Quantitative Finance, Vol. 5, No.

4, August 2005, 357–364

Order book approach to price impact

P. WEBER and B. ROSENOW*
Institut für Theoretische Physik, Universität zu Köln, D-50923 Germany

(Received 16 January 2004; in final form 23 June 2005)

Buying and selling stocks causes price changes, which are described by the price impact
function. To explain the shape of this function, we study the Island ECN orderbook.
In addition to transaction data, the orderbook contains information about potential supply
and demand for a stock. The virtual price impact calculated from this information is four
times stronger than the actual one and explains it only partially. However, we find a strong
anticorrelation between price changes and order flow, which strongly reduces the virtual price
impact and provides for an explanation of the empirical price impact function.

Keywords: Order book; Price impact; Resiliency; Liquidity

In a perfectly efficient market, stock prices change due to to be a concave function of volume imbalance, which
the arrival of new information about the underlying com- increases sublinearly for above average volume imbal-
pany. From a mechanistic point of view, stock prices ance. However, there is evidence that this concave shape
change if there is an imbalance between buy and sell cannot be a property of the true price impact a trader in an
orders for a stock. These ideas can be linked by assuming actual market would observe. Firstly, such a concave
that someone who trades a large number of stocks might price impact would be an incentive to do large trades
have private information about the underlying company, in one step instead of breaking them up into many smaller
and that an imbalance between supply and demand trans- ones as is done in practice. Secondly, in the presence
mits this information to the market. In this sense, volume of bluffers in the market, the enforcement of a strictly
imbalance and stock price changes should be connected linear price impact is the only way in which a market
causally, i.e. prices go up if demand exceeds supply and maker or liquidity trader can protect herself against
go down if supply exceeds demand. The analysis of huge suffering losses.
financial data sets (Takayasu 2002) allows a detailed One possible way to achieve a better understanding of
study of the price impact function (Hasbrouck 1991, the average price impact is the analysis of information
Hausman et al. 1992, Kempf and Korn 1999, Evans about potential supply and demand stored in the limit
and Lyons 2002, Hopman 2002, Plerou et al. 2002, order book. Using this information, one can calculate a
Rosenow 2002, Bouchaud et al. 2003, Gabaix et al. virtual or instantaneous price impact, which would be
2003, Lillo et al. 2003, Potters and Bouchaud 2003), caused by a market order matched with limit orders
which quantifies the relation between volume imbalance from the order book. This virtual price impact can be
and price changes. used as a reference point for an understanding of the
In a series of previous studies (Hasbrouck 1991, average price impact of market orders. An analysis of
Hausman et al. 1992, Kempf and Korn 1999, Evans the limit order book of the Stockholm Stock Exchange
and Lyons 2002, Plerou et al. 2002, Rosenow 2002, (Sandas 2001) suggests that the virtual price impact cal-
Bouchaud et al. 2003, Gabaix et al. 2003, Lillo et al. culated from the order book is significantly larger than
2003, Potters and Bouchaud 2003), the average price what is expected from a regression model. Similarly,
impact of an imbalance between buy and sell market a difference between hypothetical and actual price impact
orders or of individual market orders was studied. was found in Coppejans et al. (2001) and considered as
Generally, this average price impact function was found evidence for discretionary trading, e.g. large trades are
more likely to be executed when the order book has
sufficient depth.
In this paper, we calculate both the average price
*Corresponding author. Email: [email protected] impact of market orders and the virtual price impact
Quantitative Finance
ISSN 1469–7688 print/ISSN 1469–7696 online # 2005 Taylor & Francis
http://www.tandf.co.uk/journals
DOI: 10.1080/14697680500244411
358 P. Weber and B. Rosenow

Table 1. Summary statistics for the ten stocks studied. All values are calculated for five minute intervals. Numbers in brackets
are in units of the respective standard deviation.

AMAT BRCD BRCM CSCO INTC KLAC MSFT ORCL QLGC SEBL
Returns
!G 0.004 0.0053 0.0049 0.0034 0.0034 0.0036 0.0025 0.0039 0.0041 0.005
jGmax j 0.026 0.057 0.039 0.037 0.023 0.026 0.021 0.049 0.033 0.066
(6.46) (10.77) (8.1) (11.05) (6.86) (7.17) (8.3) (12.74) (8.02) (13.23)
Median of |G| 0.0022 0.0028 0.0025 0.0018 0.0018 0.002 0.0013 0.002 0.0022 0.0026
(0.55) (0.53) (0.52) (0.52) (0.55) (0.55) (0.53) (0.51) (0.55) (0.52)
Volume imbalance
!Q 9551.78 9700.37 6196.3 22652.52 16660.35 5862.25 11029.98 16010.11 6513.98 7152.31
jQmax j 344657 193633 127274 424923 323016 116790 376168 396978 82540 189966
(36.08) (19.96) (20.54) (18.76) (19.39) (19.92) (34.1) (24.8) (12.67) (26.56)
Median of |Q| 5958 6069 4086.5 14902.5 11208 3648 7307.5 10100 4078.5 4429
(0.62) (0.63) (0.66) (0.66) (0.67) (0.62) (0.66) (0.63) (0.63) (0.62)
Market orders
hNmarket i 52.38 45.31 44.97 58.34 72.62 58.89 83.59 47.41 56.82 35.02
!Nmarket 33.89 36.4 36.83 39.3 49.14 39.92 53.31 37.43 37.72 30.92
Limit orders
hNadded i 321.27 247.06 233.54 302.6 391.91 344.59 429.8 227.17 303.68 199.99
!Nadded 183.49 156.22 160.86 213.87 232.65 169.97 244.83 165 162.02 156.01
hNcancelled i 286.91 220.13 206.39 256.38 343.71 305.64 376.4 191.54 268.41 177.09
!Ncancelled 173.15 144.14 144.22 199.23 220.71 154.95 229.11 152.41 146.79 141.82

from information contained in the Island ECN order and cancellation of limit orders. If a trader is willing to
book. By comparing the two, we find that the virtual sell a certain volume (number of shares) of a stock at
price impact is more than four times stronger than the a given or higher price, she places a limit sell order.
actual one. In order to explain this surprising discrepancy, For buying at a given or lower price, a limit buy order
we study time-dependent correlations between order flow is placed. Limit orders are stored in the order book at
and returns and find strong evidence for resiliency: the their respective price with the respective volume. The
market recovers from random uninformative price shocks sell limit order with the lowest price defines the ask
as the flow of limit orders is anticorrelated with returns price Sask for that stock. Similarly, the buy limit order
in contrast to the positive correlations between returns with the highest price defines the bid price Sbid. If a buy
and market orders. Thus, limit orders placed in response limit order with a price higher than the current ask price is
to returns strongly reduce the virtual price impact placed, it is executed immediately against the sell limit
and provide for a link between virtual and actual price orders at the ask price, vice versa for sell limit orders at
impact. or below the bid price. Such marketable limit orders
We include the possible influence of discretionary will be called market orders in the following. Market
trading in our analysis by studying the average virtual orders are placed by impatient traders who want to trade
price impact, the typical virtual price impact and the vir- immediately.
tual price impact calculated from the average order book. The time period we study is the entire year 2002. Each
The average virtual price impact is strongly influenced day, we disregard the first five minutes due to the build
by low liquidity periods and hence not representative up of the order book. We checked that our results are
for the actual impact of market orders. In order to include robust with respect to this choice by removing e.g. the
the influence of discretionary trading in a semiquantita- first fifteen minutes and the last five minutes per day.
tive way, we base our analysis on the virtual price impact We consider 250 trading days and divide each trading
as calculated from the average order book, which is day into 77 intervals of five minute length.
weaker than the average and the typical price impact. We chose the 10 most frequently traded stocks for
Including the effect of resilience as apparent from the the year 2002, ticker symbols and descriptive statistics
above-mentioned anticorrelations, the actual price can be found in table 1. The volume of market buy orders
impact can be explained quantitatively. is counted as positive and the volume of market sell
We analysed limit order data from the Island ECN, orders as negative, and the volume imbalance Q(t) is
NASDAQ’s largest electronic communication network, defined as the sum of all signed market orders placed
which comprises about 20% of all trades. Island ECN is in the time interval ½t, t þ !t# with !t ¼ 5 miny. For the
a limit order driven market which allows for placement definition of stock price changes, we use the same five

yWe do not include market orders executing ‘hidden’ limit orders in the definition of QðtÞ as we want to compare with the order
book containing ‘visible’ orders only.
 Order book approach to price impact 359

minute intervals as for the volume imbalance and define

4
five minute returns
GðtÞ ¼ ln SM ðt þ !tÞ ' ln SM ðtÞ, ð1Þ 2

Return G
where the midquote price SM ðtÞ ¼ 12 ðSbid ðtÞ þ Sask ðtÞÞ is
0
the arithmetic mean of bid and ask price. We analyse
changes of the midquote price rather than transaction 1
−2
price changes to avoid a distortion of our analysis
due to the bid–ask bounce. The midquote price changes 0.1

(i) when a market order fills the current bid or ask volume −4 0.1 1 10
and changes the bid or ask price in this way, (ii) when the −10 −5 0 5 10
limit orders at the current bid or ask price are cancelled or Volume imbalance Q
(iii) new limit orders are placed inside the spread. In the
first part of the paper we will concentrate on the relation Figure 1. The price impact function Imarket ðQÞ describes returns
between the market order flow and changes of the mid- between the beginning and end of five minute intervals
quote price. In the second part, we will see that one has in response to the volume imbalance in the same time interval.
It is a monotonously increasing and concave function of the
to include the influence of limit orders in order to signed market order volume. A logarithmic plot (inset) shows
understand the relation between market order flow and that the function can be fitted by a power law.
price changes.
To make different stocks comparable, we normalize
the return time series G(t) by their standard deviation 15 min it was found to be 0.5 (Plerou et al. 2002,
! G and the volume imbalance
! time
" series Q by its first Gabaix et al. 2003). An argument by Zhang (1999)
centred moment !Q ¼ jQ ' hQij ; as for volumes the sec- based on a simple market maker model predicts a value
ond moment is not well defined due to a slow decay of the of 0.5 as well. Analysing the TAQ data basey for the year
probability distribution. This normalization ensures that 1997 instead of the years 1994 and 1995 as in Plerou et al.
the price impact functions for different stocks collapse (2002) and Gabaix et al. (2003) and for a time interval of
(Plerou et al. 2002). five minutes as compared to the fifteen minute interval in
Price impact of market orders. We define the price Plerou et al. (2002) and Gabaix et al. (2003), we find an
impact of market orders as the conditional expectation exponent 0.58 for transaction price changes and 0.75
value for midquote price changes. For the Island ECN data,
the exponent is 0.76 for midquote returns and 0.73 for
Imarket ðQÞ ¼ hG!t ðtÞiQ : ð2Þ
transaction returns, both calculated on a time scale of five
It describes the average relation between the return G in a minutes. Intuitively, one might expect a stronger price
given five minute interval and the market order flow Q in impact for the analysis of transaction prices as compared
the same time interval. The functional form of Imarket ðQÞ to the analysis of midquote prices. This intuition is in
is shown in figure 1, it is in general agreement with the agreement with the results of the empirical analysis:
results (Kempf and Korn 1999, Plerou et al. 2002, as most five minute volume imbalances Q are small with
Rosenow 2002, Lillo et al. 2003). We find that Imarket ðQÞ Q < 1 (in units of ! Q), there are many data points avail-
is a concave function of volume imbalance (Hasbrouck able in the region jQj<1. Hence, on a logarithmic scale
1991), which can be well fitted by a power law one uses as many bins in this interval as in the region
G ¼ 0:48Q0:76 . In principle, it would be desirable to cal- jQj > 1, and the region jQj < 1 contributes significantly
culate the price impact function as a dynamical and stock to the result of a logarithmic fit. However, for jQj < 1,
by stock function. Due to our averaging procedure one has jQj" > jQj# for " < #, and the price impact for
however, we obtain only one data point per five minute transaction prices is indeed stronger than the price impact
interval, and we need to aggregate over time and different for midquote prices.
stocks to obtain sufficient statistics for the calculation The concave shape of the function is very surprising:
of the conditional average in the region of large this type of price impact would theoretically be an incen-
volume imbalances. An analysis of the time dependence tive to make large trades as they would be less costly than
of liquidity is presented in Weber and Rosenow (2004). many small ones. In contrast, a convex price impact
We note that the power law exponent characterizing would encourage a trader to brake up a large trade into
the price impact function depends both on the time several smaller ones, which is what actually happens.
horizon and on the market studied. Generally, the expo- Having this in mind, we want to understand the mechan-
nent tends to increase for an increasing time horizon ism responsible for this concave shape and analyse
(Plerou et al. 2002). On a tick by tick basis, the exponent the trading information contained in the limit order
is very small (Lillo et al. 2003) or the functional relation book. There are two mechanisms conceivable for the
can be characterized by a logarithm (Potters and explanation of the concave shape: (i) we will argue that
Bouchaud 2003). On an intermediate time scale of to a large extent the concave shape can be explained by

yWe analysed the 44 most frequently traded NASDAQ stocks using data from the Trades and Quotes (TAQ) data base published by
the New York Stock Exchange.
360 P. Weber and B. Rosenow

resiliency, i.e. by the interplay of limit orders and 0.25

market orders, and (ii) one expects discretionary trading, (a)

Orderbook volume Q
0.2
i.e. large trades are made when the market is very
liquid and when the price impact is small. In this 0.15
paper, we will concentrate on mechanism (i) and will
leave the empirical study of mechanism (ii) to a future 0.1
publication.
0.05
Order book and virtual price impact. The whole market
information of the Island ECN is stored in text files where 0
each line represents a message of one of the four 0.1 1 10 100 1000
major types: add limit order, cancel limit order, execute Orderbook depth γ
limit order or trade message. While the last type 10
announces the execution of hidden orders which are
(b)
not visible in the order book, the first three types allow
for a complete reconstruction of the order book at every 5
instant of time. The placement of a market order is indi-

Return G
cated by the execution of one or several limit orders. 0
For each stock, the database contains about one to four
million messages for the whole year. We do not include
−5
‘hidden’ limit orders in our analysis as there is no infor-
mation about their placement in our data base. We have
checked that on the level of the average price impact −10
fuction Imarket ðQÞ the inclusion of hidden limit orders do −10 −5 0 5 10
not change our result as the change in ! Q accounts for the Volume imbalance Q
additional order volume.
Figure 2. (a) The average order book is characterized by a
Each message contains all necessary information: the maximum at %i ¼ 1 and a slow decay up to %i ¼ 100. The nega-
ticker symbol of the respective stock, the time past mid- tive side (buy orders) of the order book is similar to the positive
night in milliseconds, the number of shares, the limit one. (b) The virtual price impact function Ihbooki ðQÞ (circles)
price, an indicator whether the order is of buy or sell calculated from the average limit order book is a convex func-
tion of volume imbalance and much steeper than the price
type, and a unique order reference number. We use this
impact of market orders ! (squares)." Volume imbalances are
number as a key to store and identify each order in our measured in units of !Q ¼ jQ ' hQij and returns are measured
data structure and perform (partial) executions and in units of standard deviations ! G.
cancellations until an order is completely executed or
cancelled.
Since all open orders are purged from the book every For the calculation of the time-dependent density
evening, we can process the data day by day starting with functions for all ten stocks from information about place-
an empty book each morning. In our analysis, every ment, cancellation, and execution of limit orders con-
trading day is divided into intervals of five minutes length. tained in the Island ECN data base, we had to process
To make the order book information amenable to a about 60 GB of data.
statistical analysis, we calculate at the beginning of each Next, we study the average order book $hbooki ð%i Þ ¼
time interval and for each stock k the current order book h$k ð%i , tÞi, where h:::i denotes an average over both time
as a density function $k ð%i , tÞ. Due to the complexity of t and different stocks k. It is characterized by a flat
this calculation, we use a discrete coordinate % i to obtain maximum at %i ( 1 and a slow decay for large % i (see
the order book from the data structure by sorting orders figure 2 (a)). Its overall shape agrees with the results of
with respect to their limit prices and aggregating the Challet and Stinchcombe (2001), Maslov and Mills (2001)
number of shares on a lattice with spacing !%. For and Bouchaud et al. (2002).
each price Slimit, at which a limit order is placed, the We want to compare the actual price impact Imarket ðQÞ
coordinate % i is defined as to a virtual price impact function calculated from
# the average order book. To this end, we calculate
½ðlnðSlimit Þ ' lnðSbid ÞÞ=!% #!% limit buy order, the market depth for a given return and invert this
%i ¼
½ðlnðSlimit Þ ' lnðSask ÞÞ=!% #!% limit sell order. relation. We imagine a trader who wants to buy a volume
ð3Þ Q of stocks and has only offers from the order book
available. Beginning at the ask price, she executes as
Here, the function ½x# denotes the smallest integer larger many limit orders as necessary to match her market
than x. We define the density function such that order, and changes the ask price by an amount G.
$k ði!%, tÞ!% is the total volume in the price interval Traded volume (or market depth) Qhbooki ðGÞ and return
½ði ' 1Þ!%, i!%# in the order book, where i is an integer. G are related by
In our analysis, we chose !% ¼ 0:3!G as a compromise X
between computational speed and accuracy. We note Qhbooki ðGÞ ¼ $ðbookÞ ð%i Þ!%: ð4Þ
that throughout the paper % is measured in units of ! G. %i )G
 Order book approach to price impact 361

By inverting equation (4), we define the virtual price 10

impact Ihbooki ðQÞ with respect to the average order
book. Here, we assume that the bid–ask spread remains 5
constant and that the midquote price changes by the same

Return G
amount as the ask price. According to the above defini-
tion, the virtual price impact Ihbooki ðQÞ describes the price 0
change due to a single market order of arbitrary size.
Now, we want to compare the virtual price impact with −5
the actual price impact Imarket ðQÞ, which is calculated as
a function of the volume imbalance Q(t) aggregated over
a five minute interval. Using the virtual price impact to −10
−5 −4 −3 −2 −1 0 1 2 3 4 5
predict a return due to a time aggregated volume imbal-
Volume imbalance Q
ance is an approximation, which is justified if (i) the order
book is symmetric with respect to the buy and sell side Figure 3. The average virtual price impact function hIbook iðQÞ
and if (ii) its nonlinearities do not influence the final result (full squares) is steeper than the typical virtual price impact
essentially. As far as the average price impact is con- hIbook imedian (open squares) calculated by taking the median
instead of the average. The virtual price impact Ihbooki ðQÞ calcu-
cerned, the assumption of a buy–sell symmetry of the lated from the average order book (full circles) is weaker than
order book is justified and one sees from figure 2 that the other two.
the nonlinearity of the virtual price impact is weak.
We find that the virtual price impact Ihbooki ðQÞ is four
times stronger than the price impact of actual market functions Ibook ðQ, t, KÞ at these grid points are calculated
orders (see figure 2 (b)), a volume imbalance of 5!Q causes by interpolation.
a virtual price change of 8!G but only an actual price In doing so, one obtains hIbook iðQÞ as a convex function
change of 2!G . In addition, Ihbooki ðQÞ is a convex function of signed order volume which is much steeper than the
that can be fitted by a power law Ihbooki ðQÞ ¼ 1:22Q1:19 , average price impact, see figure 3. To reduce the influence
and not a concave function as Imarket ðQÞ. of low liquidity periods on the virtual price impact, we
Strictly speaking, the average virtual price impact can- have calculated a typical price impact hIbook imedian ðQÞ by
not be calculated from the average order book. Instead, replacing the average over time and different stocks by
the inversion of the depth as a function of return should the median. For large trading volumes, hIbook imedian ðQÞ
be performed at each instant of time and for each stock, is considerably smaller than hIbook iðQÞ, see figure 3.
and the average should be performed afterwards. To this hIbook imedian ðQÞ is a convex function of signed volume
end, we define a time resolved and per stock depth and quite similar to Ihbooki ðQÞ.
X As a basis for the explanation of the actual price
Qbook ðG, t, kÞ ¼ $k ð%i , tÞ!%: ð5Þ impact Imarket ðQÞ, one should not consider hIbook iðQÞ
%i )G due to the influence of discretionary trading: in low
liquidity periods, which dominate hIbook iðQÞ, one expects
By inverting this relation at each instant of time and little trading activity. On the other hand, if the liquidity
for each stock, we obtain the virtual price impact in the order book is average in the sense of the median,
Ibook ðQ, t, kÞ. We find that this function fluctuates strongly one expects discretionary trading to play a less important
in time and that its average over time and different stocks role. Therefore, hIbook imedian ðQÞ is a better reference point
hIbook iðQÞ is dominated by rare events with low liquidity. for the explanation of the actual price impact Imarket ðQÞ.
For this reason, the calculation of hIbook iðQÞ is some- As Ihbooki ðQÞ is again a little less steep than the typical
what subtle. In time intervals with very low liquidity, price impact, one can argue that due to the influence of
the domain of Ibook ðQ, t, kÞ does not even extend up to discretionary trading it is the best starting point for the
0:5!Q since the amount of limit orders stored in the order explanation of the average price impact of market orders
book is too small. In this case, the return caused by an Imarket ðQÞ. The exact influence of discretionary trading
order with signed volume Q > 0:5!Q would be undefined certainly needs to be analysed in more detail. However,
and the average over all time intervals would be undefined such an analysis is beyond the scope of the present paper
as well. In order to expand the domain of hIbook iðQÞ to as the data analysis presented here is already quite
at least 3:5!Q , we extrapolate the depth linearly by con- involved.
necting the last defined data point (with largest Q and G) The average order book and thus the virtual price
with the origin. Since this procedure is necessary only for impact can be decsribed by ‘zero intelligence models’
few time intervals, our extrapolation method does not (Bouchaud et al. 2002, Daniels et al. 2003), in which
disturb the final result. We checked this by using different orders are placed randomly. Due to the fact that we
methods, e.g. by continuing the depth function by a have calculated our virtual price impact from an average
horizontal line instead of a linear extrapolation. The influ- over time and different stocks, it is comparable to such
ence of the choice of a specific extrapolation method zero intelligence models. The disagreement between our
is clearly visible only for large volumes Q > 4!G . The virtual price impact and the actual one is evidence for
average of hIbook iðQÞ is calculated on an equidistant the necessity to include additional mechanisms describing
grid on the Q axis and the values of the individual ‘intelligent’ or collective behaviour.
362 P. Weber and B. Rosenow

Correlations between order flow and returns. Which

(a)

Correlation function cmarket

effect is responsible for the pronounced difference
0.3
between virtual and actual price impact? In the following,
we will argue that a strong anticorrelation between
returns and limit orders gives rise to resiliency and 0.2
reduces the virtual price impact. Thus, virtual and actual
price impact can be linked to each other. In our analysis, 0.1
we calculate correlation functions between returns and
the flow of market or limit orders. A technical alternative
0
to the calculation of correlation functions would be
Granger causality tests using a VAR analysis (Chen
et al. 2001) or a semi-nonparametric estimate of the (b)

Correlation function climit

joint density of price change and volume (Gallant et al.
0
1992). These techniques were used to study the relation
between volatility and trading volume.
In order to understand how order flow and price −0.1
changes are related, we study the correlation functions
−0.2
hQ" ðt þ &ÞGðtÞi ' hQ" ðtÞihGðtÞi
c" ð&Þ ¼ ð6Þ
!Q" !G
−0.3
between the volume imbalance of market orders 30
−20 −10 0 10 20
(" ¼ marketÞ or limit orders (" ¼ limit) and returns. In
Time τ [min]
order to increase the time resolution of these correlators,
the order volume is measured in intervals ½t, t þ 't# with Figure 4. Correlation functions between return and signed
width 't ¼ 50 s, and the returns are recorded for five order flow (buy minus sell orders). (a) Market orders and
minute intervals as before. For " ¼ market, Qmarket ðtÞ is returns show strong positive equal time correlations (shaded
the volume of signed market orders, and for " ¼ limit region) decaying slowly to zero. (b) Limit orders preceding
returns have weak positive correlations with them, while equal
X
1 $ % time correlations (shaded region) are strongly negative.
Qlimit ðtÞ ¼ signð'%i Þ Qadd canc
't ð%i Þ ' Q't ð%i Þ !% ð7Þ
'1
Limit order flow and feedback mechanism. The anti-
is the net volume of limit buy orders minus the net correlation between returns and limit order flow suggests
volume of limit sell orders. In equation (7), Qadd 't ð%i Þ is that dynamical effects are responsible for the difference
the volume of limit orders added to the book at a depth between virtual and actual price impact. We take into
% i and Qcanc
't ð%i Þ is the volume of orders cancelled from the account the influence of discretionary trading in a semi-
book. quantitative way by using Ihbooki ðQÞ, the least steep of the
The correlation functions are plotted in figure 4. different virtual price impact functions, as the starting
We find that cmarket ð&Þ is zero for & < '50 s as required point for the explanation of the average price impact
for an efficient market where returns cannot be predicted of market orders. The rational behind this choice is the
over extended periods of time. For times & * '50 s, we idea that trades will be made preferentially if the liquidity
find positive correlations which are strongest when is above average.
the time intervals for orders and returns overlap. For In order to connect the virtual price impact to the
& * 300 s (non-overlapping time intervals), we observe a actual one, one has to take into account both the average
slow decay of the correlation function which is probably limit order flow within a five minute interval and the limit
caused by the strong autocorrelations of the market order flow in response to price changes. The density of
order flow (Hasbrouck 1991, Hopman 2002, Lillo and the average limit order flow is described by
Farmer 2003).
The correlation function between limit orders and $flow ð%i Þ ¼ hQadd canc
!t ð%i Þ ' Q!t ð%i Þi ð8Þ
returns vanishes for negative times & < '50 s and has a
small positive value climit ð'50 sÞ ¼ 0:04. Surprisingly, with !t ¼ 5 min. Near the ask price, the net volume of
for zero and positive time differences there is a significant incoming limit orders is five times larger than the volume
anticorrelation between limit orders and returns, which stored in the average order book. More than one ! G
is strongest for & ¼ 250 s (overlapping time intervals) away from the bid and ask price, the order flow decreases
and decays slowly to zero for large positive times. We rapidly. Summation of the order flow density up to
interpret this anticorrelation as an indication that rising a given return
P G contributes the additional volume
prices cause an increased number of sell limit orders and Qflow ðGÞ ¼ %i )G $flow ð%i Þ!%, which is displayed in
vice versa for falling prices. Price changes seem to be figure 5 (a). It grows fast for small returns and saturates
counteracted by an orchestrated flow of limit orders. for larger returns.
 Order book approach to price impact 363

(a)
predictable part of the market order flow in such a way
20
that returns become uncorrelated. In order to capture this
Volume imbalance Q

complicated interplay partially, we integrate over all the

10
correlated future limit order flow. This procedure would
be exact in the (admittedly artificial) situation of ‘station-
0
ary price changes’ in which we assume that GðtÞ + G is
constant in time. Then, the choice t0 ¼ 30 min makes sure
−10
that also the additional limit order volume due to returns
in past time intervals is taken into account. The correla-
−20
tion volume corresponding to a return G is
−8 −4 0 4 8 X
Return G
Qcorr ðGÞ ¼ $c ð%i , GÞ!%: ð10Þ
%i )G
8
(b) Qcorr ðGÞ is slightly negative for small G and increases
then almost linearly for larger G (see figure 4 (a)). The
4
total volume Q(G) corresponding to a return G is the sum
Return G

QðGÞ ¼ Qbook ðGÞ þ Qflow ðGÞ þ Qcorr ðGÞ ð11Þ

0
of the volume Qbook ðGÞ of orders stored in the limit order
−4
book up to a depth G, the volume Qflow ðGÞ arriving within
a five minute interval on average, and the correlation
volume Qcorr ðGÞ. The theoretical price impact function
−8 Itheory ðQÞ calculated by inverting this relation is shown
−10 −5 0 5 10
Volume imbalance Q
in figure 5 (b).
The agreement between Itheory ðQÞ and Imarket ðQÞ is
Figure 5. (a) The average flow of limit orders integrated up to quite good, up to G ¼ 10!G there are no deviations within
an order book depth G (grey squares) changes rapidly at small the error bars of Imarket ðQÞ. Market resiliency gives rise
returns and then stays constant. The additional volume of limit
orders Qcorr in response to a return G (circles) increases linearly to additional liquidity, the influx of limit orders anti-
for large G. (b) Empirical price impact Imarket ðQÞ of market correlated with past returns has a strong influence on
orders (open circles) compared to the theoretical price impact the price impact of market orders. It strongly reduces
Itheory ðQÞ (full circles), which takes into account orders from the the virtual price impact and is responsible for the empiri-
order book, the average flow of limit orders and the additional cally observed concave shape of the price impact func-
flow Qcorr.
tion. When calculating the correlation volume Qcorr ðGÞ
over time intervals of length t0 ¼ 5 min instead of
Furthermore, there is an additional volume of incom- t0 ¼ 30 min, only half the discrepancy between virtual
ing limit orders generated by the returns G due to the and actual price impact could be explained. Whether or
anticorrelation between returns and limit orders. not our original choice of t0 ¼ 30 min is the correct one in
The density of these additional orders is described by light of the arguments related to long-range correlations
the conditional expectation value in the market order flow, can only be examined by a more
detailed analysis beyond the scope of this paper. Our
$c ð%i , GÞ ¼ hQadd canc add canc
t0 ð%i Þ ' Qt0 ð%i ÞiG ' hQt0 ð%i Þ ' Qt0 ð%i Þi: qualitative conclusion that the virtual price impact is
ð9Þ strongly reduced by limit orders is not affected by this
quantitative issue.
Here, Qadd
t0 ð%i Þ is the number of limit orders added to We note that a reduction of ‘bare’ price impact by
the book at a depth % i within the time interval ½t, t þ t0 #. liquidity providers was recently postulated in Bouchaud
We find that $c ð%i , GÞ approximately saturates for et al. (2003) in order to reconcile the strong auto-
t0 * 30 min. correlations of market orders with the uncorrelated ran-
In order to quantify the effect of this additional order dom walk of returns, and that Lillo and Farmer (2003)
flow, we must specify in which way to treat the influence explain the uncorrelated nature of returns by liquidity
of the extended time correlations of order flow. When fluctuations.
taking into account only limit orders placed in the present In summary, we find that the virtual price impact
time interval of five minutes, one underestimates the effect function as calculated from the average order book is
of correlated limit orders for the following reason: part convex and increases much faster than the concave price
of the market order volume placed in the current five impact function for market orders. This difference can
minute interval is predictable in the sense that it can be be explained by taking into account dynamical properties
explained by long correlations in the order flow of the order book, i.e. the average net order flow and the
(Bouchaud et al. 2003, Lillo and Farmer 2003). The strong anticorrelation between returns and limit order
correlations in the limit order flow which extend over flow. This anticorrelation leads to an additional influx
more than five minutes cancel the price impact of the of limit orders as a reaction to price changes, which
364 P. Weber and B. Rosenow

reduces the price impact of market orders. Including these Gabaix, X., Gopikrishnan, P., Plerou, V. and Stanley, H.E.,
dynamical effects, we quantitatively model the price A theory of power-law distributions in financial market
fluctuations. Nature, 2003, 423, 267–270.
impact of market orders. Gallant, A.-R., Rossi, P.E. and Tauchen, G., Stock prices
and volume. Rev. Financ. Stud., 1992, 5, 199–242.
Hasbrouck, J., Measuring the information content of stock
Acknowledgements trades. J. Finance, 1991, 46, 179–207.
Hausman, J., Lo, A. and MacKinlay, C., An ordered probit
analysis of transaction stock prices. J. Financ. Econ., 1992,
We thank J.-P. Bouchaud, M. Potters and D. Stauffer
31, 319–379.
for helpful discussions. Hopman, C., Are supply and demand driving stock prices?
MIT Working Paper, December 2002 (MIT).
Kempf, A. and Korn, O., Market depth and order size. J. Financ.
References Mark., 1999, 2, 29–48.
Lillo, F. and Farmer, J.D., The long memory of the efficient
market. Stud. Nonlinear Dyn. Econometrics, 2004, 8(3).
Bouchaud, J.-P., Gefen, Y., Potters, M. and Wyart, M.,
Lillo, F., Farmer, J.D. and Mantegna, R.N., Master curve
Fluctuations and response in financial markets: the subtle nat-
for price-impact function. Nature, 2003, 421, 129.
ure of ‘random’ price changes. eprint cond-mat/0307332, 2003.
Bouchaud, J.-P., Mézard, M. and Potters, M., Statistical Maslov, S. and Mills, M., Price fluctuations from the order book
properties of stock order books: empirical results and models. perspective—empirical facts and a simple model. Physica A,
Quant. Finance, 2002, 2, 251–256. 2001, 299, 234–246.
Challet, D. and Stinchcombe, R., Analysing and modelling Plerou, V., Gopikrishnan, P., Gabaix, X. and Stanley, H.E.,
1þ1d markets. Physica A, 2001, 300, 287–299 (eprint Quantifying stock price response to demand fluctuations.
cond-mat/0106114). Phys. Rev. E, 2002, 66, 027104[1]–027104[4].
Chen, G.-M., Firth, M. and Rui, O.M., The dynamic relation Rosenow, B., Fluctuations and market friction in
between stock returns, trading volume, and volatility. financial trading. Int. J. Mod. Phys. C, 2002, 13, 419–425.
Financ. Rev., 2001, 38, 153–174. Potters, M. and Bouchaud, J.-P., More statistical properties of
Coppejans, M.T., Domowitz, I.H. and Madhavan, A., Liquidity order books and price impact. Physica A, 2003, 324, 133–140.
in an Automated Auction, 21 December 2001, AFA 2002 Sandas, P., Adverse selection and competitiv market making:
Atlanta Meetings, http://papers.ssrn.com/sol3/papers.cfm? empirical evidence from a limit order market. Rev. Financ.
abstract_id¼295765. Stud., 2001, 14, 705–734.
Daniels, M.G., Farmer, J.D., Gillemot, L., Iori, G. and Takayasu, H. (Ed.), Empirical Analysis of Financial Fluctuations,
Smith, E., Quantitative model of price diffusion and market 2002 (Springer: Berlin).
friction based on trading as a mechanistic random process. Weber, P. and Rosenow, B., Large stock price changes: volume
Phys. Rev. Lett., 2003, 90, 108102[1]–108102[4]. or liquidity? preprint cond-mat/0401132, 2004.
Evans, M.D. and Lyons, R.K., Order flow and exchange rate Zhang, Y.-C., Toward a theory of marginally efficient markets.
dynamics. J. Political Economy, 2002, 110, 170–180. Physica A, 1999, 269, 30.