arXiv:1201.4580v1 [q-fin.

TR] 22 Jan 2012

A non-linear model of trading

mechanism on a financial market
N.Vvedenskaya1 , Y.Suhov1,2,3 , V.Belitsky3
Abstract
We introduce a prototype model in an attempt to capture some aspects of market dynamics simulating a trading mechanism. The model
description starts with a discrete-space, contiinuous-time Markov process
describing arrival and movement of orders with different prices. We then
perform a re-scaling procedure leading to a deterministic dynamical system controlled by non-linear odinary differential equations (ODEs). This
allows us to introduce approximations for the equilibrium distribution of
the model represented by fixed points of deterministic dynamics.

Introduction

This paper proposes a model that takes into account, in a rather stylized form,
some aspects of automated trading mechanisms adopted in modern financial
markets, in particular, the dynamics of the limit order book.
In short, a limit order book keeps records of arivals, movements and departures of market participants (traders) who declare their trading positions. An
arriving trader may wish to buy or sell at a certain price, and can move his
declared price when time progresses. If the declared price is met by a trader
with the opposite intention, a trade is recorded: this may lead to disappearance
of one or both participants from the market, due to exhaustion of their offers.
For a detailed description of some common limit order book models and their
applications, see [2, 3, 4] and references therein.

1
Institute for Information Transmission Problems, RAS, Moscow, Russia
2
DPMMS, University of Cambridge, and St Johns College, Cambridge, UK
3
IME, Universidade de Sao Paulo, Sao Paulo, Brazil

In the current paper we present a somewhat different model, including elements of queueing behavior of arrived offers; this model is studied by using
techniques of asymptotic analysis. An earlier account of this work (in its preliminary version) can be found in [1]. Compared with [1], in the present text
we adopt a continuous-time setting for the basic Markov process: it clarifies the
meaning of the main parameters of the model and shortens the proof of some
of our main results.
The model under consideration is a prototype; at this stage it does not aim
to take into account all possible aspects that can be viewed as defining, either
theoretically or practically. Instead, we opted for a simplified description which
leads to some straightforward, yet instructive, answers.
Our model differs from known models of the limit order book in a number of
aspects. Arguably, it can be a subject of criticism (as a number of other proposed
models). In particular, the strategic behavior of the model in its current form
(and some further details) do not quite match existing mechanisms governing
electronic trading on financial markets. Nevertheless, the model shows a certain
amount of flexibility, and covers a borad range of situations. Its mathematical
advantage is that in the scaling limit under consideration, it leads to a single
fixed point.
Our scaling limit is based on suppositions that (i) the number of market
participants is large, (ii) during a very short time period only part of them
makes a decision of performing a trade or maiking a move along the price range,
and (iii) the probability for any given participant to make such decision is small.
This makes it natural to change, in a suitable manner, parameters of original
Markov process.
After rescaling, a limiting dynamical system emerges, with a deterministic behavior described by a system of non-linear ordinary differential equations. The rescaling techniques greatly simplify the structure of the model,
and this phenomenon extends far beyond basic examples like the current prototype model. As we mentioned earlier, the present paper focuses on a simplified
model, with minimal number of constant parameters, where some of technically
involved issues are absent.
A similar approach is commonly used in the literature on stochastic communication networks; see, e.g., [5, 6, 7, 8] and [9]. We also find similarities,
as well as differences, with models proposed (in a different context) in a recent
paper [10]; analogies with [10] could be useful for the aforementioned purpose
of defining the prices that are appropriate for trades.
In the next section we describe the underlying Markov process. In Section
3 the rescaling of the process is presented and the main results are stated and
the proofs are given. The last section contains concluding discussion of various
aspects of the model.

The underlying Markov process

The rationale for the models below is as follows. We consider a single-commodity

market where prices may be at one of N distinct levels (say, c1 < c2 < . . . < cN ,
although the exact meaning of these values is of no importance here).
The market is operating in continuous time t R+ where R+ = [0, ). (As
was mentioned above, the earlier version [1] used a more cumbersome discrtetetime version of the underlying process.) At a given time t R+ , there are bi (t)
traders prepared to buy a unit of the commodity at price ci and si (t) traders
prepared to sell it at this price, which leads to vectors




b(t) = b1 (t), . . . , bN (t) , s(t) = s1 (t), . . . , sN (t) ZN
+.

Here and below Z+ = {0, 1, . . .} stands for a non-negative integer half-lattice

and ZN
+ for the non-negative integer N -dimensional lattice orthant. The pair
(b(t), s(t)) represents a state of a Markov process {U (t)} that will be the subject
of our analysis.
Suppose that, for given k = 1, . . . N and t R+ , we have that bk (t)
sk (t) > 0 then each of the sellers gets a trade at a given rate T > 0 and leaves
the market, together with one of the buyers. Therefore, both values bk (t) and
sk (t) decrease at rate T . In addition, each one among sk (t) sellers (i) quits
the market at rate Q > 0 or (ii) moves to the price level ck1 at rate M > 0,
if k > 1. Similarly, every buyer among the bk (t) buyers (i) quits the market at
the same rate Q as above or (ii) moves to the price level ck+1 , again at rate
M , provided that k < N .
Symmetrically, if sk (t) bk (t) > 0 then each of the buyers gets a trade
at rate T and leaves the market, together with his seller companion. The
remaining traders at the price level ck proceed in a manner as above.
Further, when k = N , a buyer leaves the system with rate Q + M . Similarly, for k = 1, a seller leaves the system with rate Q + M .
Finally, a random Poisson flow of new exogenous buyers arrives at the price
level c1 ; the rate of this arrival equals B > 0. Similarly, a Poisson random flow
of new sellers arrives at the price level cN ; the rate of this arrival is S > 0.
As usually, standard independence
are in place.
 assumptions

 This generates

the aforementioned Markov process U (t) with trajectories (b(t), s(t)) , t
R+ .
Theorem 1 For any values of parameters B/S , Q/M and T , the process
{U (t)} is irreducible, aperiodic and positive
recurrent. Therefore,
it has a unique




,
and
for any initial
set of equilibrium probabilities = b, s : b, s ZN
+

state U (0) (deterministic or random), the distribution of the random state U (t)
at time t converges weakly to as t :




lim P U (t) = (b, s) = b, s .
t

Proof of Theorem 1. Irreducibility and aperiodicity of the process is evident.

Positive recurrence follows from the following observation. The (random) time
that a given trader (a buyer or a seller) spends in the system, i.e., the time from
his arrival till exit, is majorized by a sum of N independent exponential variables. Therefore, the process {U (t)} can be majorized, in a natural fashion, by
an M/M/ queueing process. But the latter is known to be positive recurrent.
The remaining assertions of Theorem 1 are standard. N
Despite a concise description, the detailed pattern of behavior of process
{U (t)} is rather complex, particularly for large values of N . For instance,
consider the differences between vectors b(t ) and b(t) and between s(t ) and
s(t), on a time interval (t, t ) where 0 < t < t . The increments for the entries
bk ( ) and sk ( ) for 1 k N 1 are captured by the following equations:
M
bk (t ) = bk (t) + ik1
(t, t ) nk (t, t ) ikM (t, t ) ikQ (t, t ),

and

M
sk (t ) = sk (t) + jk+1
(t, t ) nk (t, t ) jkM (t, t ) jkQ (t, t ).

Here ikM (t, t ) is the number of buyers who move within time interval (t, t ) from
level k to k + 1 and jkM (t, t ) that of sellers who move from level k to k 1.
Next, ikQ (t, t ) is the number of buyers who quit the system during interval (t, t )
from level k and jkQ (t, t ) the number of sellers who quit the system from level
k. Finally, nk (t, t ) is the number of buyers and sellers who got a trade over
(t, t ) at level k. All listed quantities are non-negative integer-values random
variables. For k = 1 the structure of the expression is similar, with the term
M
M
ik1
(t, t ) being replaced by i B (t, t ) 0, while for k = N the term jk+1
(t, t ) is

replaced by j S (t, t ) 0; both i B (t, t ) and j S (t, t ) being distributed according

to a Poisson law with mean B/S (t t).
In the simplest case of a market with one price level (N = 1), the process {U (t)} is a continuous-time random walk on the two-dimensional lattice
quadrant Z2+ , where
b(t ) = b(t) + i B (t, t ) n(t, t ) i Q (t, t )
and
s(t ) = s(t) + j S (t, t ) n(t, t ) jQ (t, t ).
This already makes analytical representations for the invariant distribution
rather complicated; cf. [11] and references therein.

Scaling limit

The complexity

of the time-dynamics and of the equilibrium distribution for
process U (t) makes it desirable to develop efficient methods of approximation.
In this paper we focus on one such method based on scaling the parameters of
the process (including states and time-steps).
4

The re-scaling procedure is as follows: we fix values > 0, > 0, > 0,

B > 0 and S > 0 and set:
T =

, Q = , M = .
L
L
L

(1)

In addition, we re-scale the states and the time: pictorially,

xk

bk
sk
t
, yk , .
L
L
L

(2)

Formally, denoting the Markov process generated for a given L by U (L) , we

consider the continuous-time process
V (L) ( ) =

1 (L)
U
L), 0.
L

(3)

Let RN
+ denote a positive orthant in N dimensions. Suppose we are given
N
a pair of vectors (x(0), y(0)) RN
+ R+ where x(0) = (x1 (0), . . . , xN (0)),
y(0) = (y1 (0), . . . , yN (0)). Consider the following system of first-order ODEs
for functions xk = xk ( ) and yk = yk ( ) where > 0 and 1 i N :




x 1 = B + x1 min x1 , y1 ,




x k = xk1 + xk min xk , yk , 1 < k N,


(4)


y k = yk+1 + yk min xk , yk , 1 k < N,




y N = S + yN min xN , yN ,

with the initial data

xk (0) 0, yk (0) 0, 1 k N.


The fixed point x , y of system (4) has

(5)

)
x = (x1 , . . . , xN ) and y = (y1 , . . . , yN

where xk and yk give a solution to





B
=
+ x1 + min x1 , y1 ,




+ xk + min xk , yk ,
xk1 =
1 < k N,


 

yk+1 =
+ yk + min xk , yk ,
1 k < N,





S
=
+ yN
+ min xN , yN
.

(1 )
(2 )
(3 )

(6)

(4 )

(In (6) we noted individual equations by addition signs that are used below).
Both systems (4) and (6) are non-linear. However, the non-linearity disappears
at a local level which greatly simplifies the analysis of these systems.
In Theorems 2 and 3 below, we use the distance generated by the Euclidean
norm in RN RN .

Theorem 2 (a) For any initial data (x(0), y(0)) with xi (0) 0, yi (0) 0,
1 i N for all > 0 there exists a unique solution (x( ), y( )), to problem
(4)(5) and xi ( ) > 0, yi ( ) > 0.
(b) As , the solution approaches a fixed point which is a unique
solution to system (6):
h

i


(7)
dist x( ), y( ) , x , y 0.

Proof of Theorem 2, (a). Obviously, a unique solution exists for sufficiently

small > 0. Suppose that xi = max xk and max xk < xi . Then x i 0.
k

k<i

Similarly if yj = max yk and max yk < yj then y j 0. Also, x 1 < 0 if

k

k>j

x1 > B /( + ) and y N < 0 if yN > S /( + ). Moreover, x 1 0 when

x1 = 0 and y N 0 when yN = 0. Thus the components of the solution xi ( )
and yi ( ) are non-negative and uniformly
n bounded. oBy standard constructions
of the ODE theory, a unique solution (x( ), y( )) exists for all > 0, and
x( ), y( ) RN
+.

Before proving assertion (b), we discuss several properties of the solution to

(4)(5). The following Proposition 1 indicates that the solution to (4) possesses
a kind of the min/max principle.
Proposition 1 Suppose that we have two initial points, x = (x1 , . . . , xn ) and
x = (x1 , . . . , xn ) RN
+ such that for some k = 1, . . . , N ,
xk xk , and yk yk .

(8)

Then for the solutions (x( ), y( )), (x ( ), y ( )), to (4), with the initial conditions (x(0), y(0)) = (x, y) and (x (0), y (0)) = (x , y ), for all > 0,
xk ( ) xk ( ), and yk ( ) yk ( ).

(9)

Similarly if xk xk and yk yk then the corresponding solutions obey xk ( )

xk ( ) and yk ( ) yk ( ), for all > 0
Proof of Proposition 1. It is sufficient to consider the situation where the
strict inequalities take place. Suppose that (9) holds strictly for < 0 and
fails at = 0 . For instance, let xk (0 ) = xk (0 ) and assume that k > 1 is the
minimal index for which such an equality takes place.
If yk (0 ) > yk (0 ) then x k (0 ) < x k (0 ) and the above equality xk (0 ) =
xk (0 ) is impossible. The other case is considered in a similar manner. N
A corollary of proposition 1 is
Proposition 2 If x k (0) 0 and y k (0) 0 for all 1 k N then xk ( )
increases and yk ( ) decreases in , for all > 0 and 1 k N . Similarly if
x k (0) 0 and y k (0) 0, 1 k N , then xk ( ) decreases and yk ( ) increases
in .
6

Proof of Proposition 2. It again suffices to assume that the strict inequalities

hold true: x k (0) > 0 and y k (0) < 0. Then, for a small > 0: xk () > xk (0) and
yk () < yk (0).
Set xk (0) = xk (), yk (0) = yk (). The coefficients of equations do not
depend on , therefore xk ( ) = xk ( + ) and yk ( ) = yk ( + ) for all > 0.
By Proposition 1, xk ( + ) = xk ( ) xk ( ) and yk ( + ) = yk ( ) yk ( ),
for all > 0 and 1 k N . As may be arbitrarily small, the assertion of
Proposition 2 is valid. N
Proof of Theorem 2, (b). Given x(0) = (x1 (0), . . . , xN (0)) and
y(0) = (y1 (0), . . . , yN (0)) RN
+ , let (x( ), x( )) be the solution to (4), (5).
Consider two additional solutions, (x ( ), y ( )) and (x ( ), y ( )), to (4) with
i
h
xk (0) = 0, yk (0) = max S /( + ), max yi ,
i

and

h
i
xk (0) = max B /( + ), max xi (0) , yk (0) = 0.
i

By Propositions 1 and 2

xk ( ) xk ( ) xk ( ), yk ( ) yk ( ) yk ( ).

(10)

Further, xk ( ) increases, while yk ( ) decreases in . By the same token, xk ( )

decreases and yk ( ) increases in . Therefore, both pairs (x ( ), y ( )) and
(x ( ), y ( )) tend to limits as , which are fixed points, i.e., solutions to
Eqns (6). By (10), any solution to (4), (5) eventually lies between these limits.
To finish the proof of the theorem, we have to show that the solution (x , y )
to Eqn (6) is unique.
For convenience, we state the corresponding assertion as Lemma 1.
Lemma 1 For any values B/S , , and 0 there exists a unique solution
to Eqns (6).
Proof of Lemma 1. To start with, note that every solution to (6) has

x1 > x2 > . . . > xN , y1 < y2 < . . . < yN

.

Therefore, if x1 y1 then xk < yk , 1 < k N and if xN yN then xk >

yk , 1 < k < N .
It is convenient to introduce auxiliary variables vk , wk 0 in terms of which
Eqns (6) will be treated. Geometrically the idea is as follows: we start with
Eqn (6)(1) : B = ( + )v1 + min [v1 , w1 ] and watch how this relation between
v1 , w1 is transformed by (6)(2) ,(6)(3) to relation between v2 , w2 , then to relation
between v3 , w3 etc.
The locus of points (u, v) R2 where
v, w > 0, B = ( + )v + min[v, w]
7

coincides with a continuous broken line, L1 R2+ , formed by two pieces: 1) a

(0)
(0)
vertical ray L1 emitted from the point o1 lying on the bisectrix where




(0)
o1 = B ( + + ), B ( + + )
(1)

(0)

and 2) a line segment L1 of a negative slope ( + )/, joining o1 with the

(1)
point o1 on the horizontal axis:



(1)
o1 1 = B ( + ), 0 .

See the figure below.

w

(0)

L2

B /

L(0)
1

(1)

L2

(2)
L2

MN

(1)
L1

B /( + )

B / ( + + )

The passage from values v1 , w1 to v2 , w2 generates a 1 1 map acting on

L1 . The image of L1 under the map (v1 , w1 ) 7 (v2 , w2 ) is another continuous
(0)
broken line, L2 R2+ , formed by three pieces: 1) a vertical ray L2 issued from
point



(0)
o2 = B ( + + )2 , B ,
(0)

(1)

2) a line segment L2 joining the points o2

bisectrix and has both co-ordinates equal to

(1)

and o2

( + + )
 B
,
( + + ) ( + )( + + )2 + 2
(2)

(2)

(1)

(2)

where o2

lies on the

and 3) a line segment L2 joining o2 with the point o2 on the horizontal axis:


B
(2)
o2 =
,
0
.
( + )2
8

(1)

(2)

The slopes dw/dv of segments L2 and L2 are negative, but the slope flattens
(1)
(2)
when we pass from L2 to L2 .
In the above figure, L1 and L2 are shown on the same (v, w)-plane R2 ; in
this figure line L2 lies to the left of L1 . (The third broken line, MN present in
the figure is explained below.)
A similar picture persists when we iterate, i.e., pass from (v2 , w2 ) to (v3 , w3 )
and so on. At step k (vk , wk ) 7 (vk+1 , wk+1 ) where again the map is 1 - 1,
(vk+1 , wk+1 ) belongs to a continuous broken line Lk+1 in R2+ formed by k + 2
(0)
(1)
pieces. One piece, Lk+1 , is a vertical ray while the k + 1 others, Lk+1 , . . .,
(k+1)

Lk+1 , are line segments of negative slopes, the slope flattens when we pass
(1)

(k)

(k+1)

from Lk to Lk . The last segment, Lk+1 , joins a point on the bisectrix and
a point on the horizontal axis.
At the end of this process we obtain a continuous broken line LN , the locus
of points (vN , wN ). Our next step is to consider the intersection of LN and MN
where MN is the locus where
S = ( + )wN + min [vN , wN ].
More precisely, MN is a continuous broken line formed by a horizontal ray issued
from the point


S
S
,
,
++ ++
and a line segment joining this point with the point


S
0,
+
lying on the vertical axis. Cf. the figure.
We want to check that the point of intersection is always unique: it yields
to unique solution to (6).
Line LN may intersect the horizontal part of MN . That means that there

exist a solution to (6) where x1 > y1 and xN > yN

(this case is not presented
on our figure). For the proof of lemma it is needed to show that in this case LN
cannot intersect the sloppy part of MN , where yN > xN . For that it is sufficient

(N )
to prove that dw dv on LN \ LN , the part of LN above the bisectrix, is steeper
than ( + ), the slope of the segment of line MN .
On the other hand if LN does not intersect the horizontal part of MN it has
to intersect the sloppy part of MN and it is needed to show that in this
 case
such intersection is unique. Here again it is sufficient to show that dw dv on

(N )
LN \ LN is steeper than ( + ).

(N )
To prove the assertion
 of the lemma we show that dw dv on LN \ LN is
always steeper than ( + ). In fact, it suffices to verify that

dwN
<
on LiN , 1 i < N .
dvN
+
9

(11)

(i)

(i)

(k)

Any segment Lk , 1 < i < k, maps onto segment Lk+1 , segment Lk

onto two segments:

(k)
Lk+1

and

(k+1)
Lk+1 .

The slope of segment

(i)
Lk+1 ,

maps

i = 1, . . . , k+1

(i)
(k+1)
is steeper than that of Lk and the slope of Lk+1 is steeper than that
(i)
More, the slopes of LN , 0 < i N, are steeper then the steep segment
To show that consider three cases of the map Lk Lk+1 :

(k)

of Lk .
of MN .

1) vk > wk , vk+1 > wk+1 ;

2) vk > wk , vk+1 < wk+1 ;
3) vk < wk+1 , vk+1 < wk+1 .
In these cases we have, the following equations

Ak+1 ( + + ) dwk

2
dvk

dwk+1
2
= ( + + ) dwk ,

dvk+1

2
dvk

Bk (++) dwk ,
2

(12)

dvk

where, respectively,

( + )vk+1 + wk+1
( + + )wk

vk =
, wk+1 =
,

( + + )vk+1
( + + )wk
v =
, wk+1 =
,
k

( + )wk + vk
v = ( + + )vk+1 , w
.
k
k+1 =

vk
wk+1
and Bk = + +
.
Here Ak+1 = + +
vk+1
wk
Therefore for N > 2
 dw 
 ( + + )N 1 dw  ( + + )N 1 +
 N

1

 >
=
N 1
dvN

dv
N 1

1
 dw 



>
.
=

dv MN
+

For N = 2 we have the middle case in (12) with k = 1, k + 1 = 2. Here

+
1
again the needed inequality takes place. In fact, dw
dv1 = . By using (12),
we get that for N = 2,
dw2
( + + )2 dw1
=
on L12 .
du2
2
du2

This finishes the proof of Lemma 1. N

Theorem 3 Suppose that the re-scaled initial states converge in probability: for
any > 0,




1
lim P dist
U (0), (x(0), y(0)) = 0.
L
L
10

n
o
Then, for all T > 0, the process V (L) ( ), [0, T ] converges in probability
n
o
to the solution (x( ), y( )), 0 T . That is, > 0,
lim P

sup

0 T


n
h
io
dist V (L) ( ), (x( ), y( )) = 0.

In particular, if x(0) = x and y(0) = y then



n
h

io
(L)

= 0.
lim P sup dist V ( ), x , y
L

(13)

(14)

0 T



Moreover, if process U (t), is in equilibrium then Eqn (14) holds true.

Proof of Theorem 3. Let G(L) denote the generator of the Markov process {U (L) (t)} (with rates as in Eqn (1)). Then the action of matrix G(L) on
functions (b, s) of state variables b, s ZN
+ is determined by the equation

X 
(b ek , s) (b, s) bk
L
1kN

 
+ (b, s ek ) (b, s) sk
 X



+
(b ek + ek+1 , s) (b, s) bk
L
1k<N

X 

+
(b, s ek + ek1 ) (b, s) sk

G(L) (b, s) =

(15)

1<kN





+B (b + e1 , s) (b, s) + S (b, s + eN ) (b, s)
+


X 
(b ek , s ek ) (b, s) bk sk .
L
1kN

Here ek , 1 k N , stands for the vector in ZN

+ whose components are all 0s
except for the kth one, equal to 1.
we can take a function = (L) of the form (b, s) =

In particular,
b s
where is a smooth function on RN
,

+ . This choice agrees with the

L L


spatial scaling x b L, y s L in Eqn (2). Then x = (x1 , . . . , xN ) RN
+
and y = (y1 , . . . , yN ) RN
+ , with b = Lx, s = Ly where stands for the

11

integer part, we obtain that




G(L) Lx L,Ly L
h 
 
X


 
 i 
=
b ek L, s L b L, s L bk L
1kN

 
h  

 
 i 
+ b L, s ek L b L, s L sk L
 X h 
 


 
 i 
+
b ek + ek+1 L, s L b L, s L bk L
1k<N

 
X h  

 
 i 
+
b L, s ek + ek1 L b L, s L sk L
1<kN

h 
 


 
 i
+B b + e1 L, s L b L, s L
 
h  

 
 i
+S b L, s + eN L b L, s L

 
X h 

 
 i




b ek L, s ek L b L, s L
b k sk L .

1kN


Next, we multiply the both side by L in agreement with the time-scale t L
in Eqn (2) and pass to the limit L . This yields
lim LG(L) (x, y)
(
X 


+ yk
=
xk
xk
yk
h X1kN



X

i
+
xk
+

yk

xk+1
xk
yk1
yk
1k<N
1<kN
)


X



(x, y) .
+ S
+
xk yk
+ B
x1
yN
xi
yi

1kN

Applying Theorem 6.1 from [12], we obtain Eqn (13).

The next remark is that each scaled process {V (L) } has a unique invariant
distribution
b(L) ; the family of probability distributions
b(L) (considered on
N
R+ ) is compact in the sense of convergence in probability. This can be deduced
from the above remark that the original processes {U (L) }, and hence, the scaled
process {V (L) } can be majorized by suitable analogs of M/M/ queueing systems. It is easy to see that every limiting point for
b(L) when L is a
delta-measure sitting at a fixed point for system (5). However, the latter is
unique and coincides with (x , y ). Consequently, the distributions
b(L) converge in probability to the aforementioned delta-measure. Then, applying the
already established result, we obtain Eqn (14). N

12

Fixed points in the scaling limit. Concluding

remarks

The approximation developed in Theorem 3 calls for an analysis of solutions to

(6).
The parameter space R5+ formed by , Q/M , and b/s is partitioned into
open domains where one of the following generic patterns persists:
(i) vN > wN , (ii) v1 < w1 , and (iii) vi > wi for i = 1, . . . , and vi < wi for
i = + 1, . . . , N where 1 < < N . In each of these domains system (6) is linear.


A particular algorithm for calculating x , y is based on the following
recursion. Set
(0)

(0)

M yi+1
M xi1
s
b
(0)
(0)
(0)
, y =
, xi =
, y =
Q + M +
Q + M + N
Q + M i
Q + M


Next, let x(k) , y (k) , k = 1, 2, . . . be the solution to the system


 (k1) (k1) 
(k)
b
= q + m x1 + min x1
, y1
,


 (k1) (k1) 
(k)
(k)
, 1 < i N,
, yi
m xi1 = q + m xi + min xi


 (k) (k1) 
(k)
(k)
, 1 i < N,
m yi+1 = q + m yi + min xi , yi


 (k) (k1) 
(k)
.
s
= q + m yN + min xN , yN
(0)

x1 =

(k)

(k1)

(k)

<
, yi
> xi
These iterations converge because the inequalities xi
(k) (k)
hold true i, k 1 and, values xi , yi are uniformly bounded and
(k)
(k)
there exist lim xi , lim yi that, naturally, satisfy (6).

(k1)
yi

We conclude with the following remarks.

Our model presents also a caricature of overproduction crisis: In fact, if

is
s
P so that xi < yi , 1 i N, then the amount of trades
P sufficiently large,

min[x
,
y
]
=
i
i
i xi is not changing by increase of s , all extra sellers
i
leave the market without performing any trade.
It is interesting to investigate the dependance of trade performance on
as , that is where the trade action happens almost immediately after
the moment when traders appear at some price level. Then almost all trades
happen at two levels i0 and i0 + 1, and xi is very small as i > i0 + 1, yi ia
very small as i < i0 . But our limit model does not permit to consider the case
= , though, sure the initial Markov process can be investigated in case of
immediate trade deals. The limiting case = of our model is close to the
problems investigated in [10].
We hope that the variation of these models parameters can help to determine factors attracting or repelling various market participants. An important
aspect of any model of the market is what possibilities it gives for an accurate
13

prediction of the stochastic component in the dynamics of the market prices

and volumes.
The current set-up of the model presented here admits straightforward generalizations to the case where parameters and Q/M depend on i, 0 < i < N
and on the trader type (b/s). Another generalization emerges if these parameters and b/s become state-dependent. It is also possible to allow the exogenous
buyers and sellers to enter the system at any price level among c1 , . . ., cN . To
take into account elements of the FCFS discipline, one could introduce various
priorities into the dynamics of process U (t).
Finally, we would like to note that there are several forms of convergence for
which the assertion in Theorem 3 holds true.

Acknowledgements
The authors thank the anonymous referees for stimulating critical remarks. NV
would like to thank the grant RFBR 11-01-00485-a. YS would like to thank
the FAPESP Foundation, Sao Paulo, Brazil, for providing a grant towards this
work, and NUMEC/IME, University of Sao Paulo, for warm hospitality.

References
[1] N.D. Vvedenskaya, Y. Suhov, V. Belitsky, A non-linear model of limit order
book dynamics,http://arXiv:1102.1104 (2011); a presentation held at ISIT2011 (St Petersburg, July 31August 5, 2011).
[2] R. Cont, S. Stoikov, R. Talreja, A stochastic model for order book dynamics,
Operations Research, vol. 58, pp. 549563, 2010.
[3] R. Cont, A. Kukanov, S. Stoikov, The price impact of order book events,
Technical paper, http://ssm.com./abstract=1712822 (2011).
[4] I. Rosu, A dynamic model of the limiting order book, Rev.Financial Studies,
vol. 22, pp. 46014641, 2009.
[5] N. D. Vvedenskaya, R. L. Dobrushin, F. I.Karpelevich, A queueing system
with selection of the shortest of two queues: an asymptotical approach,
Problems of Information Transmission, vol. 32, pp. 1527, 1996.
[6] M. Mitzenmacher, The power of two choices in randomized load balancing.
PhD Thesis, UC Berkeley, 1996.
[7] J.B. Martin, Yu.M. Suhov, Fast Jackson networks, Ann. Appl. Prob., vol.
9, 854870, 1999.
[8] N. D. Vvedenskaya, Y. M. Suhov, Multy access system with many users:
stability and metastability, Problems of Information Transmission, vol. 43,
pp. 263271, 2007.
14

[9] G. Sharma, A. Ganesh, P. Key, Performance analysis of contention

based medium access control protocol, IEEE Trans. Information Theory, vol.
55, pp. 16651682, 2009.
[10] V.A. Malyshev, A.D. Manita, A.A. Zamyatin, Explicit asymptotic velocity of the boundary between particles and antiparticles,
http://arXiv.org.abs/11104776.
[11] I.A. Kurkova, Y.M. Suhov, Malyshevs theory and JS-queues, Ann. Appl.
Prob., vol. 13, pp. 13131354, 2003.
[12] S.N. Ethier, T.G. Kurtz, Markov Processes Characterization and Convergence, New York et al.: J. Wiley & Sons, 1986.
N.Vvedenskaya, Institute for Information Transmission
Problems, RAS, GSP-4, Moscow 127994, RUSSIA; [email protected]
Y.Suhov, Institute for Information Transmission Problems, RAS, GSP-4,
Moscow 127994, RUSSIA; DPMMS, University of Cambridge, and St Johns
College, Cambridge CB3 0WB, UK; IME, Universidade de Sao Paulo, C.P.
66281 CEP 05389-970 Sao Paulo, BRAZIL; [email protected]
V.Belitsky, IME, Universidade de Sao Paulo, C.P. 66281 CEP 05389-970 Sao
Paulo, BRAZIL; [email protected]

15