HAWKES MODEL FOR PRICE AND TRADES HIGH-FREQUENCY

DYNAMICS
EMMANUEL BACRY AND JEAN-FRANC
OIS MUZY
Abstract. We introduce a multivariate Hawkes process that accounts for the dynamics of
market prices through the impact of market order arrivals at microstructural level. Our model is a
point process mainly characterized by 4 kernels associated with respectively the trade arrival selfexcitation, the price changes mean reversion the impact of trade arrivals on price variations and
the feedback of price changes on trading activity. It allows one to account for both stylized facts
of market prices microstructure (including random time arrival of price moves, discrete price grid,
high frequency mean reversion, correlation functions behavior at various time scales) and the stylized
facts of market impact (mainly the concave-square-root-like/relaxation characteristic shape of the
market impact of a meta-order). Moreover, it allows one to estimate the entire market impact profile
from anonymous market data. We show that these kernels can be empirically estimated from the
empirical conditional mean intensities. We provide numerical examples, application to real data and
comparisons to former approaches.
Key words. Market Impact, Market Microstructure, Point-Processes
AMS subject classifications. 91B24, 60G55

1. Introduction. Market impact modeling (i.e, the influence of market orders

on forthcoming prices) is a longstanding problem in market microstructure literature
and is obviously of great interest for practitioners (see e.g., [6] for a recent review).
Even if there are various ways to define and estimate the impact associated with
an order or a meta-order1 , a large number of empirical results have been obtained
recently. The theory of market price formation and the relationship between the order
flow and price changes has made significant progress during the last decade [7]. Many
empirical studies have provided evidence that the price impact has, to many respects,
some universal properties and is the main source of price variations. This corroborates
the picture of an endogenous nature of price fluctuations that contrasts with the
classical scenario according to which an exogenous flow of information leads the
prices towards a true fondamental value [7].
We will not review in details all these results but simply recall the point of view of
Bouchaud et al [6]. These authors propose a model of price fluctuations by generalizing
Kyles pioneering approach [18] according to which the price is written (up to a noise
term) as the result of the impact of all trades. If n stands for the trading time,
Bouchaud et al. model is written as follows [8, 7]:
pn =

jn

G(n j)j + j

(1.1)

where j is a white noise while j = j f (vj ) with j = +1 (resp. j = 1) if a trade

occurs at the best ask (resp. at best bid) and f (vj ) describes the volume dependence
of a single trade impact. The function G(n j) accounts for the temporal dependence
of a market order impact.
CMAP,
UMR
7641
CNRS,
Ecole
Polytechnique,
91128
Palaiseau,
France
([email protected])
SPE, UMR 6134 CNRS, Universit
e de Corse, 20250 Corte, France ([email protected])
1 One usually refers to a meta-order as a set of orders corresponding to a fragmentation of a single
large volume order in several successive smaller executions

2
Even if models like (1.1) represent a real breakthrough in the understanding
of price dynamics they have many drawbacks. First, the nature and the status of
the noise j is not well defined. More importantly, these models involve discrete
events (through the trading or event time) only defined at a microstructural level
though they intend to represent some coarse version of prices (indeed, the price pn in
the previous equation can take arbitrary continuous values, moreover, calibrating its
volatility at any scale involves some additional parameter, ...). Moreover, these models
cannot account for real time (i.e physical time) dynamics or real time aggregation
properties of price fluctuations. In that respect, they are not that easy to be used in
real high frequency applications such as optimized execution. To make it short, though
being defined at finest time scales, they cannot account for the main microstructure
properties of price variations related to their discrete nature: prices live on tick grids
and jump at discrete random times.
Our aim in this paper is mainly to define continuous time version of the market impact price model discussed previously. For that purpose, point processes [9]
provide a natural framework. Let us not that point processes have been involved in
many studies in high frequency finance from the famous zero-intelligence order-book
models [24, 11] to models for trade [16, 4] or book events [21, 10] irregular arrivals.
In a recent series of papers [3, 2, 1], we have shown that self-excited point (Hawkes)
processes can be pertinent to model the microstructure of the price and in particular
to reproduce the shape of the signature plot and the Epps effect. Our goal is to extend
this framework in order to account for the market impact of market orders. In that
respect, the main ideas proposed in refs. [8, 13, 19] can be reconsidered within the
more realistic framework of point processes where correlations and impact are interpreted as cross and self excitations mechanisms acting on the conditional intensities
of Poisson processes. This allows us to make a step towards the definition of a faithful
model of price microstructure that accounts for most recent empirical findings on the
liquidity dynamical properties and to uncover new features.
The paper is organized as follows. In Section 2 we show how market order impact
can be naturally accounted within the class of multivariate Hawkes processes. Our
main model for price microstructure and market impact is presented and its stability
is studied. Numerical simulations of the model are presented. In Section 3, the
microstructure of price and market order flows are studied through the covariance
matrix of the process. Our results are illustrated on numerical simulations. An
extension of the model that accounts for labeled agents is defined in Section 4 and
results on market impact are presented in Section 5 including an explanation on how
the newly defined framework allows one to estimate the market impact profile from non
labeled data. Section 6 shows how kernels defining the dynamics of trade occurrences
and price variations can be nonparametrically estimated. All the theoretical results
obtained in the previous Sections are illustrated in Section 7 when applied on various
high frequency future data. It allows one to reveal the different dynamics involved
in price movements, market order flows and market impact. Intraday seasonalities
are shown to be taken care in a particularly simple way. We also discuss, on a semiqualitative ground, how the market efficiency can be compatible with the observed
long-range correlation in supply and demand without any parameter adjustment.
Conclusion and prospects are reported in Section 8 while technical computations are
provided in Appendices.
2. Hawkes based model for market microstructure.

3
2.1. Definition of the model. As recalled in the introduction, in order to
define a realistic microstructure price model while accounting for the impact of market
orders, the framework of multivariate self-excited point processes is well suited. A
natural approach is to associate a point process to each set of events one wants to
describe. We choose to consider all market order events and all mid-price change
events.
Let us point out that we will not take into account the volumes associated to each
market orders. Though this can be basically done within the framework of marked the
point processes, it would necessitate cumbersome notations and make the estimation
much more difficult. This issue be discussed briefly in Section 8 and addressed in a
forthcoming work.
2.1.1. Market orders and price changes as a 4 dimensional point process. The arrivals of the market orders are represented by a two dimensional counting
process
Tt =

Tt
Tt+

(2.1)

representing cumulated number of market orders arrived before time t at the best ask
(Tt+ ) and at the best bid (Tt ). Each time there is a market order, either T + or T
jumps up by 1. We suppose that the trade process2 T is a counting process that is
fully defined by
T t the conditional intensity vector of the process Tt at t.
The price is represented in the same way. Let Xt represents a proxy of the price
at high-frequency (e.g., mid-price). As in refs. [3, 2, 1] we write
Xt = Nt+ Nt

(2.2)

where Nt+ (resp. Nt ) represents the number of upward (resp. downward) price jumps
at time t 3 . Thus, each time the price goes up (resp. down), N + (resp. N ) jumps
up by 1. We set
Nt =

Nt
Nt+

(2.3)

As for the trade process, Nt is fully defined by

N t the conditional intensity vector of the process Nt at t.
The 4 dimensional counting process is then naturally defined as

Pt =

Tt
Nt

Tt
Tt+

=
Nt .
Nt+

(2.4)

2 In the following, a trade will refer to the execution of a given market order (which might
involve several counterparts). Thus the process T will be referred indifferently to as the market
order arrival process or the trade arrival process.
3 Let us point out that, in this model, we do not take into account the size of the upward or
downward jumps of the price. We just take into account the direction of the price move, +1 (reps.
1) for any upward (reps. downward) jumps

4
as well as its associated conditional intensity vector

T
 T  T t+
t

= N t
t =
N t
t
+
N t

(2.5)

2.1.2. The Model. Basically, the model consists in considering that the 4dimensional counting process Pt is a Hawkes process [14, 15]. The structure of a
Hawkes process allows one to take into account the influence of any component in Pt
on any component of t . In its general form the model is represented by the following
equation4
t = M + dPt .

(2.6)

where t is a 4 4 matrix whose elements are causal positive functions (by causal
we mean functions supported by R+ ). Moreover, we used the matrix convolution
notation,
Z
B dPt =
Bts dPs ,
R

where Bts dPs refers to the regular matrix product. M accounts for the exogenous
intensity of the trades, it has the form

(2.7)
M =
0
0
since, by symmetry we assume that the exogenous intensity of T + and T are equal
while mid-price jumps are only caused by the endogenous dynamics.
The matrix or any of its sub-matrices (or element) are often referred to as
Hawkes kernels. Each element describes the influence of a component over another
component. Thus, it is natural to decompose the 4 4 kernel t in four 2 2 matrices
in the following way
 T

t F
t
t =
(2.8)
It N
t
where
T (influence of T on T ) : accounts for the trade correlations (e.g., splitting,
herding, . . . ).
I (influence of T on N ) : accounts for the impact of a single trade on the
price
N (influence of N on N ) : accounts for the influence of past changes in
price on future changes in price (due to cancel and limit orders only, since
changes in price due to market orders are explicitly taken into account by I )
4 Let us point out that Section 4 will introduce a generalization of this model including labeled
trades.

5
F (influence of N on T ) : accounts for feedback influence of the price moves
on the trades.
If we account for the obvious symmetries between bid-ask sides for trades and up-down
directions for price jumps these matrices are naturally written as:
 T,s

 I,s

t
T,c
t
I,c
I
t
t
Tt =
,

=
(2.9)
t
T,c
T,s
I,c
I,s
t
t
t
t
and
N
t =

N,s
t
N,c
t

N,c
t
N,s
t

, F
t =

F,s
t
F,c
t

F,c
t
F,s
t

(2.10)

where all ?,?

are causal functions and the upperscripts s and c stand for self and
t
cross influences of the Poisson rates (we use the same convention which was initially
introduced in [3] for N ). Thus, for instance, on the one hand, I,s accounts for the
influence of the past buying (resp. selling) market orders on the intensity of the
future upward (reps. downward) price jumps. On the other hand, I,c quantifies the
influence of the past buying (resp. selling) market orders on the intensity of the future
downward (resp. upward) price jumps.
Remark 1 : All these 2 2 matrices commute since they diagonalize in the same
basis (independently of t). Their eigenvalues are the sum (resp. the difference) of the
self term with the cross term. This property will be used all along the paper. Most
of the computations will be made after diagonalizing all the matrices.
2.1.3. The Impulsive impact kernel model or how to deal with simultaneous jumps in the price and trade processes. It is important to point out
that a buying market order that eats up the whole volume sitting at best ask results
in an instantaneous change in the mid-price. From our model point of view, it
would mean that T and N have simultaneous jumps with a non zero probability. It is
clearly not allowed as is by the model. However, from a numerical point of view, this
can be simulated by just considering that the jump in the price takes place within a
very small time interval (e.g., of width 1ms which is the resolution level of our data)
after the market order has arrived. There is no ambiguity in the direction of the
causality : it is a market order that makes the price change and not the other way
around. This would result in an impact kernel I,s which is impulsive, i.e., localized
around 0, actually close to a Dirac distribution t .
From a practical numerical point of view, choosing I,s to be a Dirac distribution
is fairly easy. It basically amounts to considering that it is a positive function of given
L1 norm I and with a support t of the order of a few milliseconds. Let us point
out that it means that, the price increment between the moment of the trade and t
milliseconds afterwards, follows a Poisson law whose parameter is I. This actually
allows price jumps (spread over a few milliseconds) of several ticks (greater than 1).
Of course, from a mathematical point of view, this is not that simple. The limit
t 0 has to be defined properly. This will be rigorously defined and extensively discussed in a future work and is out of the scope of this paper. The practical approach
described above and the fact that we can formally replace, in all the computations,
I,s by a It is far enough for our purpose.
It is clear that we expect to find an impulsive component in I,s when estimating
on real data. Though, a priori, we do expect also a non singular component that

6
could have a large support (e.g., when the marker order eats up only a part of the
volume sitting at best ask), we will see in estimations that most of the energy of I,s
is localized around 0. Moreover, we will find that I,c is close to 0.
All these remarks will lead us to study a particularly interesting case of the
previously defined model for which
I,s = It and I,c = 0.
Consequently
I = It I, where I refers to the identity matrix.
This model will be referred in the following as the Impulsive Impact Kernel model.
Before moving on and study the conditions for our model to be well defined, we
need to introduce some notations that will be used all along the paper.
2.1.4. Notations. Let us introduce the following notations:
Notations 1. If ft is a function fbz refers to the Laplace transform of this
function, i.e.,
Z
fbz = eizt ft dt.

t refers to the Dirac distribution, consequently

bz = 1.

Moreover, we will use the convenient convention (which holds in the Laplace domain) : t t = t .
The L1 norm of f is referred to as:
Z
||f || = |ft |dt.
Thus
if

t, ft 0 then ||f || =

ft dt = fb0 .

We extend these notations to matrix of functions. Thus, if Ft is a matrix whose

element are functions of t, let Fz denote the matrix whose elements are the Laplace
transform of the elements of Ft . Following this notation, we note
!
bT
bF

z
z
bz =
.
(2.11)

bI
bN

z
z
Notations 2. If M is a matrix, M refers to the matrix M whose each element
has been replaced by its conjugate and M refers to the hermitian conjugate matrix of
M.
Notations 3. Whenever t is a stationary process, we will use the notation
T

= E(t ) =
N
N
where

T = E(T t ) = E(T t )

and

N = E(N t ) = E(N t ).

7
Let us point out that the fact that the mean intensities are equal is due to the
symmetries of the kernels in (2.9) and (2.10).
Notations 4. We define the kernels imbalance :
T = T,s T,c and bT = bT,s bT,c
I = I,s I,c and bI = bI,s bI,c
F = F,s F,c and bF = bF,s bF,c
N = N,s N,c and bN = bN,s bN,c
Let us point out that since the kernels are all positive functions, one has, replacing ?
by T , N , I or F :

and consequently

?,s
?,c
b?,c
b?,s
0 = || || and 0 = || ||,

b ?0 = b?,s b?,c = ||?,s || ||?,c ||

0
0
2.2. Stability condition - Stationarity of the price increments. The process Pt defined in Section 2.1.2 is well defined as long as the matrix t is locally
integrable on R+ . Hawkes, in his original papers [14, 15], has formalized the necessary and sufficient condition for the previously introduced model (2.6) to be stable :
the matrix made of the L1 norm of the elements of should have eigenvalues whose
modulus are strictly smaller than 1.
This condition can be expressed in terms of conditions on the L1 norm of the different
kernels :
Proposition 2.1. (Stability Condition) The hawkes process Pt is stable if
and only if the following condition holds :
b 0 have a modulus strictly smaller than 1.
(H) The eigenvalues of the matrix
In that case, Pt has stationary increments and the process t is strictly stationary.
Moreover (H) holds if and only if
c+ < (1 a+ )(1 b+ )

and

a+ , b+ < 1,

(2.12)

where

bT,c
a+ = bT,s
0 + 0 ,
N,s
and
b+ = b0 + bN,c
0
bF,c )(bI,s + bI,c ).
c+ = (bF,s
+

0
0
0
0
Moreover (2.12) implies that

bF bI
bT
bN
bT0 bN
0 1 < 0 0 < (1 0 )(1 0 ),

(2.13)

where we used Notations 4 The proof is in Section 9.1.

Let us point out that in the case there is no feedback of the price jumps on the
trades, i.e., F = 0 (or c+ = 0), then the stability condition (2.12) is equivalent to
a+ < 1 and b+ < 1, i.e., ||T,s || + ||T,c || < 1 and ||N,s || + ||N,c || < 1.
The mean intensity vector is given by the following Proposition.
Proposition 2.2. (Mean Intensity) We suppose that (H) holds (i.e., (2.12)).
Then
b 0 )1 M.
= E(t ) = (I

(2.14)

8
This can be written as


T
T

N
N

and

where v =

1
1

b 0 )(I
b N )v
= (I + D
0


b 0 )
bIv
= (I + D
0

(2.15)

(2.16)

and where Dt is defined by its Laplace transform

b z = ((I
b T )(I
bN )
bF
b I 1 I.
D
z z)
z
z

(2.17)

Proof : The proof is basically an adaptation of a proof previously presented in [2].

Let the martingale dZt be defined as
dZt = dPt t dt.
Using (2.6), we get
t = M + dPt = M + dZt + t dt.

(2.18)

(I ) t = M + dZt .

(2.19)

t = (t I + ) M + (t I + ) dZt ,

(2.20)

Thus

Consequently,

where is defined by
bz =
b z (I
b z )1 .

Taking the expectation, we get (2.14). Moreover, we have

!
bT
bF
I

b=
I
bI
bN

Using Remark 1 a the end of Section 2.1.2, one can easily check that
!
bF
bN

I
1
b
b
(I ) = (I + D)
bI
bT

(2.21)

(2.22)

where Dt is defined by (2.17). The Equations (2.15) and (2.16) are direct consequences
of this last equation combined with (2.14).
In the following we will always consider that (H) holds, i.e., that (2.12)
holds.

9
2.3. Numerical simulations. In order to perform numerical simulations of
Hawkes models, various methods have been proposed. We chose to use a thinning
algorithm (as proposed, e.g., in [22]) that consists in generating on [0, tmax ] an homo
geneous Poisson process with an intensity M > supt[0,tmax ] (Tt , N
t ). A thinning
procedure is then applied, each jump being accepted or rejected according to the ac

tual value of Tt or N
. In order to illustrate the 4-dimensional process we chose
t
to display only the price path
Xt = Nt+ Nt

(2.23)

and the cumulated trade process path as defined by

Ut = Tt+ Tt .

(2.24)

In Figure 2.1, we show an example of sample paths of both Xt and Ut on a few minutes
time interval. All the involved kernels are exponentials. Some microstructure stylized
facts of the price can be clearly identified directly on the plot : price moves arrive
at random times, price moves on a discrete grid and is strongly mean reverting (see
beginning of next section). In the large time limit, one can show that these processes
converge to correlated Brownian motions (see [2] or Section 3.3). This is illustrated
in Fig. 2.2 where the paths are represented over a wider time window (almost 2
hours). As discussed in Section 7.3, since we choose T,c = 0 and N,s = 0, the
small time increments of Ut remain correlated while the price increments correlations
almost vanish. This can be observed in Fig. 2.2 where the path of Ut appears to be
smoother than the path of Nt .

Fig. 2.1. Example of sample paths for the cumulated trade (2.24) (a) and price (2.23) (b)
processes of the model with exponential kernels. We chose T,c = N,s = I,c = R,s = 0,
T,s = 0.03 e5t/100 , N,c = 0.05 et/10 , I,s = 25 e100t and R,c = 0.1 et/2 .

In the next section we study the microstructure properties of both the price and
the market order flow.

10

Fig. 2.2. Example of sample paths for the cumulated trade (2.24) (a) and price (2.23) (b)
processes of the same model as in Fig. 2.1 over a wider range of time. One does not see the discrete
nature of time variations anymore. Ut and Xt appear as correlated processes.

3. Microstructure of price and market order flow. The model we introduced above can be seen as a generalization of the price microstructure model previously introduced in [3]. Indeed, the 2-dimensional model defined in [3] can somewhat
be seen as the projection on the price components Nt of the model defined in Section 2.1.2. Thus, it easy to show that it will account for all the price microstructure
stylized facts already accounted for by the model in [3]. This includes the characteristic decreasing shape of the mean signature plot E((Xt X0 )2 )/t (which explains
why estimating the diffusing variance using high frequency quadratic variations leads
to a systematic positive bias). As explained in [3], this effect is due to the high frequency strong mean reversion observed in real prices and can easily be modeled by
choosing kernels such that ||N,c || ||N,s ||. We refer the reader to [3] for all the
discussions concerning the signature plot, the problem of variance estimation using
high frequency data and the link with mean reversion of price time-series.
Thus, in this Section, we shall mainly focus one price correlations and market
order correlations and how they relate one to the other.
3.1. Price and Trade covariance function. Following [2], we define the covariance matrix of the normalized process at scale h and lag by
v(h) = h1 Cov (Pt+h+ Pt+ , Pt+h Pt ) ,

(3.1)

where we normalized by h in order to avoid a trivial scale dependence. Let us note

that, since the increments of Pt are stationary (we suppose that (2.12) is satisfied),
the previous definition does not depend on t. Thus, it can be rewritten as
!
Z h
Z +h
1

(h)
dPs h) .
(3.2)
dPs h)(
v = E (
h
0

11
(h)

In [1, 2], it is proven that the the Laplace transform of v can be expressed as a
(h)
(h)
+
function of the Laplace transform of g (where gt = (1 |t|
h ) ) and of :
b z )1 (I
b )1
vbz(h) = b
gz(h) (I
z

(3.3)

where is the diagonal matrix

T I
0

0
N I

(3.4)

From this result, one can easily deduce an analytical formula for the price autocovariance function.
N,(h)
Proposition 3.1. (Price auto-covariance) Let C
be the normalized autocovariance of the price increment :
CN,(h) =

1
E((Xt+h Xt )(Xt+ +h Xt+ )),
h

(3.5)

then
(h)
gz (T |bI |2 + N |1 bT |2 )
bzN,(h) =  2b
C
2


(1 bT )(1 bN ) bI bF 

(3.6)

Before proving this Proposition, we first need the following Lemma, which is a direct
consequence of (2.17) and (2.22).
Lemma 1.
!
bT E
bF
E
1
1

z
z
b
b
b
b
(I z ) (I z ) = (I + Dz )(I + Dz )
(3.7)
bzI E
bzN
E
where

b T = T (I
b N )(I
b N ) + N
bF
b F ,
E
z
b F = T
b I (I
b N ) + N
b F (I
b T ),
E
z
bzI = T (I
b N )
b I + N (I
b T )
bF ,
E
bzN = T
bI
b I + N (I
b T )(I
b T ) .
E

Proof of Proposition : Using the symmetries of all the kernels, and the fact that
Xt = Nt+ Nt , we get
CN,(h) =

2
+
+
+
+

+
+
E((Nt+
+h Nt+ )(Nt+h Nt )) E((Nt+ +h Nt+ )(Nt+h Nt ))
h

Thus, if we define dst and dct such that

b z )(I +
(I + D

b z )E
bzN
D

dbsz
dbcz

dbcz
dbsz

12
we have

Using Remark 1, we get that

which proves (3.6).

b N,(h) = 2b
C
gz(h) (dbsz dbcz )
z

T |bI |2 + N |1 bT |2
b
cs b
cc = 
2


(1 bT )(1 bN ) bI bF 

Using similar computations, one derives an analytical formula for the auto-covariance
of the cumulated trade process (2.24):
T,(h)
Proposition 3.2. (Trade auto-covariance) Let C
be the normalized covariance of the increments of the cumulated trade process Ut defined by (2.24).
CT,(h) =

1
E((Ut+h Ut )(Ut+ +h Ut+ )),
h

(3.8)

then
(h)
gz (T |1 bN |2 + N |bF |2 )
b T,(h) = 2b
C
2
z


(1 bT )(1 bN ) bI bF 

(3.9)

3.2. Numerical simulations. In order to illustrate these results, in Fig. 3.1

we have plotted both theoretical (solid lines) and estimated () correlation functions.
The sample we used for the estimation contains around 300.000 trading events and
corresponds to the same kernels as the ones used for Fig. 2.2. One clearly see that
empirical estimates closely match theoretical expressions. Moreover, one can observe
that the trade increment autocorrelation has an amplitude that is an order of magnitude larger than the price increment covariance. This issue will be addressed in
Section 7.3.
3.3. On the diffusive properties of the model. Let us point out that we
know [2] that a centered d-dimensional Hawkes process Pt diffuses at large scales
(h +) towards a multidimensional gaussian process :
1
b 0 )1 1/2 Wt
(Pht E(Pht )) law (I
h

where Wt is a standard d-dimensional Brownian motion and is defined by (3.4).

This is a very general result that can be applied here to obtain the covariance matrix
of the diffusive limit of our 4-dimensional Hawkes process Pt . Actually, if one is just
interested in the diffusive variance of Pt , one can easily
obtain it directly from the
(h)
definition of the covariance matrix (3.1) (noticing that hv0 corresponds to the
b 0 )1 (I
b )1 . It
variance of 1h Pht ). One gets that the diffusing variance is (I
0
is then straightforward to obtain the diffusive variance of the price Xt = Nt+ Nt :
q
2 T (bI0 )2 + N (1 bT0 )2
X =
(1 bT )(1 bN ) bI bF
0

(Let us point out that (2.13) shows that the denominator is positive).

13

Fig. 3.1. Empirical and theoretical (From Eqs. (3.6) and (3.9)) autoccorelation functions for
the process defined in Fig. 2.2. (a) Autocorrelation of the increments of Ut . (b) Autocorrelation of
the price increments (i.e., the increments of Xt ). In both cases, for the sake of clarity with have
removed the lag = 0.

4. Model with labeled agents.

4.1. Accounting for labeled agents. The trade arrival process T models
anonymous trades sent by anonymous agents : this corresponds to the typical trade
information sent by most exchanges (i.e., the trades are unlabeled, there is no way to
know whether two different trades involves the same agent or not). However, there
are many situations where one has access to some labeled data corresponding to some
specific labeled agents. In this case, it could be very interesting to be able to analyze
the impact of these labeled trades within the same framework. It is naturally the case
of brokers who have labels for all their clients (though some clients may have several
brokers, consequently a given broker might not be able to identify all the trades of a
given client). In general, this is the case of any financial institutions which has access
to the historic of all its own trades.
In our model, we chose, to consider only the case of a single labeled agent that is
sending market orders at some deterministic time. The case where there are several
such agents is a straightforward generalization. The trades arrival of this labeled
agent is represented by a 2d deterministic function
At =

A
t
A+
t

(4.1)

representing the market orders arrival at the best ask (A+ ) and at the best bid (A ).
Again, A+ (resp. A ) jumps upward by 1 as soon as the agent is buying (resp. selling)
(we do not take into account the associated volumes).
In all the following, we consider that the agent is active only on a finite

14
positive time interval, i.e,
T > 0,

Support(dAt ) [0, T [

(4.2)

We are ready now to reformulate the previously introduced model (see Section
2.1.2) taking into account the labeled agent.
4.2. The model. In its most general form the model writes
t = M + dPt + dAt ,

(4.3)

where is a 4 2 matrix. In the same way as we did for , we decompose the kernel
using 2 2 matrices :
 T 

=
(4.4)
I
T (influence of the labeled trades on T ) : accounts for the herding of the
anonymous trades with respect to the labeled trades.
I (influence of the labeled trades on N ) : accounts for the impact of a single
labeled trade on the price. A priori there is no reason not to consider that
the impact of a labeled trade is not the same as the impact of a anonymous
trade, i.e., we will take
I = I

(4.5)

Let us point out that the model introduced in Section 2.1.2 is a particular case (when
At = 0) of this model. In that sense, this model is a more general model.
5. Market impact in the model with labeled agents.
5.1. Computation of the market impact profile. Let us give an analytical
expression for the market impact profile.
Proposition 5.1. (Market impact profile) The expectation of the intensity
vector is given by
E(t ) = (I + ) (M + dAt ),

(5.1)

where is defined by its Laplace transform

bz =
b z (I
b z )1 .

The market impact profile between time 0 and time t of the labeled agent is defined as
M It = E(Xt ) = E(Nt+ Nt ).

(5.2)

M It = ( ) I (A+ A ),

(5.3)

It can be written

where t is defined as
c =1

(1 bT + bT )
(1 bT )(1 bN ) bI bF

(5.4)

15
Proof of (5.1). It is very similar to the proof of (2.14). Let dZt be the martingale
defined as
dZt = dPt t dt.
Using (4.3), we get
t = M + dAt + dPt = M + dAt + Y dZt + Y t dt.

(5.5)

which gives
(I ) t = M + dAt + dZt ,

(5.6)

t = (I + ) (M + dAt ) + (I + ) t dZt .

(5.7)

which is equivalent to

Taking the expectation, we get (5.1).

Proof of (5.4)
Using (2.22), we get
E(N t ) = (I + D) K (M + dAt ).
where K is the 2 4 matrix defined by
K=

I T

Thus
E(dNt+ dNt ) = u(I + D) K

(5.8)



(5.9)

T dAt
I dAt

(5.10)

where u stands for the vector u = (1, 1). Let us recall that we chose I = I . Using
Remark 1, all the 2 2 matrices involved in this equation are commuting since they
are symmetric along both diagonals. Thus one gets
E(dNt+ dNt ) = u(I + D) (I T + T ) I dAt

(5.11)

Thus the market impact is

M It = u(I + D) (I T + T ) I At

(5.12)

Using Remark 1 again, we deduce (5.4)

Let us state a trivial corollary of this last proposition that will be of particular
interest in order to study the permanent impact in the following Section:
Corollary 1. In the case of an impulsive impact kernel (see Section 2.1.3) and
T = 0, the market impact profile of a single buying market order sent at time t = 0

(i.e., dA+
t = t and dAt = 0) is given by
Z t
u du
(5.13)
M It = 1
0

16

Fig. 5.1. Example of two market impact profiles according to Eq. (5.3) with T = 0 and
dAt = T 1 1tT dt. The kernel shapes and the parameters have been chosen to roughly match the
bT . (a)
values observed empirically in Section 7. The two figures only differ by the values of
0
T
T
b
b
0 = 0.45 (b) 0 = 0.95.

where t is defined as
c =1

1 bT
(1 bT )(1 bN ) IbF

(5.14)

In Fig. 5.1 are displayed 2 examples of market impact profiles computed according
to Eq. (5.3) for a meta-order consisting in buying a constant amount of shares during
a period T (i.e. dAt = T 1 dt if t T and dAt = 0 otherwise). The parameters Fig.
5.1(b) correspond to the ones estimated from empirical data as discussed in Section
7 (notably bT0 0.9). One can see that Fig. 5.1(b) reproduces fairly well the
empirical shape measured using labeled database (as, e.g., in [20, 5]): a concave shape
of the profile M It for times t T (the so-called square-root law [13]) and a convex
relaxation towards the permanent impact when t > T . This shape is qualitatively
explaind in Section 7.4.
For illustration purpose, the parameters of Fig. 5.1(a) are the same as the ones
of Fig. 5.1(b) except that they have been tweaked in order to get bT0 < 1/2. The
fact that the asymptotic value of the market impact is larger is explained by point
iii) of the next section.
5.2. Permanent versus non permanent market impact. One important
consequence of (5.14) is that the asymptotic market impact of a single labeled market
buying order at time t = 0 is of the form (assuming an impulsive market impact
kernel) :
M I+ = (1 b0 ).

(5.15)

Thus controlling how permanent is the market impact essentially amounts in controlling how large b0 is where
c =1

1
,
N
b
(1 0 ) IbF /(1 bT0 )

(5.16)

17
thus
M I+ =

1
.
N
b
(1 0 ) IbF /(1 bT0 )

(5.17)

Let us note that, the stability condition of the process (see Proposition 2.1) implies
bT
that |1 bN
0 | < 1 and |1 0 | < 1.
This last equation allows us to identify three different dynamics that can lead
to a decrease of the permanent impact. They all correspond to different ways of
introducing mean reversion into the price :
i) Mean reversion due to microstructure. This case corresponds to the case bN
0
is very negative, i.e., ||N,c || much larger than ||N,s ||. However this effect is
limited by the fact that we know that we cannot go below bN
0 > 1 due to
the stability condition.
ii) Feedback cross influence of the price moves on the trades. This case correF,c
sponds to the case bF
|| large and F,s = 0.
0 is very negative, e.g., ||
This is definitely an effect that is present in real life (see Section 7) : as the
price goes up, traders are sending more and more selling buying orders.
iii) Auto-correlation of the trades . The previous effect i) is even stronger if there
is a strong correlation in the signs of the market orders (i.e., ||bT0 || 1).
Let us point out that Fig. 5.1 shows a clear illustration of this point : the left
market impact curve has a stronger permanent market impact than the one
of the right, this is due to the fact that it corresponds to a smaller ||bT0 ||.
As we will see in Section 7.1, all these effects are present in real data. And they will
all play a part in reducing the asymptotic market impact (see Section 7.4)
5.3. Estimation from non labeled data. Response function versus Market impact function. As already pointed out, most markets do not provide labeled
data. The order flows are made of anonymous orders sent by anonymous agents. A
priori, in that case, the only way of quantifying the impact a given market order
has on the price is to estimate numerically the response function Rt . The response
function is defined as the variation of the price from time 0 to time t knowing there
was a (e.g., buying) trade at time 0. Thus it can be written
Rt = E(Nt+ Nt |dT0+ = 1), for t > 0,

(5.18)

and Rt = 0 otherwise. It can be written as

Z t
Z t
+
+
Rt =
E(dNt |dT0 = 1)
E(dNt |dT0+ = 1), for t > 0,
0

Notice that within the model (1.1) of Bouchaud et al., this response function can be
explicitly related to the bare impact function G(n) and therefore used to estimate its
shape [8]. It is important to understand that this function is fundamentally different
from the market impact profile of a single market order. The market impact isolates
all the market orders of a single agent (e.g., a meta-order) and quantifies what the
impact of these market orders are. In order to compute it, one, a priori, needs to
identify all the market orders of a particular agent. This is not the case of the
response function which is polluted by the impact of all the market orders that are
in the same meta-order as the market order that is under consideration.
A very important consequence of our model (as it will be explained in Section 7.4)
is that our model allows estimation of the market impact profile even if no labeled

18
data are available. It will basically consist in first estimating all the kernels (see
Section 6) and then using the analytical formula (5.4).
Let us point out that one can easily obtain an analytical formula for the response
function using Proposition 9.1 of Section 9.2. We prove that if gt is the 4 4 matrix
defined by gt = {gtij }1i,j4 with
gtij dt = E(dPti |dP0j = 1) ij t i dt,
then
b z )1 (I
b )1 1 I.
gz = (I
b
z

Using Lemma 1, if we define

Rts

E(dNu+ |dT0+

= 1) du and

Rtc

t
0

E(dNu |dT0+ = 1) N du,

and if we define dRts = rts dt, dRtc = rtc dt, then,

 s 
rbz
b +D
b )(E
bI ) /T
= (I + D)(I
rbzc

Using Remark 1, we get

rbz = rbzs rbzc =

(I bN )bI + (N /T )(I bT )bF

,

2


T
N
I
F
b
b
b
b

)(1

(1


In the case of an impulsive impact kernel, it gives

Rt = I(1

u du),

t > 0

where t is defined by
bN ) + (N /T )(1 bT )bF I 1
d = 1 (1


2
z


(1 bT )(1 bN ) IbF 

6. Non parametric estimation of the kernel functions. In this section we

provide a new method to estimate the shape of the kernels involved in the definition
of the model5 . A former non parametric estimation method has been introduced in
[1]. This method relies on the expression (3.3) and mainly consists in extracting the
square root of the autocorrelation matrix. However, in order to do so univocally, one
has to suppose that the process is fully symmetric and in particular that T and N
have the same laws. This assumption is clearly unrealistic, therefore the method of [1]
is not suited to estimate the process defined in Section 2. For that reason we introduce
5 Notice that during the completion of this paper, we have been aware of a work by M. Kirchner
introducing a non parametric estimation method very similar to the one presented in this section.
Let us point out that he has performed a comprehensive statistical study of the main properties of
this estimator [17].

19

Fig. 6.1. Numerical estimation of the Hawkes kernel matrix by solving the Fredholm equation
(9.22) for an exponential model. Estimations are represented by symbols () for the self kernels
and () for the cross kernels. Exact exponential kernels are represented by the solid lines. The
model corresponds to T,c = N,s = I,c = R,c = 0, T,s = 0.04 et/5 , N,c = 0.02 et/5 ,
I,s = 0.02 et/20 and R,s = 0.06 et/10 .

an alternative method that does not require any symmetry hypothesis. This method
relies on the Proposition 9.2 of Annex 9.2 (Eq. (9.22)):
gt = (I + gt ),

t > 0.

where gt is the matrix of conditional expectations defined in Eq. (9.17). This above
equation is a Fredholm equation of 2nd kind. Since gt can be easily estimated from
empirical data, the matrix can be obtained as a numerical solution of the Fredholm system. We thus implemented a classical Nystrom method that amounts to
approximate the integrals by a quadrature and solve a linear system [23]. In order
to illustrate the method, we have reported in Fig. 6.1 the estimated kernels in the
case when all functions t have an exponential shape. The sample used has typically
3 105 trade events and 1.5 105 price change events. One can see that for such sample
length, numerical estimates and the real kernels are close enough to determine a good
fit of the latter. Let us note that we have checked that the method is reliable for
various examples of kernels like exponential, power-laws or constant over an interval.
In order to investigate the efficiency of the method, one can evaluate the estimation error behavior as a function of the number of events. This error can be defined as
the supremum of the mean square error associated with each kernel: if Ne stands for
the estimated kernel matrix for Ne trading events, then one can consider the following

20

Fig. 6.2. Estimation squared error as a function of the number of events for the model of Fig.
6.1 in log-log scale. The slope is 1.

square error:


e2 (Ne ) = E sup ||ij ij ||2
i,j

A log-log plot of E 2 (Ne ) as a function of Ne is reported in Fig. 6.2 where E 2 (Ne )

has been estimated, for each Ne , using 500 trials of the model. The measured slope
is close to 1 in agreement with the standard behavior of the estimation error:
E0
e(Ne )
.
Ne
A comprehensive study of the statistical properties of the method is out of the scope
of the present paper and will be addressed in a forthcoming work.
7. Application to real data. In this section we apply the main theoretical results we obtained in the previous sections to real data. We consider intraday data associated with the most liquid maturity of EuroStoxx (FSXE) and Euro-Bund (FGBL)
future contracts. The data we used are trades at best bid/ask provided by QuantHouse Trading Solution 6 . Each time series covers a period of 800 trading days going
from 2009 May to 2012 September. The typical number of trades is around 40.000
per day while the number of mid price changes is 20.000 per day.
7.1. Kernel matrix estimation. In this section we apply the non parametric kernel estimation algorithm presented in Section 6 to our data. Since intraday
statistics are well known to be non stationary, estimations are based on a 2 hours
liquid period from 9 a.m. to 11 a.m. (GMT). The results of the kernel estimation of
EuroStoxx are displayed in Figs.7.1, 7.2 and 7.3.
One first sees that T,c is small as compared to T,s . This appears more clearly
on the zoom presented in Fig. 7.2 where we have removed the first point in order to
get a finer scale. The fact that T,s is larger than T,c confirms the well known strong
correlation observed in trade signs (see Section 7.3). The mid-price dynamics is well
known to be mainly mean-reverting [3]. This is confirmed by Fig. 7.3 where a greater
6 http://www.quanthouse.com

21

Fig. 7.1. Numerical estimation of the Hawkes kernels of EuroStoxx in the intraday slot [9 a.m.,
11 a.m.]

Fig. 7.2. Numerical estimation of the Hawkes kernels T,s (a) and T,c (b) in the intraday slot
[9 a.m., 11 a.m.] for EuroStoxx. This is a zoom of Fig. 7.1. One sees that T,s slowly decreases
and is very large as compared to T,c .

intensity of the kernel N,c as compared to N,s can be observed. Both kernels T,s
and N,c are found to decrease as a power-law:
t = (c + t)
with an exponent 1.2 for T,s and 1.1 for N,c . The cut-off c insures the
finiteness of the L1 norm of and is found to be smaller than 102 s while the values
of are typically between 0.05 and 0.1. This power-law behavior is illustrated in
Fig. 7.4 where we have reported in log-log scale the estimates T,s and N,c for both
EuroStoxx ( ) and Euro-Bund (solid lines). Let us notice that power-law behavior of
Hawkes kernels have already been observed in [1] using a simple 2D Hawkes model of

22

Fig. 7.3. Numerical estimation of the Hawkes kernels N,s (a) and N,c (b) in the intraday
slot [9 a.m., 11 a.m.] for EuroStoxx. This is a zoom of Fig. 7.1. One sees that N,c is larger
than N,s .

Fig. 7.4. Scaling of Hawkes kernels T,s and N,c . The kernels are represented in double
logarithmic representation for Eurostoxx () and Euro-Bund (solid lines) estimates in the intraday
slot [9 a.m., 11 a.m.]. (a) T,s (b) N,c . These plots suggest that the kernels have the same
power-law behavior for both EuroStoxx and Euro-Bund.

price jumps. This important property suggests the existence of some scale invariance
properties underlying the order book dynamics. Let us point out that Fig. 7.4 also
suggests that the kernels are the same for EuroStoxx and Euro-Bund.
As far as the impact is concerned, one can see in Fig. 7.1 that only I,s is
significant and well modeled by an impulsive kernel (see Section 2.1.3). This confirms
the fact that a buy (resp. sell) market order increases the probability of an upward
(resp. downward) movement of the mid-price but only within a very small time
interval after the trade. The feed-back kernel that accounts for the influence of a
price move on the trading intensity is also found to be well localized but only the
cross term is non negligible. It seems that an upward (resp. downward) move of the
mid price triggers an higher trading activity on the bid side (resp. on the ask side).
7.2. Introducing a model with intraday seasonalities. In order to check
the stationarity of the kernels, we performed various estimations on different intraday
slices of 2 hours. The results for T,s and N,c are reported in Fig. 7.5(a) and Fig.
7.5(b) in the case of EuroStoxx. One can see that the shape of the kernels does not

23

Fig. 7.5. Numerical estimation of the Hawkes kernel matrix for various 2 hours slices during
the day for EuroStoxx. In (a) and (b) 7 estimated kernels are represented in doubly logarithmic
scale corresponding to the 7 slots: [8 a.m., 10 a.m.], [9 a.m., 11 a.m.], [11 a.m., 1 p.m], [12 a.m.,
2 p.m.], [1 p.m., 3 p.m], [2 p.m., 4 p.m] and [3 p.m., 5 p.m.]. In (a) T,s and in (b) N,c . One
sees that the kernel shapes are remarkably stable power-laws across intraday time periods. In (c) we
reported the estimation of the constant rate as a function of the intraday time. One recognizes
the classical U-shaped behavior.

depend on the intraday market activity and are remarkably stable throughout the day.
An estimation of the norms of all kernels allows one to estimate the rate through
(2.15) or (2.16). In Fig. 7.5(c), we see that this parameter follows the classical Ushaped intraday curve. It thus appears that, in the model, the exogenous intensity of
trades fully accounts for the intraday modulation of market activity.
It is therefore natural to consider the following model with seasonal variations
that generalizes, in a particularly simple way, the definition (2.6):
t = Mt + dPt ,

(7.1)

with

t
t

Mt =
0 .
0

(7.2)

where t is a U-Shaped 1-day periodic function. Let us point out that this is particularly elegant and simple way of accounting for an a priori pretty complex phenomenon.
7.3. Trades correlations, Price correlations and Efficiency. Few years
ago, Bouchaud et al. [8] and Lillo and Farmer [19] independently brought evidence

24
that the market order flow is a long memory process. By studying the empirical
correlation function (in trading time) of the signs of trades they have shown, for
different markets and assets that:
C(n) n

(7.3)

where n is the lag in number of trades (i.e., using trading time). This scaling law
remains valid over 2 or 3 decades and the exponent is in the interval [0.4, 0.7]. The
aim of impact price models as described in the introductory section (Eq. (1.1)) was
to solve this long-memory puzzle. Indeed, since trades impact prices, if trades are
long-range correlated, in order to maintain efficiency, the price have to respond to
market order through a long-memory kernel. This is precisely the meaning of the
kernel G(j) in (1.1). Bouchaud et al. [8] have shown that, provided the specific shape
of G(j) is adjusted as respect to the behavior (7.3), trade correlations impact on
prices can be canceled (see also [7] for an interpretation of this price model in terms
of prediction error or surprise).
In this section, we explain how this issue can be addressed within our model. It
is important to point out that, as we will see, the apparent long-range correlation of
the trade signs and the price efficiency will be a consequence of 3 empirical findings
(cf Section 7.1):
T t is power-law
bT0 is close to 1, i.e., it almost saturates the stability condition bT0 < 1
bN
0 < 0, i.e., the price at high frequency is mainly mean-reverting.

About the long-range memory of the trade sign process. The exact expression of the autocorrelations of the increments of Ut = Tt+ Tt and Xt = Nt+ Nt
as a function of the lag can be hardly deduced from (3.6) and (3.9). In order to
discuss the shape of theses correlation functions one can however use qualitative arguments based on classical Tauberian theorems [12]. In what follows the argument z
of Laplace transforms is assumed to be real and positive (z > 0).
Let us first remark that, within the range of parameters observed in empirical
data (and notably the relationship T 2N ), one can show that the terms involving
bF both at the numerator and at the denominator in (3.9) are subdominant. In this
case, this equation, reads (we neglect the g (h) factor and drop the (h) upper-script
everywhere):
2T
bzT 
C
2


(1 bTz )

(7.4)

Let us note that, according to this expression, the stability condition bT0 < 1 implies
b0T < , i.e., the correlation function of the trades CT is integrable and that,
that C
strictly speaking, there is no long-range memory in supply and demand. Let us show
however that, under the conditions we observe empirically, CT can reproduce, on a
pretty large (though finite) range of , a slow-decay with an exponent smaller than
1, leading various numerical estimations to conclude long-range memory.
As observed in previous section, T is close to a power-law (t > 0):
Tt = (c + t)

(7.5)

25
were = 1 + is the scaling exponent (empirically we found 0.2), c is a small
scale cut-off (empirically c 102 s), and is a factor such that the norm bT0 1.
On has obviously:
=

( 1)bT0
c1

If one computes the Laplace transform (with z 0) of expression (7.5), it is easy to

show that, in the limit of small z (in practice that means z < c1 )
bTz bT0 (1 (1 )(cz) )

Thus, according to (7.4), one has

2T
bT 
C
2
z


(1 bT0 ) + bT0 (1 )(cz) 

(7.6)

b T C C z and therefore, thanks to Tauberian Theorems (limit

and consequently C
z
of small z corresponds to limit of large time) :
CT

However, if we suppose that not only z is small but that bT0 is close enough to 1
such that
then (7.6) becomes

1 bT0 bT0 (1 )(cz)

(7.7)

bzT z 2
C

or equivalently

CT 21

(7.8)

Let us point out that the inequality (7.7) holds, as long as:
z>c

1 bT0
bT (1 )
0

! 1

c1 105

(7.9)

using the estimates 0.2 and bT0 0.9. Since z < c1 , this means that it
can take 5 decades to see the short range nature of the trade correlation function.
Considering that c 102 s, for an inter event mean time of 1s, the scaling (7.8) can
extend over 3 decades. Since 1 2 0.6, this exponent is precisely in the range of
values reported empirically in [8] and [19].
About price efficiency. From an empirical point of view, in agreement with
market efficiency hypothesis, it is well known that price variations are almost uncorrelated after few seconds. Let us show, that, under the conditions observed in
empirical data, CN decreases very fast around = 0. We provide the same kind of

26
pedestrian arguments as in previous discussion: we neglect the influence of F and
suppose that we are in the impulsive case of the impact kernel, i.e., It = It .
Let us remark that, since T /N is bounded (empirically T /N 2), if I is small
enough (empirically I < 101 ), the same arguments invoked for C T shows that in the
intermediate range:
c1 105 z c1
one has T |bIz |2 N |1 bT |2 . It results that (3.6) reduces, in this range, to:
2N
bzN 
C
2


(1 bN
z )

In order to study the behavior of this function, we use again a power-law expression
for N
t :

N
t = (c + t)

(7.10)

where empirically < 0 (and thus bN

0 < 0) and = 1 + with 0.1. It
follows that

bzN C
C

in the range
z c1

1 bN
0

(1 )|bN
0 |

!1/

bN
Since bN
0 < 0, |0 | < 1, (1 ) 1, 1/ 10, we have 1

bN
1
0
bN |
(1 )|
0

1/

and the previous inequality always holds. This shows that the price increments correlation functions decreases very fastly around zero without any fine tuning of the
kernel shapes.
Numerical illustrations. All these considerations are illustrated in Fig. 7.6,
where we have simulated a Hawkes process with the parameters close to the ones we
found empirically. Let us notice that, in order to mimic the experiments performed
in [8] and [19], the correlation functions are computed in trading time, i.e., a discrete
time which is incremented by 1 at each jump of Ut . We checked that setting a lag n in
trading time roughly amounts to consider a lag in physical time such that: = nT .
In that respect, the shapes of the correlation functions in trading time and in physical
time are very close to each other, up to a scaling factor. In Fig. 7.6(a) are displayed
CnT /C0T () and CnN /C0N (solid line) as functions of the lag n (in trading time). On
clearly sees that CnT is slowly decaying while CnN almost vanishes after few lags. In
Fig. 7.6(b) CnT is represented in log-log. As expected, it behaves as power-law with
the exponent 2 1 (0.6 in the example we choose) represented by the solid line.
7.4. Market impact profile estimation from anonymous data. As discussed in Section 5, once one has estimated all the model parameters, it is possible to
compute the shape of the market impact of some particular (meta-) order (Proposition
5.1 with T = 0). If one uses the parameters reported previously, one gets a shape of
the market impact profile associated with the meta-order dAt = T 1 1tT dt similar

27

Fig. 7.6. Empirical correlation functions of the increments Ut and Xt in trading time. The
process is a numerical simulation of a Hawkes process where the kernels have been chosen to fit
the empirical observations found in 7.1. The correlation functions where estimated on a single
T /C T () and C N /C N (solid line) as
realization containing around 40.000 market orders. (a) Cn
n
0
0
T in double logarithmic reprensentation
functions of the lag n (expressed in trading time). (b) Cn
(). The solid line represents the power-law fit with expression (7.8) where 2 1 = 0.6.

to the one displayed in Fig. 5.1(b). One can use similar qualitative arguments than
in previous section to explain the shape observed in Fig. 5.1(b). Indeed, according to
(5.3) and using the same notations, parameter values (notably the fact that bT0 is
close to 1) and approximations as discussed previously, one can show that in a wide
intermediate range of laplace parameters z, we have:

bz =
If dAt = T 1 1tT dt, then A

cI z z A
bz .
M

1ezT
T z2

(7.11)

and therefore

cI z z 1 if z T 1
M
cI z z 2 if z T 1 .
M

It follows that the market impact profile behavior reads:

M It t if t T
M It t1 if t T.
that corresponds to the behavior observed in Fig. 5.1(b). Notice that a strict squareroot would correspond to = 1/2 while it seems we rather observe 0.2 empirically.
In general, the determination of the market impact profile is a hard task that
requires to possess agent labeled (e.g. broker) data. The model presented in this
paper allows one to recover a market profile using anonymous order flow data as
explained in Section 5.3.
8. Conclusion and prospects. From our knowledge, the model we developed
in this paper is the first model that accounts for both stylized facts of market prices
microstructure (including random time arrival of price moves, discrete price grid,
high frequency mean reversion, correlation functions behavior at various time scales)
and the stylized facts of market impact (mainly the concave/relaxation characteristic

28
shape of the market impact of a meta-order). Analytical closed formula can be obtained for most of these stylized facts. Not only it allows us to reveal (through the
estimations of the different kernels) the dynamics involved between trade arrivals and
price moves, but it also allows us to estimate the entire market impact profile from
anonymous market data.
As far as trade and price dynamics are concerned, we have provided evidence
of a power-law behavior of the kernels T and N and that the model is close to
instability (i.e. bT0 is smaller but close to 1). This suggests the existence of some
self-similarity properties in the order-book dynamics and sharply contrasts with the
usual exponential kernels used in former parametric Hawkes modeling in Finance. The
cross-kernels associated with impact of trade on prices (I ) and feed-back (F ) are well
localized in time (i.e. of impulsive nature). Thus, upward (resp. downward) price
moves are mainly impacted by trades on the ask (resp. bid) side. In turn, positive
(resp. negative) mid-price variations imply an increase of the trading intensity on the
bid (resp. ask) side.
Besides these important points, qualitative arguments showed that, provided
bT0 1 and bN
0 < 0, the long memory puzzle of the order flow raised by
Bouchaud et al. [8] can be addressed without any fine tuning of the model parameters: trades naturally appear as long-range correlated over a wide range of lags while
price variations are almost uncorrelated. Moreover, the same kind of arguments can
explain the concave (square-root law) / relaxation typical market impact shape and
an almost vanishing permanent impact.
Let us first point out that the model, as is, can be used as a stochastic price
replayer using as an input the true market order arrivals in place of the stochastic
process Tt . This allows one to replay the price of a given historical period and, for
instance, using it as an input price for any algorithm designed to estimate or manage
a risk.
In this paper, we have presented the most basic form of the model. It can be
seen as a building block for more elaborate models depending on what it is meant
to be used for. There are many ways for extensions. Let us just go through some of
them we have already developed or we are still working on. For instance, it would
be important to have a model which not only takes into account the arrival times of
the market orders but their volumes too. This is a pretty easy extension since it can
be done within the framework of marked Hawkes processes for which straightforward
extensions of all computations presented in this paper can be obtained. In the simplest
form, one could use i.i.d. volumes vt for the market orders : at any time t a market
order arrives (dTt+ = 1 or dTt = 1) a volume vt is chosen randomly (using a given
law). In its simplest form the new model consists in replacing the projection of (2.6)
on the last two components by
N
I
N
t = dNt + f (vt )dTt ,

where f is a function that describes how the volume impacts the price. It basically
corresponds to what is generally referred into the literature by the instantaneous
impact function.
In order to go further into the understanding of the underlying dynamics of the
order-book, a very natural extension of the model, consists in using more dimensions
in order to take into account limit/cancel
 orders.
 Thus, for instance, one way would be
L

t
to introduce a new point process Lt =
where L+
t (resp. Lt ) is incremented
L+
t

29
by 1 whenever a limit order arrives at the best ask (resp. best bid) or a cancel order
arrives at best bid (resp. best ask). One would then need to introduce the different
kernels that account for the influence of L on T and N and the kernels that account
for the influence of T and N on L itself. The estimation procedure of the kernels can
follow the exact same procedure described in Section 6. Along the same line, i.e., by
adding new dimensions to the 4d-Hawkes model presented in this paper, one could
quantify the impact of a given exogenous news on the market order flow or directly on
the price by estimating the corresponding kernel. Or, alternatively, one could consider
a multi-agent models (e.g., adding 2 dimensions for each agent) and model/estimate
the influence of a given agent on another one or on all the anonymous agents (as T
does in (4.3)). These extended framework would open the door to precise estimations
and obvious interpretations in order to get better insights into to full order book
dynamics.
9. Annexes.
9.1. Stability condition : proof of Proposition 2.1. In this section, we give
the proof to the Proposition 2.1 For the process to be stable, we need the eigenvalues
b 0 to have a modulus smaller than 1 (cf hypothesis (H)).
of the matrix
b 0)
Lemma 2. (Eigenvalues of
b 0 , then it satisfies
If x is an eigenvalue of
(a x)(b x) c = 0

(9.1)

(a+ x)(b+ x) c+ = 0,

(9.2)

or

bT,c
bN,s bN,c and c = (bF,s bF,c )(bI,s bI,c ). (Let
where a = bT,s
0 0 , b = 0
0
0
0
0
0

bI
= bF
us point out that, using Notations 4, a = bT0 , b = bN
0 0 ).
0 and c
Proof
From (2.11) we get
!
b T xI
bF

0
0
b
det(0 xI) = det
= 0.
(9.3)
bI
b N xI

I
0
0

Since all the matrices are bi-symmetric, they all commute and thus


b 0 xI) = det (
b T xI)(
b N xI)
bF
bI
det(
0
0
0 0 = 0

(9.4)

Moreover, they all diagonalize in the same basis and their eigenvalues are the sum
and the difference between the self term and the cross term. Thus the eigein values
b I satisfy either (9.1) or (9.2).
bF
b T xI)(
b N xI)
of (
0

We now have to study when the modulus of the roots of these equations are smaller
than 1.
Lemma 3. (Condition for the roots to have a modulus smaller than 1)
The roots of the equation
x2 x(a + b) + ab c = 0
have a modulus smaller than 1 iff

30
(i) either
c < (a b)2 /4

and

ab c < 1

(9.5)

|a + b| < min(1 + ab c, 2)

(9.6)

(ii) or
c > (a b)2 /4 and

Proof
The discriminant of this second-order equation is (a b)2 + 4c, thus there are two
cases
(i) c < (a b)2 /4. In this case the roots are complex and conjugated one of
the other. Thus their modulus is smaller than 1 iff their product is smaller
than 1, i.e., ab c < 1 which gives the result for (ii)
(ii) c > (a b)2 /4. In this case the roots are real. The condition for them to
be smaller than 1 is
1 (a + b) + ab c > 0 and

a+b
<1
2

and the condition for them to be greater than -1 is

1 + (a + b) + ab c > 0 and

a+b
> 1
2

This synthesizes into case ii of Lemma

Using these two lemmas, let us study the stability condition corresponding to (9.1)
and (9.2). Before starting, let us note that, since all the functions are positive
functions (thus all the b0 are real positive), it is clear that
|a | a+ , |b | b+ , |c | c+ .

(9.7)

Proving (2.12) is a necessary condition for (H) to hold.

For this purpose, we just need to look at the condition for the roots of (9.2) to have
a modulus smaller than 1. Indeed, (9.2) falls in the case of (ii) of Lemma 3, since
c = c+ 0. Moreover (9.6) is equivalent to |a+ + b+ | < min(1 + a+ b+ c+ , 2) which
is equivalent to
c+ < (1 a+ )(1 b+ ) and a+ , b+ < 1

(9.8)

which proves (2.12)

Proving (2.12) is a sufficient condition for (H) to hold. (and proving (2.13))
For that purpose, we suppose that (2.12) holds. We want to prove that the roots of
both (9.1) and (9.2) have modulus strictly smaller than 1. We have just seen this is
the case for the roots of (9.2), we just need to check that it is also the case for the
roots of (9.1). For (9.2) the case (i) reads
a b 1 < c < (a b )2 /4
and the case (ii) (since |a + b | a+ + b < 2 according to (9.8))
(a b )2 /4 < c < 1 + a b |a + b |

31
Thus, merging these last two inequations, we get the following condition for the roots
of (9.1) to have a modulus strictly smaller than 1
a b 1 < c < 1 + a b |a + b |,

(9.9)

which is nothing but (2.13). Thus, in order to complete the proof of Proposition 2.1,
we just need to prove that this last inequation holds (i.e., (2.13) holds).
Let us first notice that, since |a | a+ < 1 and |b | b+ < 1, one has 2a b
|a + b |, and consequently
1 + a b |a + b | 1 a b .
Since 1 a b > 0 the following inequation
|c | < 1 + a b |a + b |.
is a sufficient condition for (9.9) to hold. Moreover since |c | c+ , using (2.12), a
sufficient condition for (9.9) to hold is
(1 a+ )(1 b+ ) 1 + a b |a + b |.
which is equivalent to
a+ + b+ a+ b+ |a + b | a b .

(9.10)

a+ + b+ a+ b+ = a+ (1 b+ ) + b+ |a |(1 b+ ) + b+
b+ (1 |a |) + |a | |b |(1 |a |) + |a |

(9.11)
(9.12)

Since, one has

|a | + |b | |a ||b |,

(9.13)

inequation (9.10) is implied by

|a | + |b | |a ||b | |a + b | a b .

(9.14)

Thus, in order to complete the proof of Proposition 2.1, we just need to prove this
last inequation.
In the case a and b have the same sign, this inequation is a actually a strict
equality, so it obviously holds. Let us suppose that a and b do not have the same
sign. Without loss of generality, we can suppose that a 0 b . We distinguish
two cases :
either a b 0, in which case
|a +b |a b = b (1+a )a b (1+a )a = |a |+|b ||a ||b |.
or b < a < 0, in which case
|a +b |a b = a (1b )+b a (1b )+b = |a |+|b ||a ||b |.

32
9.2. Conditional expectation of a Hawkes process. In this Section, we
establish general results on N -dimensional Hawkes process P = {P i }1iN and more
particularly on the expectations of dP i at time t conditioned by the fact that the P j
jumped at time 0. We mainly establish two results. The fist one (Prop. 9.1) links these
conditional expectations with the auto-covariance function of dPt and, using previous
results [1, 2], allows us to derive an analytical formula as a function of the kernel.
These results are used in the Section 5.18 for characterizing the response function.
The second one (Prop. 9.2) proves that these expectations satisfy a Fredholm system
that will be used in Section 6 for elaborating a general procedure for the kernel
estimation.
The Hawkes process is defined by its kernel = {i }1i,jN and the exogenous
intensity = {i }1iN through the equation
t = + dPt ,

(9.15)

where t is the conditional intensity of P at time t. We consider that the process is

b 0 have modulus smaller than 1. We set
stable, i.e., the eigenvalues of the matrix
1

b0
.
(9.16)
= E(t ) = I
Finally for all t and i, j such that 1 i, j N , we define gt = {gtij }1i,jN with
gtij dt = E(dPti |dP0j = 1) ij t i dt

(9.17)

where E(dPti |dP0j = 1) is the conditional expectation of dNti knowing that Ntj jumps
at t = 0, t is the dirac distribution and ij is always 0 except for i = j for which it
is equal to 1. Since E(dNti |dN0i = 1) is singular at t = 0 we substracted this singular
component. In the following gt will refer to the matrix whose elements are the gtij ,
i.e.,
gt = {gtij }1i,jN .

(9.18)

We are ready to state our first result.

Proposition 9.1. Let Ct,t+ be the infinitesimal covariance matrix (without the
singular part) as defined by
ij
Ct,t+ = {Ct,t+
}1i,jN ,

with
ij
i
Ct,t+
dtd = Cov(dPt+
, dPtj ) ij i dt.

Then gt and Ct,t+ are linked through the relation

g = Ct,t+ 1 .

(9.19)

Using the analytical formula proved in [1, 2] for C ij ,one gets

1 +
1 ,
g = +

(9.20)

= and is defined such that (I

b z )1 = (I +
b z ). In the Fourier
where
domain, this last equation corresponds to
b z )1 (I
b )1 1 I.
gz = (I
b
z

(9.21)

33
Proof of the Proposition.
This is a pretty straightforward computation :
ij
i
Ct,t+
dtd = E(dPt+
dPtj ) i j dtd ij i dt.

which gives
ij
i
Ct,t+
dtd = E(dPt+
|dPtj = 1)P rob(dPtj = 1) i j dtd ij i dt.

Using the stationnarity and dividing by dt, one gets

ij
Ct,t+
d = E(dPi |dP0j = 1)j i j d ij i .

Then, using the definition (9.17) of g ij

ij
Ct,t+
d = gij j d

which gives (9.19). From [1, 2], we know that

+
,
Ct,t+ = +

This last equation along with (9.19) leads to (9.20) and then to (9.21).
The next result corresponds to the following proposition :
Proposition 9.2. Using the notations above, the density gt satisfies the following
fredholm system of integral equations for positive t
gt = (I + gt ),

t > 0.

(9.22)

Moreover, for t < 0 we have

gt = gt
.

(9.23)

Proof of the Proposition.

Proof of (9.22). We consider t > 0
gtij = E(dPti |dP0j = 1) i = E(it |dP0j = 1) i
Using (9.15),
gtij = i +

N
X

k=1

ik E(dPtk |dP0j = 1) i .

Using (9.17), we get

gtij = i +

N
X

k=1

ik (gtkj + kj t + k ) i .

And consequently
gtij = i +

N
X

k=1

||ik ||k i + ij
t +

N
X

k=1

ik gtkj

34
PN
However, the vector formulation of i + k=1 ||ik ||k i is nothing but
+ (b0 I) which is 0 according to (9.16), thus
gtij = ij
t +

N
X

ik gtkj

k=1

which gives (9.22).

Let us note that we could have derived directly the system (9.22) from (9.21).
Indeed, (9.21) gives
b z )b
b )1 1 I +
bz.
(I
gz = (I
z

b z is supported by R , thus going back to the time

In the time domain,
domain and restricting to t > 0 directly leads to (9.22).
Proof of (9.23). Let t < 0.
E(dPti |dP0j = 1) = P rob(dPti = 1|dP0j = 1) =

P rob(dPti = 1, dP0j = 1)
P rob(dP0j = 1)

Since P is stable, dP is stationary, P rob(dP0j = 1) = P rob(dPtj = 1) = j ,

thus
j
j E(dPti |dP0j = 1) = P rob(dPti = 1, dP0j = 1) = E(dPt
|dP0i = 1)i

Consequently :
ji
j gtij = i gt

Acknowledgments. The authors thank Sylvain Delattre, Marc Hoffmann, CharlesAlbert Lehalle and Mathieu Rosenbaum for useful discussions. We gratefully acknowledge financial support of the chair Financial Risks of the Risk Foundation and of the
chair QuantValley/Risk Foundation: Quantitative Management Initiative. The financial data used in this paper have been provided by the company QuantHouse
EUROPE/ASIA, http://www.quanthouse.com.
REFERENCES
[1] E. Bacry, K. Dayri, and J.F. Muzy, Non-parametric kernel estimation for symmetric hawkes
processes. application to high frequency financial data, Eur. Phys. J. B, 85 (2012), p. 157.
[2] E. Bacry, S. Delattre, M. Hoffmann, and J. Franc
ois Muzy, Scaling limits for Hawkes
processes and application to financial statistics, ArXiv e-prints, (2012).
[3] E. Bacry, S. Delattre, M. Hoffmann, and J. F. Muzy, Modelling microstructure noise with
mutually exciting point processes, Quantitative Finance, 13 (2013), pp. 6577.
[4] L. Bauwens and N. Hautsch, Modelling financial high frequency data using point processes.,
In T. Mikosch, J-P. Kreiss, R. A. Davis, and T. G. Andersen, editors, Handbook of Financial
Time Series, Springer Berlin Heidelberg, 2009.
[5] N. Bershova and D. Rakhlin, The non-linear market impact of large trades: Evidence from
buy-side order flow, (preprint).
[6] J.P. Bouchaud, Price impact, in Encyclopedia of Quantitative Finance, R. Cont, ed., John
Wiley & Sons Ltd., 2010.
[7] J.P. Bouchaud, J.D. Farmer, and F. Lillo, How markets slowly diggest changes in supply
and demand, in Handbook of Financial Markets, Elsevier, 2009.
[8] J.P. Bouchaud, M. Potters, Y. Gefen, and M. Wyart, Fluctuations and response in financial markets: the subtle nature of random price changes., Quantitative Finance, 4
(2004), pp. 176190.

35
[9] D.J. Daley and D. Vere-Jones, An Introduction to the Theory of Point Processes: Elementary theory and methods, Probability and its applications, Springer, 2003.
[10] P. Embrechts, T. Liniger, and L. Lu, Multivariate hawkes processes: an application to
financial data, To appear in Journal of Applied Probability, (2011).
[11] J.D. Farmer, P. Patelli, and I. Zovko, The predictive power of zero intelligence in financial
markets, Proceedings of the National Academy of Sciences of the United States of America,
102 (2005), pp. 22542259.
[12] W. Feller, An Introduction to Probability Theory and Its Applications, vol. 1, Wiley, January
1968.
[13] J. Gatheral, No dynamic arbitrage and market impact, Quantitative Finance, 10 (2010),
pp. 749759.
[14] A.G. Hawkes, Point spectra of some mutually exciting point processes, Biometrika, 58 (1971),
pp. 8390.
[15]
, Spectra of some self-exciting and mutually exciting point processes, Journal of the Royal
Statistical Society. Series B (Methodological), 33-3 (1971), pp. 438443.
[16] P. Hewlett, Clustering of order arrivals, price impact and trade path optimisation, in Workshop on Financial Modeling with Jump processes, Ecole Polytechnique, 2006.
[17] M. Kirchner, in preparation, 2012.
[18] A.S. Kyle, Continuous auction and insider trading, Econometrica, 53 (1985), p. 1315.
[19] F. Lillo and J.D. Farmer, The long memory of the efficient market, Studies in Nonlinear
Dynamics & Econometrics, 8 (2004).
[20] E. Moro, J. Vicente, L.G. Moyano, A. Gerig, J.D. Farmer, G. Vaglica, F. Lillo, and
R.N. Mantegna, Market impact and trading profile of hidden orders in stock markets,
Phys. Rev. E, 80 (2009), p. 066102.
[21] I. Muni Toke, market making behaviour in an order book model and its impact on the
bid-ask spread, in Econophysics of Order-driven Markets, F. Abergel, B.K. Chakrabarti,
A. Chakraborti, and M. Mitra, eds., Springer Verlag, 2011.
[22] Y. Ogata, On lewis simulation method for point processes, Ieee Transactions On Information
Theory, 27 (1981), pp. 2331.
[23] W.H. Press, S.A. Teukossky, W.T. Vetterling, and B.P. Flannery, Numerical Recipes
in C. The Art of Scientific Computing, Cambridge University Press, 1992.
[24] E. Smith, J.D. Farmer, L. Gillemot, and S. Krishnamurthy, Statistical theory of the continuous double auction, Quantitative Finance, 3 (2003), pp. 481514.