costs
(Harris, 1991). The results in Table 2
confirm this idea that the decision to submit a limit order on a
round increment is contingent upon
whether the limit order is sell- or buy-initiated.
Ball, Torous, and Tschoegl (1985) present the price resolution
hypothesis and propose that price
clustering is a result of investor uncertainty about the
underlying value of a security. Coupling the price
resolution hypothesis with the theoretical arguments of Saar
(2001), in which buy trades are shown to be
more informative than sell trades, we document that less
informative limit-sell orders are more apt to
be placed on round numbers, and therefore increasing the
percentage of price clustering. Conversely
we find limit-buy orders to be more informative as they cluster
significantly less, resulting in a more
accurate reflection of fundamental prices. In general, we
contend that investors submitting limit-sell
orders are more uncertain about fundamental values and thus more
likely to place orders on round
increments.
Observable in Table 2, we pose that the decrease in clustering
is a result of differing limit-order
flows within the specified price levels. We find a monotonic
decrease in the difference in abnormal
clustering percentage between limit-buy and -sell orders as the
nominal security price is increasing.
The difference between limit-buy and -sell orders is 1.05% for
orders submitted at $10 or less, and
insignificantly different from zero for orders placed at a price
above $200. In unreported results, the
-
8
amount of limit-order volume in the quintile below $10 far
exceeds that of the volume exhibited for
orders in the quintile above $200.
We continue our examination of buy- and sell-limit order
differences by separating the data by
exchange listing following empirical research showing that
prices cluster more among NASDAQ-
listed firms (Christie and Schultz, 1994; Christie, Harris, and
Schultz, 1994; Bessembinder, 1999). The
results in panels B and C display clustering percentages for
stocks segmented by exchange listing. We
group NYSE and AMEX firms together and compare those with stocks
listed on the NASDAQ
exchange. We find strong univariate evidence that NASDAQ-listed
stock order prices cluster more
frequently than NYSE/AMEX stock prices. In particular, the
average firm listed on the NASDAQ
exhibits a mean clustering percentage of 26.73%, compared to
23.70% for NYSE/AMEX stocks.
In addition, the difference between limit-buy and -sell order
clustering is more pronounce among
NASDAQ-listed firms. For the NASDAQ sample, limit-sell orders
cluster on $0.05 increments 1.23%
more frequently than limit-buy orders, on average. For the
NYSE/AMEX sample, limit-sell orders
only cluster 0.48% more often than limit-buy orders. The
observed differences between limit-buy and
-sell orders across both exchanges are statistically significant
at the 1% level. Again, we interpret these
results as evidence that limit-buy orders are relatively more
informative than limit-sell orders, which
is particularly true among NASDAQ-listed stocks. NASDAQ stocks
are generally smaller and more
difficult to value, which may explain why the observed
difference in clustering frequencies between
limit-buy and -sell orders are more pronounce.
3.1.2. Price clustering in altering price conditions
Our next set of tests examine clustering differences in
limit-order flows during periods of high
price volatility. Harris (1991) shows evidence that transaction
level price clustering significantly
increases following the stock market crash of 1987. However,
bearish markets lead to downward price
pressure which, as discussed previously, should reduce overall
price clustering. We examine whether
price clustering exhibits a significant increase during periods
of extreme price volatility. According to
our theoretical understanding of price clustering, we would
expect to see a decrease in clustering as
-
9
prices are falling, and conversely an increase in clustering as
prices are rising. As a measure of
completeness, we categorize advancing and declining periods
using three separate methods.
The first approach is to examine the differences between extreme
market advances and declines.
We define a market as declining (advancing) when the NASDAQ
index experiences a sizeable decrease
(increase). Again, we are concerned with the NASDAQ exchange
only because the ITCH database
consists of firms that trade solely on the NASDAQ. Within our
sample time frame, we identify
periods when the NASDAQ dropped by 11.93% (April 2, 2012-June
1,2012) and rose by 17.78%
(January 3, 2012-April 1, 2012). Secondly, we measure declining
versus advancing markets as weekly
fluctuations in average returns on the S&P 500. Therefore we
denote a declining period as a week in
which the average return on the S&P 500 is negative, and an
advancing week when the average return
is positive. Our final method for approximating price volatility
is measured on the firm level. We
define a declining period as a week in which a stock experiences
negative average raw returns, and an
advancing period in which a stock exhibits positive average raw
returns. Together, we can disentangle
what level of price volatility is driving clustering
fluctuations, if any.
Panel A of Table 3 presents the univariate results for price
clustering around extreme market
movements. We find that the average limit-order clusters
significantly more, 0.49 percentage points ,
during the extreme advancing period, relative to the extreme
declining period. This difference is
significant at the 1% level. We document that this finding does
not seem to be isolated to stocks listed
on a particular exchange. The difference in price clustering
between limit orders is not significant
when the measure for market movements is the average weekly
return on the S&P 500. Panel C
reports results similar to that of Panel A, in that limit-order
price clustering is higher during firm-level
advancing periods, but only marginally. These results provide
evidence that when price clustering
behavior differs among market conditions, it is in the direction
theory predicts (Ball, Torous, and
Tschoegl; 1985). Contrary to the findings of Harris (1991), this
study provides evidence that limit-
order clustering is significantly greater during advancing
periods as opposed to declining periods. As
security prices increase, i.e. advancing period, limit-order
clustering also increases. The converse is
also true with regards to declining periods.
-
10
Baker and Wurgler (2007) and Delong et al. (1990) argue that
investors are subject to general
belief about future cash flows and risks not justifiable by
available information. Therefore, investors
can drive prices to unjustifiable levels by their inability to
determine the informativeness of a limit
order. We contend that price clustering distracts market
participants from finding “true” equilibrium
values. Thus during price advances and declines, limit orders
submitted on discrete sets continue to
drive prices away from fundamental values. Limit orders may also
act as price barriers, for example,
hindering additional upward movement. (Ohta, 2006; Chiao and
Wang, 2009).
3.1.3. Asymmetry in price clustering between limit order flows
within altering price
conditions
Investors setting limit orders on discrete price sets well
outside the bid-ask spread, may be hopeful
that when the market turns in their desired direction, their
submitted orders will have the greatest
probability of being executed. Cooney, Van Ness and Van Ness
(2003) show that execution rates of
even-priced limit orders exceed those of odd-priced limit
orders. The authors make the argument that
investors setting limit orders may prefer to use round prices to
increase the probability of their order
being executed. Chiao and Wang (2009) provide evidence that
limit orders submitted by individual
investors tend to cluster on integer and even prices more often
than those of other investor classes.
Therefore, individual investors may be arbitrarily setting limit
orders on round increments around the
best bid and best offer in hopes of capitalizing on sudden price
changes.
3.1.4. Limit-sell orders
Similar to Tables 1 and 2, we are interested in the
discrepancies in price clustering effects among
limit-order flows. We therefore separate the data into limit-buy
versus limit-sell orders. The results
in Table 3 suggest that limit-sell orders cluster on round
increments significantly more during
advancing markets than that of limit-buy orders. During an
extreme market advance, limit-sell orders
cluster 26.07% of the total weekly limit-sell orders executed,
compared to 25.02% for limit-buy orders.
This difference is significant at the 1% level3. We find similar
patterns when observing the clustering
3 We conduct a simple t-test comparing means between limit-buy
and limit-sell orders. We find that the difference
1.05% is statistically significant at the 1% level.
-
11
differences under the alternative definitions of price advances.
Limit-sell orders cluster 1.37% more
frequently than buy orders during weeks in which the return on
the S&P 500 is positive4. In addition,
limit-sell orders cluster 1.38% more often than that of
limit-buy orders during weeks in which
individual firm returns are positive5. We conclude, from a
univariate standpoint, that limit-sell orders
cluster considerably more during advancing periods than
limit-buy orders.
We also report that the clustering percentage of limit-sell
orders experiences a significant decrease
when moving from an advancing period to a declining period. In
Panel A we find that limit-sell orders
cluster 1.26% more frequently during extreme market advances
relative to extreme market declines.
In Panel B we report that the clustering percentage for
limit-sell orders during weeks in which the
S&P 500 raw returns are positive is 0.65% greater than the
clustering frequency when the S&P 500
raw returns are negative. Similarly, Panel C reports that the
difference between sell-limit order
clustering frequencies between firm-level price inclines and
declines is 0.70%. We evaluate the
difference in means using sample t-tests and find that each of
the reported differences are significant
at the 1% level. As before, we partition the data by exchange
listings and report qualitatively similar
results. Therefore, limit-sell orders appear to cluster
substantially more during advancing periods
relative to declining periods.
We pose that limit-sell orders that are sitting on the
limit-order book above the best ask will begin
to transact as prices rise. Investors setting limit-sell orders
on round prices well above the asking price,
are causing limit-order clustering frequencies to increase as
prices advance. Therefore price clustering
during market increases may be explained, in part, by liquidity
suppliers’ sell limit orders. In addition,
if there is a plethora of closely priced sell limit orders
waiting above the ask price, advances will be
decelerated when moving through order barriers (Ohta, 2006;
Chiao and Wang, 2009). Therefore,
investors’ tendencies to set orders on round increments may make
it increasingly more difficult to
evaluate fundamentals, in an already uncertain environment.
Harris (1991) expounds upon the price
resolution hypothesis and states that price clustering depends
on how well known is the intrinsic value of
4 A simple t-test shows that the difference in means is
significant at the 1% level. 5As done previously, we conduct a
t-test showing that this difference is significant at the 1%
level.
-
12
the security. If the value is not well know, then the frequency
with which prices cluster will increase.
Thus limit-order clustering can prolong market uncertainty about
fundamental values.
Limit-sell orders that are set above the asking price transact
more frequently when prices are on
an upswing. When the market is bullish and prices are on the
rise, investors may view it as an optimal
time to sell due to theuncertainty about future prices. One
feasible explanation can be attributed to
loss aversion. This well documented economic decision theory
assumes that losses and disadvantages
have a greater impact on preferences than gains and advantages
(Tversky and Kahneman, 1991).
Therefore during advancing periods, investors submitting
limit-sell orders may be more concerned
with the possibility of the market turning and losing profits
than the possibility of prices continuing
in their upward fashion. Consistent with Cooney, Van Ness, and
Van Ness (2003) investors may be
submitting limit-sell orders on round numbers more frequently
during price advances in hopes of
increasing the probability of their orders being executed. We
therefore contend that limit-sell orders
set above the asking price, and on a discrete price set, will
execute at an increasing rate.
3.1.5. Limit-buy orders
In Table 3 we find that the clustering percentage differences
between limit-buy orders and limit-
sell orders within each of the defined declining periods are
insignificantly different. In other words,
limit-buy orders do not appear to cluster more frequently when
prices are declining relative to limit-
sell orders. However, limit-buy orders do appear to cluster more
frequently during declining periods
compared to advancing periods. These differences are
statistically significant at the 1% level and
robust to each of the defined periods. Empirical research shows
that the price impact of buy trades
exceeds that of sell trades (e.g. Kraus and Stoll, 1972;
Holthausen, Leftwich and Mayers, 1987 & 1990;
Gemmill, 1996, etc.). We contribute to this strand of research
by showing that the informativeness
of the limit-buy order is dependent upon the condition of the
market and stock. An increase in price
clustering is argued to be a result of price uncertainty
(Alexander and Peterson, 2007). Therefore the
increase in limit-buy order clustering during declining periods
is evidence that buy orders may be less
informative when prices are decreasing.
-
13
We propose that limit-buy orders that are sitting on the book
well below the best bid will execute
only when prices begin to decline. Liquidity suppliers setting
limit-buy orders well below the bid price
have the intention of buying stock when they believe prices are
at their lowest. Investors may be
arbitrarily setting limit orders on discrete sets under the bid
price as a mechanism to take advantage
of temporarily undervalued stocks. An investor looking to buy
stock, whether correctly or incorrectly,
expects that the price of the security will increase following
the purchase.
During a market downturn, investors are aware that any
particular stock price will be deflated,
but whether that be temporary or permanent is unknown.
Therefore, an investor looking to buy a
security may be less concerned about finding the equilibrium
price because he/she assumes that the
stock is already underpriced. This could cause prices to deviate
even further from fundamental
values. Inaccurate information may be imbedded into stock
prices, from a swarm of limit-buy orders
simply being executed because they were set on discrete price
sets before the market began to decline.
Within a univariate setting, we report that the difference in
price clustering among order flows is
primarily driven by the rise in stock prices. Although limit-buy
orders seem to demonstrategreater
price clustering frequency during declining periods, the
magnitude is fairly trivial in most cases.
3.2. Multivariate Analysis
We recognize the need to control for variables that might
influence limit-order clustering when
attempting to determine whether differences exist among order
flows and whether those differences
are affected by market conditions. We begin by estimating the
following equation using weekly pooled
cross-sectional limit order data.
𝐶𝑙𝑢𝑠𝑡𝑒𝑟%𝑖,𝑡 = 𝛼𝑖 + 𝛿𝑡 + 𝛽1𝑆𝑖𝑧𝑒𝑖,𝑡 + 𝛽2𝑃𝑟𝑖𝑐𝑒𝑖,𝑡 + 𝛽31/√𝑁𝑇𝑖,𝑡
+ 𝛽4𝑆𝑝𝑟𝑒𝑎𝑑%𝑖,𝑡+𝛽5𝑇𝑟𝑎𝑑𝑒𝑆𝑖𝑧𝑒𝑖,𝑡 + 𝛽6𝑆𝑒𝑙𝑙𝑖,𝑡 + 𝜀𝑖,𝑡 1
The dependent variable, Cluster%, is the frequency with which
limit-orders fall on increments of
$0.05. We follow several papers that examine price clustering in
a multivariate framework and include
the following independent variableswhich are mainly motivated by
the price resolution hypothesis (e.g.
Harris, 1991; Ikenberry and Weston, 2008; Alexander and
Peterson, 2007). Size, is the market
-
14
capitalization or closing price times number of shares
outstanding. Size should be negatively related
to price clustering because more information is available for
large firms (more analyst coverage) and
large firms are better diversified. Price is the daily closing
price obtained from CRSP. Price is included
because, as reported in Table 2, clustering is increasing with
stock price. 1/√𝑁𝑇is the inverse of the
square root of the total number of transactions for each firm by
week. We expect clustering frequency
to decrease as transaction activity increases. Therefore we
should find a positive relation between our
inverse transaction measure and clustering frequency.
Two other measures used as explanatory variables are the bid-ask
spread and trade size. Spread%
is calculated as the difference in closing bid-ask prices
divided by the midpoint6. In accordance with
the price resolution hypothesis, as the spread increases
(widens) price uncertainty also increases. We
therefore expect to see a positive relation between the bid-ask
spread and limit-order clustering.
TradeSize is the average trade size, and is constructed by
dividing the number of executed shares by
the total number of transactions. Trade size should be
positively related to clustering frequency
because large size trades are generally more informed.Finally,
an indicator variable that captures
whether a limit order is buy or sell is also included as a
regressor. We include both firm fixed effects
and time fixed effects in each model specification.
Column 1 of Table 4 reports the results from estimating equation
(1). We find that limit-order
clustering is increasing in price, trade size, and spread and
decreasing with transaction frequency.
These results are consistent with previous findings (e.g.
Ikenberry and Weston, 2007) and support the
price resolution hypothesis. We therefore focus our efforts on
the independent variable, Sell, which displays
a significant positive relation with the dependent variable,
clustering percentage. Therefore limit-order
clustering is significantly affected by the type of order flow,
whether it be buy or sell, holding all other
factors constant. This supports our univariate findings that
limit-sell orders cluster more frequently
than that of limit-buy orders. Because the indicator variable
Sell takes on the value of 1 for a limit-sell
order and 0 for a limit-buy order. This suggests, holding all
else constant, a one unit increase in limit-
6 This simple measure of spread has been shown to be a reliable
approximation for microstructure spreads (Chung
and Zhang, 2013; Roll and Subrahmanyam, 2010). We replicate the
analysis using Corwin and Schultz (2012) simple spread
approximation and find qualitatively similar results.
-
15
buy orders leads to a 0.87 percentage point decrease in price
clustering. To the extent that price
clustering is negatively related to price resolution (Ball,
Torous, and Tschoegl, 1985), this coincides
with the theory that buy orders carry more information about
fundamental values.
Our next set of tests examine the affects of price volatility on
limit-order clustering. We expand
equation (1) to incorporate such effects. Therefore we estimate
the following equation:
𝐶𝑙𝑢𝑠𝑡𝑒𝑟%𝑖,𝑡 = 𝛼𝑖 + 𝛿𝑡 + 𝛽1𝑆𝑖𝑧𝑒𝑖,𝑡 + 𝛽2𝑃𝑟𝑖𝑐𝑒𝑖,𝑡 + 𝛽31/√𝑁𝑇𝑖,𝑡
+ 𝛽4𝑆𝑝𝑟𝑒𝑎𝑑%𝑖,𝑡+𝛽5𝑇𝑟𝑎𝑑𝑒𝑆𝑖𝑧𝑒𝑖,𝑡 + 𝛽6𝑆𝑒𝑙𝑙𝑖,𝑡 + 𝛽7𝐹𝑖𝑟𝑚𝐷𝑒𝑐𝑖,𝑡
+ 𝛽8𝑀𝑘𝑡𝐷𝑒𝑐𝑖,𝑡 + 𝛽9𝐸𝑥𝐷𝑒𝑐𝑖,𝑡 + 𝛽10𝑆𝑒𝑙𝑙 × 𝐷𝑒𝑐𝑖,𝑗,𝑡 + 𝜀𝑖,𝑗,𝑡
2
We construct indicator variables which take on the value of 1
for each of the three-defined
declining periods. FirmDecline captures firm-level price
movements. MarketDecline measures average
weekly market movements (estimating using S&P 500 weekly
returns). ExtremeDecline indicates
extreme unprecedented market movements discussed in detail
previously. We also include interaction
terms, Sell*Declineij, which captures the effects of each
specified market condition and sell orders on
clustering frequency. Within each regression we include both
firm- and time- fixed effects.
The results in Columns 2, 3, and 4 report evidence that the
frequency of limit-order clustering
decreases in declining markets. This is perhaps surprising given
the evidence provided by Harris
(1991), in which trade price clustering increases sharply
following a significant price decline. We do
not observe a significant increase in limit-order clustering
following a severe economic downturn. To
the contrary, we find that limit-order clustering is
significantly decreasing during both firm-declines
and severe market declines. The coefficients on FirmDecline and
ExtremeDecline are statistically
significant, and negative; however the economic effect may be
trivial. Our results therefore
consistently indicate that clustering is either negatively
affected by price declines or unaffected. We
fail to find any evidence of clustering increases following
price declines when examining limit orders.
Most theoretical and empirical models of price clustering
however do predict a positive relation
between security prices and the frequency with which they
cluster (Niederhoffer and Osborne, 1966;
Ball, Torous, and Tschoegl, 1985; Harris, 1991; Alexander and
Peterson, 2007).
-
16
The results in Table 4, columns 5, 6, and 7, also provide
evidence that there exists a significant
difference in limit-order clustering among order flows, which is
amplified during price movements.
We find that the interaction terms between limit-sell orders and
the dummy variables for price declines
are significantly negative. The results are qualitatively
similar regardless of the measure of
declining/advancing periods. Not only are the coefficients
statistically significant but they are also
economically meaningful. We interpret these results as evidence
that a significant difference exists
between price clustering among limit-order flows, which in turn,
appears highly correlated to market
and firm conditions. Limit-sell orders cluster substantially
more during advancing markets, while
limit-buy order cluster on round increments more during
declining markets, whether it be evaluated
at the firm-level or market level These results suggest that
limit-sell orders are less informative than
limit-buy orders during price advances, as clustering frequency
is increasing Therefore, limit-buy
orders appear more indicative of fundamental values during price
upswings. The converse is also true
during price declines, in that limit-buy orders cluster more
frequently than limit-sell orders and can
therefore be considered less informative about fundamental
values.
We run post-hoc tests on the findings reported in Table 5 by
partitioning the data into limit-buy
and -sell orders. The fact that the interaction terms between
sell orders and declining periods are
significant, implies that further investigation is needed. We
therefore estimate equation (2) using
separate samples for limit-buy orders and limit-sell orders.
Again, we control for both firm- and time-
fixed effects.
The first three columns examine the effects of pricing
conditions on limit-sell order clustering.
We find that sell limit orders cluster significantly less during
declining periods and, inversely, more
during advancing periods. Column 1 shows that clustering among
sell orders is decreasing as firm-
level prices are increasing. Similarly, Columns 2 and 3 report
that limit-sell orders cluster substantially
less as market prices are on the rise.
The final three columns estimate equation (2) using the sample
of limit-buy orders. We report
that the effects of price movements on limit-buy order
clustering is an almost mirror image of that of
-
17
sell-order clustering. We find that limit-buyorders cluster
significantly more when prices are declining.
This suggests that limit-buy orders may not always be more
informative than limit-sell orders, when
taking into account price conditions. To the extent that price
clustering is inversely related to price
discovery, limit-buy orders appear to become less informative as
prices are decreasing, both aggregate
market prices and firm-level prices. To reiterate, these results
hint that limit-buy orders may not always
be more informative than limit-sell orders. This interpretation
of order flow needs to be done with
caution, as significant differences exist among altering price
conditions.
As a measure of robustness, we partition the data by exchange
listing (i.e. NYSE/AMEX and
NASDAQ) and re-evaluate the main results of this paper. Prior
literature documents a significant
difference in price clustering between firms listed on the
NASDAQ versus those listed on the
NYSE/AMEX (Christie and Schultz, 1994; Christie, Harris, and
Schultz, 1994; Bessembinder, 1999;
and others). We find that limit-buy orders and limit-sell orders
exhibit similar clustering behavior
across exchanges. The frequency of limit order clustering
increases during declining markets and falls
during advancing markets. As before, limit-sell orders cluster
more often than limit-buy orders,
however that relationship is conditioned upon the state of the
market.
4. Concluding remarks
Extant empirical research shows that markets react differently
to buy and sell orders and that the
price impact of purchases exceeds that of sales (e.g. Kraus and
Stoll, 1972; Holthausen, Leftwich and
Mayers, 1987 & 1990; and Gemmill, 1996) We seek to provide
evidence that limit order flows can
explain a significant portion of the observed price clustering
in equity markets. If some quantity of
these limit orders are submitted on round increments in an
effort to minimize cognitive processing
costs, the transactions occurring at these prices are less
representative of the security’s fundamental
value. Thus, asymmetry in price clustering between executed
limit-buy and –sell orders will reduce
the relative informativeness of sell orders.
Consistent with our expectations, we find that there exists a
significant difference in the frequency
with which buy and sell limit orders cluster on increments of
$0.05. From our multivariate analysis,
-
18
we find that buy/sell limit orders can explain a significant
portion of the observed price clustering
within our sample, after controlling for factors that have been
hypothesized to explain the variability
in clustering frequency. We report that limit-sell orders
cluster significantly more, on average, than
limit-buy orders. However, this result is conditional upon the
state of the market and the firm. Prior
literature has found that price clustering is negatively related
to price discovery.
To the extent that clustering is negatively related to price
resolution, we interpret the initial results
as limit-buy orders being more informative than that of
limit-sell ordersbecause of the observed lower
clustering frequency. However, we find that limit-buy orders do
not always appear more informative
than limit-sell orders. For example, as firm-level prices,
and/or market prices, are falling, limit-buy
orders cluster significantly more than that of limit-sell
orders. The reverse is also true during price
advances, with limit-sell orders clustering more.
We provide several explanations as to why limit-sell orders
cluster more during bullish
market/firm price conditions and limit-buy orders cluster more
during bearish market/firm price
conditions. For instance, we contend that limit-buy (sell)
orders sitting on the limit order book well
below (above) the best bid (ask) will exhibit an increase in
price clustering as prices decline (advance).
These limit-orders may have been set on round numbers
strategically to increase the probability of
execution (Cooney, Van Ness and Van Ness, 2003) or less
strategically to reduce cognitive processing
costs (e.g. Wyckoff, 1963; Niederhoffer 1965, 1966). Whether it
be the former or the latter, we witness
significant differences in price clustering for limit-order
flows during market/firm price shifts
(advances and declines).
We contribute to previous literature by showing that limit-order
clustering is clearly dependent
upon the type of order that is being submitted, whether it be a
buy or sell. If increased levels of
clustering are associated with lower price resolution, we
provide evidence that limit-buy orders are not
always more informative than limit-sell orders, and the
informativeness can be explained in part by
the overall state of the market. In other words, the frequency
with which limit-buy and -sell orders
cluster is conditional upon the state of the market and firm.
Limit-buy orders cluster relatively more
-
19
during price declines and limit-sell order cluster relatively
more during price advances. In addition,
we do not find a significant increase in limit-order clustering
following extreme market declines. In
fact, we witness a slight decrease in the frequency with which
limit order prices fall on round
increments. .
5. References
Ahn, H. J., Cai, J., & Cheung, Y. L. (2005). Price
clustering on the limit-order book: Evidence from
the Stock Exchange of Hong Kong. Journal of financial Markets,
8(4), 421-451.
Alexander, G.J. and M.A. Peterson, 2007, An Analysis of
Trade-Size Clustering and its Relation to
Stealth Trading. Journal of Financial Economics, 84,
435-471.
Baker, M., & Wurgler, J. (2007). Investor sentiment in the
stock market. Journal of
EconomicPerspectives, 21(2): 129-151.
Ball, C.A., W.N. Torous, and A.E. Tschoegl, 1985, The Degree of
Price Resolution: The Case of the
Gold Market. Journal of Futures Markets, 5, 29-43.
Bessembinder H., 1999, Issues in assessing trade execution
costs. Journal of Financial Markets, 6,233-
257.
Chan, L. K., & Lakonishok, J. (1993). Institutional trades
and intraday stock price behavior.Journal
ofFinancial Economics, 33(2), 173-199.
Chiao, C., & Wang, Z. M. (2009). Price Clustering: Evidence
Using Comprehensive Limit‐
OrderData. Financial Review, 44(1), 1-29.
Christie W.G. and P.H. Schultz, 1994, Why do NASDAQ Market
Makers Avoid Odd-Eight Quotes?.
The Journal of Finance, 49, 1813-1840.
Christie, W.G., J.H., Harris, and P.H. Schultz, 1994, Why Did
NASDAQ Market Makers
StopAvoiding Odd-Eighth Quotes?. The Journal of Finance, 49,
1841-1860.
-
20
Cooney Jr, J. W., Van Ness, B., & Van Ness, R. (2003). Do
investors prefer even-eighth
prices?Evidence from NYSE limit orders. Journal of banking &
finance, 27(4), 719-748.
Delong, J. B., Shleifer, A., Summers, L, and Waldmann, R.
(1990). Noise Trader Risk in
FinancialMarkets." Journal of Political Economy, 98(4):
703-738.
Gemmill, G. (1996). Transparency and liquidity: A study of block
trades on the London
StockExchange under different publication rules. The Journal of
Finance, 51(5), 1765-1790.
Harris, L., 1991, Stock Price Clustering and Discreteness.
Review of Financial Studies, 4, 389-415.
Harris, L., 1994, Minimum price variations, discrete bid-ask
spreads, and quotation sizes. Reviewof
Financial Studies, 7, 149-178.
Holthausen, R. W., Leftwich, R. W., & Mayers, D. (1987). The
effect of large block transactions on
security prices: A cross-sectional analysis. The Journal of
Finance, 19, 237-267.
Holthausen, R. W., Leftwich, R. W., & Mayers, D. (1990).
Large-block transactions, the speedof
response, and temporary and permanent stock-price effects.
Journal of Financial
Economics, 26(1),71-95.
Ikenberry, D., and J.P. Weston, 2008, Clustering in US Stock
Prices after Decimalization.European
Financial Management, 14, 30-54.
Kraus, A., and Stoll, H. R. (1972). Price impacts of block
trading on the New York StockExchange. The
Journal of Finance, 27, 569-588.
Niederhoffer, V., 1965, Clustering of Stock Prices, Operations
Research, 13, 258-265.
Niederhoffer, V., 1966, A New Look at Clustering of Stock
Prices. Journal of Business, 39, 309-313.
Niederhoffer, V., and M.F.M. Osborne, 1966, Market Making and
Reveral on the Stock
Exchange.Journal of the American Statistical Association, 61,
897-916.
Ohta, W. (2006). An analysis of intraday patterns in price
clustering on the Tokyo
StockExchange. Journal of Banking & Finance, 30(3),
1023-1039.
-
21
Osborne, M.F.M., 1962, Periodic Structure in the Brownian Motion
of Stock Prices, Operations Research,
10, 345-379.
Saar, G. (2001). Price impact asymmetry of block trades: An
institutional trading explanation. Review
of Financial Studies, 14(4), 1153-1181.
Tversky, A., & Kahneman, D. (1991). Loss aversion in
riskless choice: A reference-
dependentmodel. The Quarterly Journal of Economics, 106(4),
1039-1061.
Wyckoff, P., 1963, Psychology of Stock Market Timing.
Prentice-Hall, Englewood Cliffs, NJ.
-
22
Table 1
Descriptive Statistics
This table provides summary statistics for our sample of firms.
We observe 2,626 sample firms. The information is segmented into
firms listed on the NYSE/Amex (1,243 firms) and NASDAQ (1,383
firms) exchanges. The sample consists of executed orders on the
NASDAQ exchange for the period January 1, 2012 to December 31,
2012. Market value is constructed for each firm as the closing
price times the number of shares outstanding, in $billions. Price
is the average daily closing price obtained from CRSP. Number of
Orders is calculated for each firm as the average number of limit
orders executed daily on the NASDAQ exchange. Buy (Sell) orders is
the average number of trades executed daily on the NASDAQ exchange
for buy (sell) orders for each firm. Executed shares is the average
number of shares traded daily on the NASDAQ exchange by firm. Buy
(Sell) shares is the average daily number of shares executed on the
NASDAQ exchange for buy (sell) initiated orders for each firm.
Volume is average daily trading volume for each firm obtained from
CRSP. Return volatility is calculated as the standard deviation of
daily returns for each firm over the sample period. Spread is the
percentage bid/ask spread calculated as the difference between the
closing Bid-Ask prices obtained from CRSP, divided by the
midpoint.
Mean Median Std. Dev.
Panel A: All Stocks
Market Value 6.0748 1.0838 22.9354
Price 33.26 22.57 58.97
Number ofOrders 1,706.14 503.00 3,249.70
Buy Orders 852.53 250.00 1,631.66
Sell Orders 853.83 247.00 1,635.73
Executed Shares 236,429.83 47,829.00 735,321.25
Buy Shares 118,226.33 23,778.00 367,565.03
Sell Shares 118,225.27 23,453.00 370,700.77
Volume 1,505,333.46 310,807.00 6,149,121.29
Return Volatility 0.0220 0.0203 0.0095
Spread% 0.0016 0.0007 0.0033
Panel B: NYSE & AMEX listed Stocks
Market Value 9.0478 2.2235 25.5455
Price 40.01 29.45 73.30
Number of Orders 2,018.79 707.00 3409.65
Buy Orders 1,008.07 351.00 1,710.12
Sell Orders 1,011.14 349.00 1,716.63
Executed Shares 277,679.57 65,722.00 817,838.78
Buy Shares 138,869.62 32,804.00 407,974.40
Sell Shares 138,848.66 32,379.00 412,529.78
Volume 2117074.30 585700.00 8028596.76
Return Volatility 0.0202 0.0189 0.0087
Spread% 0.0008 0.0004 0.0013
-
23
Panel C: NASDAQ listed stocks
Market Value 3.2406 0.5718 19.7215
Price 26.83 17.76 39.83
Number of Orders 1,408.07 388.00 3,059.91
Buy Orders 704.30 193.00 1,538.65
Sell Orders 703.81 189.00 1,539.76
Executed Shares 197,105.41 36,627.00 644475.52
Buy Shares 98,553.41 18,153.00 323,180.21
Sell Shares 98,557.61 17,814.00 324,628.36
Volume 922,149.34 159,580.00 3,422,296.78
Return Volatility 0.0237 0.0219 0.0098
Spread% 0.0024 0.0012 0.0044
-
24
Table 2
Clustering frequency
Mean clustering frequencies for buy and sell prices in limit
order data from the NASDAQ exchange for the period January 1, 2012
to December 31, 2012. We separate the statistics by exchange
listing into NYSE/AMEX and NASDAQ listed stocks. Clustering
percentage is constructed for each firm by week, as the total
number of prices (buy/sell) that rounded on $0.05 increments
divided by the total number of trades. Price is the weekly average
buy (sell) price for each firm. Simple pooled test-statistic are
constructed to assess the difference of means between buy and sell
clustering percentages.
Clustering Frequency
Price level subsample
Orders Buy Sell Difference Sell-
Buy Avg. Buy
Price Avg. Sell
Price
Panel A. NYSE, AMEX, and NASDAQ listed stocks
All stocks 25.26% 24.82% 25.69% 0.87%*** $33.08 $33.11
Over $200 33.16 32.99 33.33 0.33 514.70 514.89
$40-$200 23.89 23.58 24.20 0.63*** 63.77 63.82
$20-$40 24.61 24.20 25.01 0.81*** 28.90 28.92
$10-$20 26.13 25.62 26.65 1.03*** 14.65 14.66
Under $10 26.52 25.99 27.04 1.05*** 7.36 7.37
Panel B. NYSE & AMEX listed stocks
All stocks 23.70 23.46 23.93 0.48*** 40.01 40.04
Over $200 33.38 33.22 33.54 0.32 541.98 542.64
$40-$200 22.93 22.68 23.17 0.50*** 64.54 64.60
$20-$40 23.40 23.18 23.62 0.44*** 29.58 29.60
$10-$20 24.42 24.18 24.66 0.48*** 14.82 14.82
Under $10 24.50 24.24 24.76 0.52*** 7.43 7.43
Panel C. NASDAQ listed stocks
All stocks 26.73 26.11 27.35 1.23*** 26.54 26.56
Over $200 32.77 32.60 32.94 0.34 465.30 464.78
$40-$200 25.72 25.28 26.16 0.87*** 62.32 62.34
$20-$40 25.96 25.35 26.58 1.23*** 28.14 28.16
$10-$20 27.22 26.53 27.91 1.38*** 14.54 14.55
Under $10 27.50 26.85 28.16 1.31*** 7.34 7.35
***,**, * denote statistical significance at the 1%, 5%, and 10%
levels, respectively.
-
Table 3
Clustering Frequency Advancing vs. Declining
Mean clustering frequencies for executed limit orders separated
by buys and sells. Clustering percentage is constructed weekly for
each, and is the total number of prices that rounded on $0.05
increments divided by the number of executed orders. We categorize
advancing and declining markets in three separate ways: First in
Panel A, we report the results under extreme market conditions
which we define as follows. The advancing market period is when the
NASDAQ index increased by an unprecedented 17.78% (471 points),
which occurred between January 3, 2012 and April 01, 2012.
Conversely, the declining market is defined as when the NASDAQ
index experienced a significant decrease of 11.93% between the
period April 2, 2012 to June 1, 2012. Second in Panel B, we report
the results from defining advancing markets as weeks when the
average return on the S&P 500 is positive, and declining
markets as weeks when the average return on the S&P 500 is
negative. Lastly in Panel C, for the firm-specific measure, we
define a firm decline when the average weekly returns for a
particular stock is negative, and the opposite holds true for firm
advances. Weekly averaged buy/sell executed shares are also
reported. Simple pooled test-statistics are measured to assess the
difference of means between the buy and sell clustering percentages
across market conditions, which are reported in parentheses.
All Stocks NYSE/AMEX NASDAQ
Advance Decline Difference Advance Decline Difference Advance
Decline Difference
Panel A. Extreme Market Advances vs. Declines
Orders 25.55% 25.05% 0.49%***
23.83% 23.44% 0.39%***
27.16% 26.56% 0.60%***
(9.21) (7.46) (6.65)
Buy 25.02 25.29 -0.27***
23.51 23.60 -0.09
26.44 26.88 -0.44***
(-3.71) (-1.25) (-3.58)
Sell 26.07 24.81 1.26***
24.16 23.29 0.87***
27.87 26.25 1.63***
(16.11) (11.61) (12.42)
Buy Shares 620,604.34 636,743.00 742,229.68 764,959.52
506,179.48 516,562.52
Sell Shares 624,933.70 627,211.53 744,122.66 755,854.69
512,861.79 506,576.80
Panel B. S&P 500 Declining vs. Advancing Weeks
Orders 25.27% 25.24% 0.04%
23.75% 23.62% 0.13%***
26.71% 26.76% -0.05
(1.06) (3.45) (-0.87)
Buy 24.59 25.16 -0.57***
23.32 23.65 -0.34***
25.79 26.57 -0.79***
(-12.17) (-6.45) (-10.46)
Sell 25.96 25.31 0.65***
24.18 23.58 0.60***
27.63 26.95 0.69***
(12.18) (10.60) (8.00)
Buy Shares 549,603.18 552,900.49 649,874.75 654,399.44
454,850.61 457,310.36
Sell Shares 555,438.85 544,825.45 656,642.50 645,300.03
459,911.11 450,043.95
-
2
Panel C. Firm-Specific Advancing vs. Declining Weeks
Orders 25.32% 25.18% 0.14%***
23.72% 23.67% 0.05%
26.84% 26.60% 0.25%***
(3.86) (1.24) (4.35)
Buy 24.63 25.06 -0.43***
23.37 23.57 -0.19***
25.83 26.45 -0.62***
(-9.21) (-3.76) (-8.29)
Sell 26.01 25.31 0.70***
24.06 23.78 0.29***
27.85 26.74 1.11***
(13.27) (5.12) (13.03)
Buy Shares 534,498.53 570,767.45 631,893.93 675,890.15
441,653.04 472,774.70
Sell Shares 550,464.76 551,718.35 646,211.91 658,914.12
459,459.18 451,435.81
***,**, * denote statistical significance at the 1%, 5%, and 10%
levels, respectively.
-
Table 4 Cross-section Regressions
This table reports the results from estimating the following
equation using cross-sectional intra-day limit order data from
January to December 2012:
𝐶𝑙𝑢𝑠𝑡𝑒𝑟%𝑖,𝑡 = 𝛼𝑖 + 𝛿𝑡 + 𝛽1𝑆𝑖𝑧𝑒𝑖,𝑡 + 𝛽2𝑃𝑟𝑖𝑐𝑒𝑖,𝑡 + 𝛽31/√𝑁𝑇𝑖,𝑡 +
𝛽4𝑆𝑝𝑟𝑒𝑎𝑑%𝑖,𝑡+𝛽5𝑇𝑟𝑎𝑑𝑒𝑆𝑖𝑧𝑒𝑖,𝑡 + 𝛽6𝑆𝑒𝑙𝑙𝑖,𝑡+ 𝛽7𝐹𝑖𝑟𝑚𝐷𝑒𝑐𝑖,𝑡 + 𝛽8𝑀𝑘𝑡𝐷𝑒𝑐𝑖,𝑡
+ 𝛽9𝐸𝑥𝐷𝑒𝑐𝑖,𝑡 + 𝛽10𝑆𝑒𝑙𝑙 × 𝐷𝑒𝑐𝑖,𝑗,𝑡 + 𝜀𝑖,𝑗,𝑡
The dependent variable𝐶𝑙𝑢𝑠𝑡𝑒𝑟%𝑖,𝑡 is the observed frequency of
transaction prices that fall on $0.05 increments for
firm i in week t. We include the following as independent
variables: 𝑆𝑖𝑧𝑒𝑖,𝑡 is the market capitalization, or the daily
closing price multiplied by the number of shares outstanding, in
$billions. 𝑃𝑟𝑖𝑐𝑒𝑖,𝑡 is the average daily closing price
obtained from CRSP. 1/√𝑁𝑇𝑖,𝑡 is the inverse of the square root
of the number of transactions for each firm by week.
𝑆𝑝𝑟𝑒𝑎𝑑%𝑖,𝑡 is the difference between the CRSP closing bid and
ask prices divided by the midpoint. 𝑇𝑟𝑎𝑑𝑒𝑆𝑖𝑧𝑒𝑖,𝑡 is the average
trade size, and is constructed by dividing the number of executed
shares by the total number of transactions.
𝑆𝑒𝑙𝑙𝑖,𝑡 is an indicator variable labeled 1 if it is a sell order
and 0 for a buy order. We construct indicator variables which
take values of 1 for each of the three-defined declining
periods:Extreme Decline (𝐸𝑥𝐷𝑒𝑐𝑖,𝑡), Market Decline (𝑀𝑘𝑡𝐷𝑒𝑐𝑖,𝑡), and
Firm Decline (𝐹𝑖𝑟𝑚𝐷𝑒𝑐𝑖,𝑡). 𝑆𝑒𝑙𝑙 × 𝐷𝑒𝑐𝑖,𝑗,𝑡 is the independent
interaction between the 𝑆𝑒𝑙𝑙𝑖,𝑡 indicator variable and each of the
three specified declining period dummy variables. Intra-day limit
order data are used to measure daily values for all executed orders
that we average across weeks to calculate clustering percentage
(𝐶𝑙𝑢𝑠𝑡𝑒𝑟%𝑖,𝑡), number of trades, and executed shares. We report
t-statistics in parentheses after controlling for firm and time
fixed effects.
(1) (2) (3) (4) (5) (6) (7)
𝑆𝑖𝑧𝑒𝑖,𝑡 0.0001 0.0001 0.0001 0.0002 0.0001 0.0001*** 0.0002
(1.26) (1.27) (1.28) (1.05) (1.26) (1.27) (1.05)
𝑃𝑟𝑖𝑐𝑒𝑖,𝑡 0.0002*** 0.0002*** 0.0002*** 0.0001* 0.0002***
0.0002*** 0.0001*
(9.99) (10.02) (10.03) (1.93) (10.01) (10.03) (1.92)
1/√𝑁𝑇𝑖,𝑡 0.0432*** 0.0427*** 0.0428*** -0.1071*** 0.0425***
0.0430*** -0.1064***
(5.91) (5.84) (5.85) (-8.40) (5.81) (5.88) (-8.36)
𝑆𝑝𝑟𝑒𝑎𝑑%𝑖,𝑡 0.0204*** 0.0203*** 0.0203*** 0.0040*** 0.0204***
0.0204*** 0.0040***
(23.64) (23.51) (23.52) (2.97) (23.67) (23.65) (3.00)
𝑇𝑟𝑎𝑑𝑒𝑆𝑖𝑧𝑒𝑖,𝑡 0.0003*** 0.0003*** 0.0003*** 0.0004*** 0.0003***
0.0003*** 0.0004***
(56.56) (56.41) (56.49) (39.79) (56.47) (56.58) (39.81)
𝑆𝑒𝑙𝑙𝑖,𝑡 0.0087*** 0.0133*** 0.0136*** 0.0104***
(28.64) (32.64) (34.54) (18.49)
𝐹𝑖𝑟𝑚𝐷𝑒𝑐𝑖,𝑡 -0.0006* 0.0046*** (-1.93) (10.63)
𝑀𝑘𝑡𝐷𝑒𝑐𝑖,𝑡 -0.0004 0.0056*** (-1.41) (12.86)
𝐸𝑥𝐷𝑒𝑐𝑖,𝑡 -0.0045*** 0.0030*** (-10.20) (4.83)
𝑆𝑒𝑙𝑙 × 𝐷𝑒𝑐𝑖,𝑗,𝑡 -0.0103*** -0.0120*** -0.0151*** (-17.01)
(-19.59) (-17.20)
Observations 258,520 258,520 258,520 108,076 258,520 258,520
108,076
Adj R2 0.2598 0.2574 0.2574 0.3415 0.2606 0.2609 0.3440
***,**, * denote statistical significance at the 1%, 5%, and 10%
levels, respectively.
-
Table 5
Cross-section Regressions Partitioned Across Buy and Sell
Orders
This table reports the results from partitioning the data into
buy and sell trades and estimating the following equation:
𝐶𝑙𝑢𝑠𝑡𝑒𝑟%𝑖,𝑡 = 𝛼𝑖 + 𝛿𝑡 + 𝛽1𝑆𝑖𝑧𝑒𝑖,𝑡 + 𝛽2𝑃𝑟𝑖𝑐𝑒𝑖,𝑡 + 𝛽31/√𝑁𝑇𝑖,𝑡 +
𝛽4𝑆𝑝𝑟𝑒𝑎𝑑%𝑖,𝑡+𝛽5𝑇𝑟𝑎𝑑𝑒𝑆𝑖𝑧𝑒𝑖,𝑡+ 𝛽6𝐹𝑖𝑟𝑚𝐷𝑒𝑐𝑖,𝑡 + 𝛽7𝑀𝑘𝑡𝐷𝑒𝑐𝑖,𝑡 +
𝛽8𝐸𝑥𝐷𝑒𝑐𝑖,𝑡 + 𝜀𝑖,𝑗,𝑡
The dependent variable 𝐶𝑙𝑢𝑠𝑡𝑒𝑟%𝑖,𝑡 is the observed frequency of
transaction prices that fall on $0.05 increments for firm i in week
t. The independent variables have been defined previously. We
partition the sample by limit-sell and limit-buy orders. We report
t-statistics in parentheses after controlling for firm and time
fixed effects.
Sell Buy
(1) (2) (3) (4) (5) (6)
𝑆𝑖𝑧𝑒𝑖,𝑡 0.00002 0.00003 -0.0002 0.0002 0.0002 0.0005** (0.14)
(0.22) (-0.99) (1.52) (1.44) (2.49)
𝑃𝑟𝑖𝑐𝑒𝑖,𝑡 0.0004*** 0.0004*** 0.0004*** 0.0001* 0.0001*
-0.0002**
(11.74) (11.92) (4.93) (2.10) (1.94) (-2.25)
1/√𝑁𝑇𝑖,𝑡 0.0465*** 0.0464*** -0.2547*** 0.0462*** 0.0471***
0.0467***
(4.22) (4.22) (-13.32) (4.85) (4.95) (2.79)
𝑆𝑝𝑟𝑒𝑎𝑑%𝑖,𝑡 0.0214*** 0.0215*** 0.0029 0.0219*** 0.0218***
0.0064***
(16.54) (16.61) (1.44) (18.78) (18.71) (3.42)
𝑇𝑟𝑎𝑑𝑒𝑆𝑖𝑧𝑒𝑖,𝑡 0.0004*** 0.0004*** 0.0003*** 0.0003*** 0.0003***
0.0004***
(44.19) (44.46) (26.32) (35.23) (35.08) (29.03)
𝐹𝑖𝑟𝑚𝐷𝑒𝑐𝑖,𝑡 0.0053*** 0.0043*** (-11.72) (10.45)
𝑀𝑘𝑡𝐷𝑒𝑐𝑖,𝑡 -0.0065*** 0.0057*** (-14.46) (13.87)
𝐸𝑥𝐷𝑒𝑐𝑖,𝑡 -0.0126*** 0.0035*** (-20.06) (5.68)
Observations 129,270 129,270 54,043 129,250 129,250 54,033
Adj R2 0.2983 0.2986 0.3989 0.2569 0.2574 0.3501
***,**, * denote statistical significance at the 1%, 5%, and 10%
levels, respectively.
-
Table 6
Cross-section Regressions Separated Across Buy/Sell Orders and
Exchange Listings
This table reports the results from estimating the following
equation:
𝐶𝑙𝑢𝑠𝑡𝑒𝑟%𝑖,𝑡 = 𝛼𝑖 + 𝛿𝑡 + 𝛽1𝑆𝑖𝑧𝑒𝑖,𝑡 + 𝛽2𝑃𝑟𝑖𝑐𝑒𝑖,𝑡 + 𝛽31/√𝑁𝑇𝑖,𝑡 +
𝛽4𝑆𝑝𝑟𝑒𝑎𝑑%𝑖,𝑡+𝛽5𝑇𝑟𝑎𝑑𝑒𝑆𝑖𝑧𝑒𝑖,𝑡+ 𝛽6𝐹𝑖𝑟𝑚𝐷𝑒𝑐𝑖,𝑡 + 𝛽7𝑀𝑘𝑡𝐷𝑒𝑐𝑖,𝑡 +
𝛽8𝐸𝑥𝐷𝑒𝑐𝑖,𝑡 + 𝜀𝑖,𝑗,𝑡
All variables have been defined previously. For robustness, we
partition the sample by sell and buy orders and also by exchange
listings (NYSE/AMEX, NASDAQ). We note that the sample consists of
executed orders on the NASDAQ exchange only, but the exchange
listings differ among stocks. We draw seemingly identical
coefficients on all control variables; therefore, we report only
the coefficients on the declining market indicator variables.
T-statistics are reported in parentheses after controlling for firm
and time fixed effects.
Panel A. NYSE/AMEX
Sell Buy
(1) (2) (3) (4) (5) (6)
𝐹𝑖𝑟𝑚𝐷𝑒𝑐𝑖,𝑡 -0.0023*** 0.0018***
(-4.42) (3.80)
𝑀𝑘𝑡𝐷𝑒𝑐𝑖,𝑡 -0.0060*** 0.0034*** (-11.56) (7.23)
𝐸𝑥𝐷𝑒𝑐𝑖,𝑡 -0.0079*** 0.0017*** (-11.84) (2.67)
Control Variables Yes Yes Yes Yes Yes Yes
FE Yes Yes Yes Yes Yes Yes
Observations 62,762 62,762 26,175 62,751 62,751 26,172
Adj R2 0.1931 0.1946 0.2661 0.1798 0.1803 0.2544
Panel B. NASDAQ
Sell Buy
(1) (2) (3) (4) (5) (6)
𝐹𝑖𝑟𝑚𝐷𝑒𝑐𝑖,𝑡 -0.0081 0.0065 (-11.16) (9.95)
𝑀𝑘𝑡𝐷𝑒𝑐𝑖,𝑡 -0.0070 0.0078 (-9.67) (11.86)
𝐸𝑥𝐷𝑒𝑐𝑖,𝑡 -0.0172 0.0052 (-16.36) (5.08)
Control Variables Yes Yes Yes Yes Yes Yes
FE Yes Yes Yes Yes Yes Yes
Observations 66,508 66,508 27,868 66,499 66,499 27,861
Adj R2 0.3068 0.3065 0.4095 0.2639 0.2644 0.3525
***,**, * denote statistical significance at the 1%, 5%, and 10%
levels, respectively.