Math Finan Econ (2013) 7:477–507

DOI 10.1007/s11579-012-0087-0

Dealing with the inventory risk: a solution to the market

making problem

Olivier Guéant · Charles-Albert Lehalle ·

Joaquin Fernandez-Tapia

Received: 31 March 2012 / Accepted: 14 August 2012 / Published online: 4 September 2012
© Springer-Verlag 2012

Abstract Market makers continuously set bid and ask quotes for the stocks they have under
consideration. Hence they face a complex optimization problem in which their return, based
on the bid-ask spread they quote and the frequency at which they indeed provide liquidity, is
challenged by the price risk they bear due to their inventory. In this paper, we consider a sto-
chastic control problem similar to the one introduced by Ho and Stoll (J Fin Econ 9(1): 47–73,
1981) and formalized mathematically by Avellaneda and Stoikov (Quant Fin 8(3):217–224,
2008). The market is modeled using a reference price St following a Brownian motion with
standard deviation σ , arrival rates of buy or sell liquidity-consuming orders depend on the
distance to the reference price St and a market maker maximizes the expected utility of its
P&L over a finite time horizon. We show that the Hamilton–Jacobi–Bellman equations asso-
ciated to the stochastic optimal control problem can be transformed into a system of linear
ordinary differential equations and we solve the market making problem under inventory
constraints. We also shed light on the asymptotic behavior of the optimal quotes and propose
closed-form approximations based on a spectral characterization of the optimal quotes.

Keywords Stochastic optimal control · High-frequency market making ·

Avellaneda–Stoikov problem

This research has been conducted within the Research Initiative “Microstructure des Marchés Financiers”
under the aegis of the Europlace Institute of Finance.

O. Guéant (B)
Laboratoire Jacques-Louis Lions, UFR de Mathématiques, Université Paris-Diderot,
175, rue du Chevaleret, 75013 Paris, France
e-mail: olivier.gueant@gmail.com

C.-A. Lehalle
Head of Quantitative Research, Crédit Agricole Cheuvreux, 9, Quai du Président Paul Doumer,
92400 Courbevoie, France
e-mail: clehalle@cheuvreux.com

J. Fernandez-Tapia
LPMA, Université Pierre et Marie Curie, 4 Place Jussieu, 75005 Paris, France

123
478 Math Finan Econ (2013) 7:477–507

JEL Classification C61 · G10 · G11

1 Introduction

From a quantitative viewpoint, market microstructure is a sequence of auction games between

market participants. It implements the balance between supply and demand, forming an equi-
librium traded price to be used as reference for valuation. The rule of each auction game
(fixing auction, continuous auction, etc.) is fixed by the firm operating each trading venue.
Nevertheless, most of all trading mechanisms on electronic markets rely on market partic-
ipants sending orders to a “queuing system” where their open interests are consolidated as
“liquidity provision” or form transactions [1]. The efficiency of such a process relies on an
adequate timing between buyers and sellers, to avoid too many non-informative oscillations
of the transaction price (for more details and modeling, see for example [20]).
In practice, it is possible to provide liquidity to an impatient buyer (respectively seller)
and maintain an inventory until the arrival of the next impatient seller (respectively buyer).
Market participants focused on this kind of liquidity-providing activity are called “market
makers”. On one hand they are buying at the bid price and selling at the ask price they choose,
making money out of this “bid-ask spread”. On the other hand, their inventory is exposed to
price fluctuations mainly driven by the volatility of the market (see [2,5,10,11,17,24]).
The recent evolution of both technology and market regulation reshaped the nature of
the interactions between market participants during continuous electronic auctions, one con-
sequence being the emergence of “high-frequency market makers” who are said to be part
of 70 % of the electronic trades in the US (40 % in the EU and 35 % in Japan) and have a
massively passive (i.e. liquidity-providing) behavior—a typical balance between passive and
aggressive orders for such market participants being around 80 % of passive interactions (see
[22]).
From a mathematical modeling point of view, the market making problem corresponds
to the choice of optimal quotes (i.e. the bid and ask prices) that such agents provide to other
market participants, taking into account their inventory limits and their risk constraints often
represented by a utility function (see [9,16,19,21,23,25]).
Avellaneda and Stoikov proposed, in a widely cited article [3], an innovative framework
for “market making in an order book”. In their approach, rooted to an old article by Ho
and Stoll [18], the market is modeled using a reference price or fair price St following a
Brownian motion with standard deviation σ , and the arrival of a buy or sell liquidity-con-
suming order at a distance δ from the reference price St is described by a point process with
intensity A exp(−kδ), A and k being two positive constants which characterize statistically
the liquidity of the stock.
We consider the same model as in [3]—adding inventory limits—and we show, using
a new change of variables, that the Hamilton–Jacobi–Bellman equations associated to the
problem boil down to a system of linear ordinary differential equations. This new change of
variables (i) simplifies the computation of a solution since numerical approximation of partial
differential equations is now unnecessary, and (ii) allows to study the asymptotic behavior of
the optimal quotes. In addition to these two contributions, we use results from spectral anal-
ysis to provide an approximation of the optimal quotes in closed-form. Finally, we provide
in the case of our model with inventory limits a verification theorem that was absent from the
original article (the admissibility of the quotes obtained in the original Avellaneda–Stoikov
model appears in fact to be an open problem!).

123
Math Finan Econ (2013) 7:477–507 479

Since Avellaneda and Stoikov seminal article, other authors have considered related mar-
ket making models. Cartea et al. [8] consider a more sophisticated model inspired from the
Avellaneda–Stoikov one,1 including richer dynamics of market orders, impact on the limit
order book, adverse selection effects and predictable α. They obtained closed-form approxi-
mations of the optimal quotes using a first-order Taylor expansion. Cartea and Jaimungal [7]
recently used a similar model to introduce risk measures for high-frequency trading. Earlier,
they used a model inspired from Avellaneda–Stoikov [6] in which the mid-price is modeled
by a Hidden Markov Model. Guilbaud and Pham [14] also used a model inspired from the
Avellaneda–Stoikov framework but including market orders and limit orders at best (and next
to best) bid and ask together with stochastic spreads. Very recently, Guilbaud and Pham [15]
used another model inspired from the Avellanada-Stoikov one in a pro-rata microstructure.
It is also noteworthy that the model we use to find the orders a market maker should opti-
mally send to the market has been used in a totally different domain of algorithmic trading:
optimal execution. Bayraktar and Ludkovski [4] and Guéant et al. [13] used indeed a similar
model to optimally liquidate a portfolio.
This article starts in Sect. 2 with the description of the model. Section 3 is dedicated to
the introduction of our change of variables and solves the control problem in the presence
of inventory limits. Section 4 focuses on the asymptotic behavior of the optimal quotes and
characterizes the asymptotic value using an eigenvalue problem that allows to propose a
rather good approximation in closed-form. Section 5 generalizes the model in two different
directions: (i) the introduction of a drift in the price dynamics and (ii) the introduction of
market impact that may also be regarded as adverse selection. Section 6 carries out the com-
parative statics. Section 7 provides backtests of the model. Adaptations of our results are in
use at Cheuvreux.

2 Setup of the model

Let us fix a probability space (, F , P) equipped with a filtration (Ft )t≥0 satisfying the
usual conditions. We assume that all random variables and stochastic processes are defined
on (, F , (Ft )t≥0 , P).
We consider a high-frequency market maker operating on a single stock2 . We suppose
that the mid-price of this stock or more generally a reference price3 of the stock moves as an
arithmetic Brownian motion:4
d St = σ dWt
The market maker under consideration will continuously propose bid and ask prices
denoted respectively Stb and Sta and will hence buy and sell shares according to the rate
of arrival of market orders at the quoted prices. His inventory q, that is the (signed) quantity
of shares he holds, is given by
qt = Ntb − Nta

1 The objective function is different.

2 We suppose that this high-frequency market maker does not “make” the price in the sense that he has no
market power. In other words, we assume that the size of his orders is small enough to consider price dynamics
exogenous. Market impact will be introduced in Sect. 5.
3 This reference price for the stock can be thought of as a smoothed mid-price for instance.
4 Since we will only consider short horizon problems, this assumption is almost equivalent to the usual
Black-Scholes one.

123
480 Math Finan Econ (2013) 7:477–507

where N b and N a are the point processes (independent of W ) giving the number of shares the
market maker respectively bought and sold (we assume that transactions are of constant size,
scaled5 to 1). Arrival rates obviously depend on the prices Stb and Sta quoted by the market
maker and we assume, in accordance with the model proposed by Avellaneda and Stoikov
[3], that intensities λb and λa associated respectively to N b and N a depend on the difference
between the quoted prices and the reference price (i.e. δtb = St − Stb and δta = Sta − St ) and
are of the following form6 :

λb (δ b ) = Ae−kδ = A exp(−k(s − s b )) λa (δ a ) = Ae−kδ = A exp(−k(s a − s))

b a

where A and k are positive constants that characterize the liquidity of the stock. In particular,
this specification means—for positive δ b and δ a —that the closer to the reference price an
order is posted, the faster it will be executed.
As a consequence of his trades, the market maker has an amount of cash evolving according
to the following dynamics:
d X t = (St + δta )d Nta − (St − δtb )d Ntb
To this original setting introduced by Avellaneda and Stoikov (itself following partially
Ho and Stoll [18]), we add a bound Q to the inventory that a market maker is authorized to
have. In other words, we assume that a market maker with inventory Q(Q > 0 depending
in practice on risk limits) will never set a bid quote and symmetrically that a market maker
with inventory −Q, that is a short position of Q shares in the stock under consideration, will
never set an ask quote. This realistic restriction may be read as a risk limit and allows to solve
rigorously the problem.
Now, coming to the objective function, the market maker has a time horizon T and his
goal is to optimize the expected utility of his P&L at time T . In line with [3], we will focus
on CARA utility functions and we suppose that the market maker optimizes:
 
sup E − exp (−γ (X T + qT ST ))
(δta )t ,(δtb )t ∈A

where A is the set of predictable processes bounded from below, γ is the absolute risk aver-
sion coefficient characterizing the market maker, X T is the amount of cash at time T and
qT ST is the evaluation of the (signed) remaining quantity of shares in the inventory at time
T (liquidation at the reference price ST ).7

3 Characterization of the optimal quotes

The optimization problem set up in the preceding section can be solved using the classical
tools of stochastic optimal control. The first step of our reasoning is therefore to introduce
the Hamilton–Jacobi–Bellman (HJB) equation associated to the problem. More exactly, we
introduce a system of Hamilton–Jacobi–Bellman partial differential equations which consists
of the following equations indexed by q ∈ {−Q, . . . , Q} for (t, s, x) ∈ [0, T ] × R2 :

5 The only important hypothesis is the constant size of orders since we can easily replace 1 by any positive
size Δ.
6 Some authors also used a linear form for the intensity functions—see [18] for instance.
7 Our results would be mutatis mutandis the same if we added a penalization term −b(|q |) for the shares
T
remaining at time T . The rationale underlying this point is that price risk prevents the trader from having
important exposure to the stock. Hence, qt should naturally mean-revert around 0.

123
Math Finan Econ (2013) 7:477–507 481

For |q| < Q:

1
∂t u(t, x, q, s) + σ 2 ∂ss2
u(t, x, q, s)
 2 
+ sup λb (δ b ) u(t, x − s + δ b , q + 1, s) − u(t, x, q, s)
δb
 
+ sup λa (δ a ) u(t, x + s + δ a , q − 1, s) − u(t, x, q, s) = 0
δa

For q = Q:

1
∂t u(t, x, Q, s) + σ 2 ∂ss
2
u(t, x, Q, s)
 2 
+ sup λ (δ ) u(t, x + s + δ a , Q − 1, s) − u(t, x, Q, s) = 0
a a
δa

For q = −Q:

1
∂t u(t, x, −Q, s) + σ 2 ∂ss 2
u(t, x, −Q, s)
 2 
+ sup λb (δ b ) u(t, x − s + δ b , −Q + 1, s) − u(t, x, −Q, s) = 0
δb

with the final condition:

∀q ∈ {−Q, . . . , Q}, u(T, x, q, s) = − exp (−γ (x + qs))

To solve these equations we will use a change of variables based on two different ideas.
First, the choice of a CARA utility function allows to factor out the Mark-to-Market value
of the portfolio (x + qs). Then, the exponential decay for the intensity functions λb and λa
allows to reduce the Hamilton–Jacobi–Bellman (HJB) equations associated to our control
problem to a linear system of ordinary differential equations:

Proposition 1 (Change of variables for (HJB)) Let us consider a family (vq )|q|≤Q of positive
functions solution of:
 
∀q ∈ {−Q + 1, . . . , Q − 1}, v̇q (t) = αq 2 vq (t) − η vq−1 (t) + vq+1 (t)
v̇ Q (t) = α Q 2 v Q (t) − ηv Q−1 (t)
v̇−Q (t) = α Q 2 v−Q (t) − ηv−Q+1 (t)

−(1+ γk )
with ∀q ∈ {−Q, . . . , Q}, vq (T ) = 1, where α = 2k γ σ 2 and η = A(1 + γk ) .
γ
Then, u(t, x, q, s) = − exp(−γ (x + qs))vq (t)− k is solution of (HJB).

Then, the following proposition proves that there exists such a family of positive
functions:

Proposition 2 (Solution of the ordinary differential equations) Let us introduce the matrix
M defined by:

123
482 Math Finan Econ (2013) 7:477–507

⎛ 2 ⎞
αQ −η 0 ··· ··· ··· 0
⎜ . .. .. ⎟
⎜ −η α(Q − 1)2
⎜ −η 0 . . . . ⎟ ⎟
⎜ .. .. .. .. .. .. ⎟
⎜ 0 . . . . . . ⎟
⎜ ⎟
⎜ . . .. .. .. .. .. ⎟
M =⎜
⎜ .. .. . . . . ⎟
. ⎟
⎜ . ⎟
⎜ . .. .. .. .. .. ⎟
⎜ . . . . . . 0 ⎟
⎜ ⎟
⎜ .. .. .. ⎟
⎝ . . . 0 −η α(Q − 1)2 −η ⎠
0 ··· ··· ··· 0 −η α Q2
−(1+ γk )
where α = k2 γ σ 2 and η = A(1 + γk ) .
Let us define
v(t) = (v−Q (t), v−Q+1 (t), . . . , v0 (t), . . . , v Q−1 (t), v Q (t))
= exp(−M(T − t)) × (1, . . . , 1)
Then, (vq )|q|≤Q is a family of positive functions solution of the equations of Proposition 1.
Using the above change of variables and a verification approach, we are now able to solve
the stochastic control problem, that is to find the value function of the problem and the optimal
quotes:
Theorem 1 (Solution of the control problem) Let consider (vq )|q|≤Q as in Proposition 2.
γ
Then u(t, x, q, s) = − exp(−γ (x + qs))vq (t)− k is the value function of the control
problem.
Moreover, the optimal quotes are given by:
 
1 vq (t) 1  γ
s − s (t, q, s) = δ (t, q) = ln
b∗ b∗
+ ln 1 + , q = Q
k vq+1 (t) γ k
 
1 vq (t) 1  γ
s a∗ (t, q, s) − s = δ a∗ (t, q) = ln + ln 1 + , q  = −Q
k vq−1 (t) γ k
and the resulting bid-ask spread quoted by the market maker is given by:
 
∗ 1 vq+1 (t)vq−1 (t) 2  γ
ψ (t, q) = − ln + ln 1 + , |q|  = Q
k vq (t)2 γ k

4 Asymptotic behavior and approximation of the optimal quotes

To exemplify our findings and in order to motivate the asymptotic approximations we shall
provide, we plotted on Figs. 1 and 2 the behavior with time and inventory of the optimal
quotes. The resulting bid-ask spread quoted by the market maker is plotted on Fig. 3.
We clearly see that the optimal quotes are almost independent of t, as soon as t is far
from the terminal time T . This observation is at odds with the approximations proposed8 in
8 In [3], the approximations obtained by the authors using an expansion in q leads to the following expressions:

1  γ  1 + 2q
δtb∗ ln 1 + + γ σ 2 (T − t)
γ k 2
1  γ  1 − 2q
δta∗ ln 1 + + γ σ 2 (T − t)
γ k 2
One can easily show, using the results of Theorem 1, that these approximations are nothing but the Taylor
expansions of the optimal quotes for t close to T .

123
Math Finan Econ (2013) 7:477–507 483

5.5
5
4.5
4
s − s [Tick]

3.5
3
b

2.5
2
1.5
1
600
500
400
30
Tim 300 20
e[ 200 10
Se 0
c] 100 −10
−20 to ry
Inven
0 −30

Fig. 1 Behavior of the optimal bid quotes with time and inventory. σ = 0.3 Tick s−1/2 , A = 0.9 s−1 , k =
0.3 Tick −1 , γ = 0.01 Tick −1 , T = 600 s

5.5
5
4.5
s a − s [Tick]

4
3.5
3
2.5
2
600
1.5 500
1 400
−30 300
−20 c]
−10 200 [ Se
0
10 100 Time
Invento 20 0
ry 30

Fig. 2 Behavior of the optimal ask quotes with time and inventory. σ = 0.3 Tick s−1/2 , A = 0.9 s−1 , k =
0.3 Tick −1 , γ = 0.01 Tick −1 , T = 600 s

Avellaneda and Stoikov [3] using an expansion in q. It motivates however the study of the
asymptotic behavior of the quotes.

Theorem 2 (Asymptotics for the optimal quotes) The optimal quotes have asymptotic limits

lim δ b∗ (0, q) = δ∞
b∗
(q)
T →+∞
lim δ a∗ (0, q) = δ∞
a∗
(q)
T →+∞

123
484 Math Finan Econ (2013) 7:477–507

6.64
6.63
6.62
6.61
ψ [Tick]

6.6
6.59
6.58
6.57
600
6.56
500
6.55 400
−30 300
−20 c]
−10
0
200 [ Se
10 100
Ti me
Invento 20 0
ry 30

Fig. 3 Behavior of the resulting bid-ask spread with time and inventory. σ = 0.3 Tick s−1/2 , A =
0.9 s−1 , k = 0.3 Tick −1 , γ = 0.01 Tick −1 , T = 600 s

that can be expressed as:

   
1  γ 1 f q0 1  γ 1 f q0
δ∞
b∗
(q) = ln 1 + + ln δ∞
a∗
(q) = ln 1 + + ln
γ k k 0
f q+1 γ k k 0
f q−1

where f 0 ∈ R2Q+1 is an eigenvector corresponding to the smallest eigenvalue of the matrix

M introduced in Proposition 2 and characterized (up to a multiplicative constant) by:

Q 
Q−1
f0 ∈ argmin αq 2 f q 2 + η ( f q+1 − f q )2 + η f Q 2 + η f −Q 2
f ∈R2Q+1 , f 2 =1 q=−Q q=−Q

The resulting bid-ask spread quoted by the market maker is asymptotically:

 0 
2  γ
0
f q+1 f q−1
∗ 1
ψ∞ (q) = − ln + ln 1 +
k f q0 2 γ k

The above result, along with the example of Figs. 1, 2, and 3, encourages to approxi-
mate the optimal quotes and the resulting bid-ask spread by their asymptotic value. These
asymptotic values depend on f 0 and we shall provide a closed-form approximation for f 0 .
The above characterization of f 0 corresponds to an eigenvalue problem in R2Q+1 and
we propose to replace it by a similar eigenvalue problem in L 2 (R) for which a closed-form
solution can be computed. More precisely we replace the criterion

Q 
Q−1
f0 ∈ argmin αq 2 f q 2 + η ( f q+1 − f q )2 + η f Q 2 + η f −Q 2
f ∈R2Q+1 , f 2 =1 q=−Q q=−Q

by the following criterion for f˜0 ∈ L 2 (R):

+∞ 
f˜0 ∈ argmin αx 2 f˜(x)2 + η f˜ (x)2 d x
 f˜ L 2 (R) =1−∞

123
Math Finan Econ (2013) 7:477–507 485

The introduction of this new criterion is rooted to the following proposition that states (up
to its sign) the expression for f˜0 in closed form:

Proposition 3 Let us consider

  
f˜0 ∈ argmin αx 2 f˜(x)2 + η f˜ (x)2 d x
 f˜ L 2 (R) =1
R

Then:
 1   
1 α 8 1 α 2
f˜0 (x) = ± 1 exp − x
π4 η 2 η

From the above proposition, we expect f q0 to behave, up to a multiplicative constant, as

  
exp − 21 αη q 2 . This heuristic viewpoint induces an approximation of the optimal quotes
and the resulting optimal bid-ask-spread:

1  γ 1 α
δ∞
b∗
(q) ln 1 + + (2q + 1)
γ k 2k η

1  γ  2q + 1 σ 2 γ  γ 1+ γk
ln 1 + + 1+
γ k 2 2k A k
  
1 γ 1 α
δ∞
a∗
(q) ln 1 + − (2q − 1)
γ k 2k η

1  γ  2q − 1 σ 2 γ  γ 1+ γk
ln 1 + − 1+
γ k 2 2k A k

∗ 2  γ σ 2γ  γ 1+ γk
ψ∞ (q) ln 1 + + 1+
γ k 2k A k
We exhibit on Figs. 4 and 5 the values of the optimal quotes, along with their associated
approximations. Empirically, these approximations for the quotes are satisfactory in most
cases and are always very good for small values of the inventory q. In fact, even though f 0
appears to be well approximated by the gaussian approximation, we cannot expect a very
good fit for the quotes when q is large because we are approximating expressions that depend
f q0 f q0
on ratios of the form 0
f q+1
or 0
f q−1
.

5 Extensions of the model

5.1 The case of a trend in the price dynamics

So far, the reference price was supposed to be a Brownian motion. In what follows we extend
the model to the case of a trend in the price dynamics:

d St = μdt + σ dWt

In that case we have the following proposition (the proof is not repeated):

123
486 Math Finan Econ (2013) 7:477–507

10 7

8 6

6 5

s − s [Tick]
s − sb [Tick]

4 4

b
2 3

0 2

−2 1

−4 0
−30 −20 −10 0 10 20 30 −30 −20 −10 0 10 20 30
Inventory Inventory

Fig. 4 Asymptotic behavior of optimal bid quote (bold line). Approximation (dotted line). Left: σ =
0.4 Tick s−1/2 , A = 0.9 s−1 , k = 0.3 Tick −1 , γ = 0.01 Tick −1 , T = 600 s. Right: σ = 1.0 Tick s−1/2 , A =
0.2 s−1 , k = 0.3 Tick −1 , γ = 0.01 Tick −1 , T = 600 s
10 7

8 6

6 5
s − s [Tick]
sa − s [Tick]

4 4

2 3
a

0 2

−2 1

−4 0
−30 −20 −10 0 10 20 30 −30 −20 −10 0 10 20 30
Inventory Inventory

Fig. 5 Asymptotic behavior of optimal ask quote (bold line). Approximation (dotted line). Left: σ =
0.4 Tick s−1/2 , A = 0.9 s−1 , k = 0.3 Tick −1 , γ = 0.01 Tick −1 , T = 600 s. Right: σ = 1.0 Tick s−1/2 , A =
0.2 s−1 , k = 0.3 Tick −1 , γ = 0.01 Tick −1 , T = 600 s

Proposition 4 (Solution with a drift) Let us consider a family of functions (vq )|q|≤Q solution
of the linear system of ODEs that follows:
 
∀q ∈ {−Q + 1, . . . , Q − 1}, v̇q (t) = (αq 2 − βq)vq (t) − η vq−1 (t) + vq+1 (t)
v̇ Q (t) = (α Q 2 − β Q)v Q (t) − ηv Q−1 (t)
v̇−Q (t) = (α Q 2 + β Q)v−Q (t) − ηv−Q+1 (t)
−(1+ k )
with ∀q ∈ {−Q, . . . , Q}, vq (T ) = 1, where α = k2 γ σ 2 , β = kμ and η = A(1 + γk ) γ .

− γk
Then, u(t, x, q, s) = − exp(−γ (x + qs))vq (t) is the value function of the control
problem.
The optimal quotes are given by:
 
1 vq (t) 1  γ
s − s (t, q, s) = δ (t, q) = ln
b∗ b∗
+ ln 1 +
k vq+1 (t) γ k
  
1 vq (t) 1 γ
s a∗ (t, q, s) − s = δ a∗ (t, q) = ln + ln 1 +
k vq−1 (t) γ k
and the resulting bid-ask spread of the market maker is:

123
Math Finan Econ (2013) 7:477–507 487

 
∗ 1 vq+1 (t)vq−1 (t) 2  γ
ψ (t, q) = − ln + ln 1 +
k vq (t)2 γ k
Moreover,
 
1  γ 1 f q0
lim δ (0, q) = ln 1 +
b∗
+ ln
T →+∞ γ k k 0
f q+1
 
1  γ 1 f q0
lim δ (0, q) = ln 1 +
a∗
+ ln
T →+∞ γ k k 0
f q−1
 0 
2  γ
0
f q+1 f q−1
1
lim ψ ∗ (0, q) = − ln + ln 1 +
T →+∞ k f q0 2 γ k

where f 0 is an eigenvector corresponding to the smallest eigenvalue of:

⎛ 2 ⎞
αQ + βQ −η 0 ··· ··· ··· 0
⎜ . .. .. ⎟
⎜
⎜ −η α(Q − 1)2 + β(Q − 1) −η 0 . . . . ⎟
⎟
⎜ . . . . . . ⎟
⎜ . . . .
. . . . . . .
. ⎟
⎜ 0 ⎟
⎜ .. .. .. .. .. .. .. ⎟
⎜ . . . . . ⎟
⎜ . . ⎟
⎜ .. ⎟
⎜ .. .. .. .. .. ⎟
⎜ . . . . . . 0 ⎟
⎜ ⎟
⎜ .. .. .. ⎟
⎝ . . . 0 −η α(Q − 1) − β(Q − 1)
2 −η ⎠
0 ··· ··· ··· 0 −η αQ − βQ
2

In addition to this theoretical result, we can consider an approximation similar to the

approximation used for the initial model with no drift. We then obtain the following approx-
imations for the optimal quotes and the bid-ask spread:

1  γ   μ 2q + 1

σ 2γ  γ 1+ γk
δ∞
b∗
(q) ln 1 + + − 2 + 1+
γ k γσ 2 2k A k

 
1  γ μ 2q − 1 σ 2γ  γ 1+ γk
δ∞
a∗
(q) ln 1 + + − 1 +
γ k γσ2 2 2k A k

∗ 2  γ σ 2γ  γ 1+ γk
ψ∞ (q) ln 1 + + 1+
γ k 2k A k

5.2 The case of market impact

Another extension of the model consists in introducing market impact. The simplest way to
proceed is to consider the following dynamics for the price:
d St = σ dWt + ξ d Nta − ξ d Ntb , ξ >0
When a limit order on the bid side is filled, the reference price decreases. On the contrary,
when a limit order on the ask side is filled, the reference price increases. This is in line with
the classical modeling of market impact for market orders, ξ being a constant since the limit
orders posted by the market maker are all supposed to be of the same size.
Adverse selection is another way to interpret the interaction we consider between the
price process and the point processes modeling execution: trades on the bid side are often

123
488 Math Finan Econ (2013) 7:477–507

followed by a price decrease and, conversely, trades on the ask side are often followed by a
price increase.
In this modified framework, the problem can be solved using a change of variables that is
slightly more involved than the one presented above but the method is exactly the same and
we have the following proposition (the proof is not repeated):

Proposition 5 (Solution with market impact) Let us consider a family of functions

(vq )|q|≤Q solution of the linear system of ODEs that follows:

k  
∀q ∈ {−Q + 1, . . . , Q − 1}, v̇q (t) = αq 2 vq (t) − ηe− 2 ξ vq−1 (t) + vq+1 (t)
k
v̇ Q (t) = α Q 2 v Q (t) − ηe− 2 ξ v Q−1 (t)
k
v̇−Q (t) = α Q 2 v−Q (t) − ηe− 2 ξ v−Q+1 (t)

with ∀q ∈ {−Q, . . . , Q}, vq (T ) = exp(− 21 kξ q 2 ), where α = 2γσ

k 2 and η = A(1 +
γ −(1+ γk )
k ) .
γ
Then, u(t, x, q, s) = − exp(−γ (x + qs + 21 ξ q 2 ))vq (t)− k is the value function of the
control problem.
The optimal quotes are given by:

 
1 vq (t) ξ 1  γ
s − s b∗ (t, q, s) = δ b∗ (t, q) = ln + + ln 1 +
k vq+1 (t) 2 γ k
 
1 vq (t) ξ 1  γ
s a∗ (t, q, s) − s = δ a∗ (t, q) = ln + + ln 1 +
k vq−1 (t) 2 γ k

and the resulting bid-ask spread of the market maker is :

 
∗ 1 vq+1 (t)vq−1 (t) 2  γ
ψ (t, q) = − ln + ξ + ln 1 +
k vq (t)2 γ k

Moreover,

 
1  γ ξ 1 f q0
lim δ (0, q) = ln 1 +
b∗
+ + ln
T →+∞ γ k 2 k 0
f q+1
 
1  γ ξ 1 f q0
lim δ (0, q) = ln 1 +
a∗
+ + ln
T →+∞ γ k 2 k 0
f q−1
 0 
2  γ
0
f q+1 f q−1
1
lim ψ ∗ (0, q) = − ln + ξ + ln 1 +
T →+∞ k f q0 2 γ k

123
Math Finan Econ (2013) 7:477–507 489

where f 0 is an eigenvector corresponding to the smallest eigenvalue of:

⎛ k
⎞
α Q2 −ηe− 2 ξ 0 ··· ··· ··· 0
⎜ .. .. .. ⎟
⎜ k k ⎟
⎜−ηe− 2 ξ α(Q − 1)2 −ηe− 2 ξ 0 . . . ⎟
⎜ ⎟
⎜ .. .. .. .. .. .. ⎟
⎜ 0 . . . . . . ⎟
⎜ ⎟
⎜ .. . . . . . . . . . . .
. ⎟
⎜ . . . . . . . ⎟
⎜ ⎟
⎜ . . . . . . ⎟
⎜ .. .. .. .. .. .. 0 ⎟
⎜ ⎟
⎜ .. .. .. ⎟
⎜ k k ⎟
⎝ . . . 0 −ηe− 2 ξ α(Q − 1)2 −ηe− 2 ξ ⎠
k
0 ··· ··· ··· 0 −ηe− 2 ξ α Q2

In addition to this theoretical result, we can consider an approximation similar to the

approximation used for the initial model. We then obtain the following approximations for
the optimal quotes and the bid-ask spread:

1  γ  ξ 2q + 1 k σ 2γ  γ 1+ γk
δ∞
b∗
(q) ln 1 + + + e4ξ 1+
γ k 2 2 2k A k

1  γ ξ 2q − 1 k ξ σ 2 γ  γ 1+ γk
δ∞
a∗
(q) ln 1 + + − e4 1+
γ k 2 2 2k A k

∗ 2  γ k σ 2γ  γ 1+ γk
ψ∞ (q) ln 1 + + ξ + e4ξ 1+
γ k 2k A k

6 Comparative statics

We argued in Sect. 4 that the value of the optimal quotes was almost independent of t for
t sufficiently far from the terminal time T and we characterized, using spectral arguments,
the asymptotic value of the optimal quotes. In order to obtain closed-form formulae, we
also provided approximations for the asymptotic value of the optimal quotes. Although these
closed-form formulae are only approximations, they provide a rather good intuition about
the influence of the different parameters on the optimal quotes.

6.1 Dependence on σ 2

The dependence of optimal quotes on σ 2 depends on the sign of the inventory. More precisely,
we observe numerically, in accordance with the approximations, that:
⎧ b∗
⎪
⎪ ∂δ∞ ∂δ∞a∗
⎪
⎪ < 0, > 0, if q < 0
⎪
⎨ ∂σb∗
2 ∂σ 2
∂δ∞ ∂δ∞a∗

⎪ > 0, > 0, if q = 0
⎪
⎪ ∂σ 2 ∂σ 2
⎪
⎪ ∂δ b∗ ∂δ a∗
⎩ ∞ > 0, ∞
< 0, if q > 0
∂σ 2 ∂σ 2
For the bid-ask spread, we obtain:
∗
∂ψ∞
>0
∂σ 2

123
490 Math Finan Econ (2013) 7:477–507

The rationale behind this is that an increase in σ 2 increases inventory risk. Hence, to
reduce this risk, a market maker that has a long position will try to reduce his exposure and
hence ask less for his stocks (to get rid of some of them) and accept to buy at a lower price
(to avoid buying new stocks). Similarly, a market maker with a short position tries to buy
stocks, and hence increases its bid quote, while avoiding short selling new stocks, and he
increases its ask quote to that purpose. Overall, due to the increase in price risk, the bid-ask
spread widens as it is well instanced in the case of a market maker with a flat position (this
one wants indeed to earn more per trade to compensate the increase in inventory risk).

6.2 Dependence on μ

The dependence of optimal quotes on the drift μ is straightforward and corresponds to the
intuition. If the agent expect the price to increase (resp. decrease) he will post orders with
higher (resp. lower) prices. Hence we have:

∂δ∞
b∗ ∂δ∞
a∗
< 0, >0
∂μ ∂μ
6.3 Dependence on A

Because of the form of the system of equations that defines v, the dependence on A must be
the exact opposite of the dependence on σ 2 :
⎧ b∗
⎪
⎪ ∂δ∞ ∂δ∞
a∗
⎪
⎪ > 0, < 0, if q < 0;
⎪
⎨ ∂A ∂A
∂δ∞
b∗ ∂δ∞
a∗

⎪ < 0, < 0, if q = 0
⎪ ∂A
⎪ ∂A
⎪
⎩ ∂δ∞ < 0, ∂δ∞
b∗ a∗
⎪
> 0, if q > 0
∂A ∂A

For the bid-ask spread, we obtain:

∗
∂ψ∞
<0
∂A

The rationale behind these results is that an increase in A reduces the inventory risk. An
increase in A indeed increases the frequency of trades and hence reduces the risk of being
stuck with a large inventory (in absolute value). For this reason, an increase in A should have
the same effect as a decrease in σ 2 .

6.4 Dependence on γ

Using the closed-form approximations, we see that the dependence on γ is ambiguous. The
market maker faces indeed two different risks that contribute to inventory risk: (i) trades
occur at random times and (ii) the reference price is stochastic. But if risk aversion increases,
the market maker will mitigate the two risks: (i) he may set his quotes closer to one another
to reduce the randomness in execution and (ii) he may widen his spread to reduce price risk.
The tension between these two roles played by γ explains the different behaviors we may
observe, as on Figs. 6 and 7 for the bid-ask spread resulting from the asymptotic optimal
quotes:

123
Math Finan Econ (2013) 7:477–507 491

6.5
Spread [Tick]

5.5

4.5 0
0.1
4
0.2
30
20 0.3
10
0 0.4
γ
−10
Inven −20 0.5
tory −30

Fig. 6 Bid-ask spread resulting from the asymptotic optimal quotes for different inventories and different
values for the risk aversion parameter γ .σ = 0.3 Tick s−1/2 , A = 0.9 s−1 , k = 0.3 Tick −1 , T = 600 s

2.5

2.4

2.3
Spread [Tick]

2.2

2.1

1.9 0
0.1
1.8
0.2
30
20 0.3
10 γ
0 0.4
Inven −10
tory −20 0.5
−30

Fig. 7 Bid-ask spread resulting from the asymptotic optimal quotes for different inventories and different
values for the risk aversion parameter γ .σ = 0.6 Tick s−1/2 , A = 0.9 s−1 , k = 0.9 Tick −1 , T = 600 s

6.5 Dependence on k

From the closed-form approximations, we expect δ∞ b∗ to be decreasing in k for q greater

than some negative threshold. Below this threshold, we expect it to be increasing. Similarly
we expect δ∞ a∗ to be decreasing in k for q smaller than some positive threshold. Above this

threshold we expect it to be increasing.

Eventually, as far as the bid-ask spread is concerned, the closed-form approximations
indicate that the resulting bid-ask spread should be a decreasing function of k.

123
492 Math Finan Econ (2013) 7:477–507

10
8
s − s b [Tick]

6
4
2
0
−2
2 30
20
1.5
10
1 0
k [T
ic ory
k −1 0.5
−10
ent
] −20 Inv
0 −30

Fig. 8 Asymptotic optimal bid quotes for different inventories and different values of k.σ = 0.3 Tick s−1/2 ,
A = 0.9 s−1 , γ = 0.01 Tick −1 , T = 600 s

∂ψ∞∗
<0
∂k

In fact several effects are in interaction. On one hand, there is a “no-volatility” effect that
is completely orthogonal to any reasoning on the inventory risk: when k increases, in a sit-
uation where δ b and δ a are positive, trades occur closer to the reference price St . For this
reason, and in absence of inventory risk, the optimal bid-ask spread has to shrink. However,
an increase in k also affects the inventory risk since it decreases the probability to be executed
(for δ b , δ a > 0). Hence, an increase in k is also, in some aspects, similar to a decrease in A.
These two effects explain the expected behavior.
Numerically, we observed that the “no-volatility” effect dominated for the values of the
inventory under consideration (see Fig. 8 for the case of the bid quote).9

6.6 Dependence on the market impact ξ

The market impact introduced in 5.2 has two effects on the optimal quotes. In the absence
of price risk, given the functional form of the execution intensities, the direct effect of ξ is
approximately to add ξ2 the each optimal quote: the market maker approximately maintains
his profit per round trip on the market but the probability of occurrence of a trade is reduced.
This adverse selection effect has a side-effect linked to inventory risk: since adverse selec-
tion gives the market maker an incentive to post orders deeper in the book, it increases the
risk of being stuck with a large inventory for a market maker holding such an inventory. As
a consequence, for a trader holding a positive (resp. negative) inventory, there is a second
effect inciting to buy and sell at lower (resp. higher) prices. These two effects are clearly
highlighted by the closed-form approximations exhibited in the previous section:

9 The case of the ask quote is obviously similar.

123
Math Finan Econ (2013) 7:477–507 493


1  γ ξ 2q + 1 k σ 2γ  γ 1+ γk
δ∞
b∗
(q) ln 1 + + + 4ξ 1+
γ k e !"
 2!" 2
side−effect
2k A k
adverse selection

1  γ ξ 2q − 1 k σ 2γ  γ 1+ γk
δ∞
a∗
(q) ln 1 + + − 4ξ 1+
γ k e !"
 2!" 2
side−effect
2k A k
adverse selection

7 Backtests

Before using the above model on historical data, we need to discuss some features of the
model that need to be adapted before any backtest is possible.
First of all, the model is continuous in both time and space while the real control problem
is intrinsically discrete in space, because of the tick size, and in time, because orders have
a certain priority and changing position too often reduces the actual chance to be reached
by a market order. Hence, the model has to be reinterpreted in a discrete way. In terms of
prices, quotes must not be between two ticks and we decided to round the optimal quotes
to the nearest tick. In terms of time, an order of size ATS10 is sent to the market and is
not canceled nor modified for a given period of time Δt, unless a trade occurs and, though
perhaps partially, fills the order. Now, when a trade occurs and changes the inventory or when
an order stayed in the order book for longer than Δt, then the optimal quote is updated and,
if necessary, a new order is inserted.
Concerning the parameters, σ, A and k can be calibrated on trade-by-trade limit order
book data while γ has to be chosen. However, it is well known by practitioners that A and
k have to depend at least on the actual market bid-ask spread. Since we do not explicitly
take into account the underlying market, there is no market bid-ask spread in the model.
For the backtest example we present below, A and k have been chosen independent of the
spread but, in practice, the value of A and k are function of the market bid-ask spread. As
far as γ is concerned, we decided in our backtests to assign γ an arbitrary value for which
the inventory stayed between −10 and 10 during the day we considered (the unit being the
ATS).
Turning to the backtests, they were carried out with trade-by-trade data and we assumed
that our orders were entirely filled when a trade occurred at or above the ask price quoted by
the agent. Our goal here is just to exemplify the use of the model11 and we considered the
case of the French stock France Telecom on March 15th 2012.
We first plot the price of the stock France Telecom on March 15th 2012 on Fig. 9, the
evolution of the inventory12 on Fig. 10 and the associated P&L (the stocks in the portfolio
are evaluated at mid-price) on Fig. 11.
This P&L can be compared to the P&L of a naive trader (Fig. 12) who only posts orders
at the first limit of the book on each side, whenever he is asked to post orders—that is when
one of his orders has been executed or after a period of time Δt with no execution.

10 ATS is the average trade size.

11 We, voluntarily, do not give full details about the algorithm based on the model.
12 The ATS is 1105 for the day we considered.

123
494 Math Finan Econ (2013) 7:477–507

11.29

11.28

11.27

11.26

11.25
Price

11.24

11.23

11.22

11.21

11.2
10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time

Fig. 9 Price of the stock France Telecom on 15/03/2012, from 10:00 to 16:00

2
Inventory

−2

−4

−6

−8
10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time

Fig. 10 Inventory (in ATS) when the strategy is used on France Telecom (15/03/2012) from 10:00 to 16:00

1400

1200

1000
PnL [in EUR]

800

600

400

200

0
10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time

Fig. 11 P&L when the strategy is used on France Telecom (15/03/2012) from 10:00 to 16:00

123
Math Finan Econ (2013) 7:477–507 495

800

600

400
PnL [in EUR]

200

−200

−400

−600
10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time
Fig. 12 P&L of a naive market maker on France Telecom (15/03/2012) from 10:00 to 16:00

11.24

11.23
Price

11.22

11.21

11.2

12:00 12:10 12:20 12:30 12:40 12:50 13:00

Time
Fig. 13 Details for the quotes and trades when the strategy is used on France Telecom (15/03/2012). Thin lines
represent the market while bold lines represent the quotes of the market maker. Dotted lines are associated to
the bid side while plain lines are associated to the ask side. Black points represent trades in which the market
maker is involved

Now, to better understand the details of the strategy, we focused on a subperiod of 1 hour
and we plotted the state of the market along with the quotes of the market maker (Fig. 13).
Trades occurrences involving the market maker are signalled by a dot.

8 Conclusion

In this article we present a model for the optimal quotes of a market maker. Starting from
a model in line with Avellaneda and Stoikov [3] and rooted to Ho and Stoll [18], we intro-
duce a change of variables that allows to transform the HJB equation into a system of linear
ordinary differential equations. This transformation allows to find the optimal quotes and to

123
496 Math Finan Econ (2013) 7:477–507

characterize their asymptotic behavior. Closed-form approximations are also obtained using
spectral analysis.
The change of variables introduced in this article can also be used to solve the initial
equations of Avellaneda and Stoikov [3] and we provide a complete mathematical proof
in [12]. However, in the absence of inventory limits, no proof of optimality is available
for the quotes claimed to be optimal in [3] and their admissibility appears to be an open
problem.
An important topic for future research consists in generalizing the model to any inten-
sity function. This is particularly important because the exponential form of the intensity is
only suited to liquid stocks with a small bid-ask spread. Another important topic consists in
introducing “passive market impact” (i.e. the perturbations of the price formation process by
liquidity provision). This is a real modeling challenge since no quantitative model for this
type of impact has been proposed in the literature.

Acknowledgements The authors wish to acknowledge the helpful conversations with Yves Achdou (Uni-
versité Paris-Diderot), Vincent Fardeau (London School of Economics), Thierry Foucault (HEC), Jean-Michel
Lasry (Université Paris-Dauphine), Antoine Lemenant (Université Paris-Diderot), Pierre-Louis Lions (Collège
de France), Albert Menkveld (VU University Amsterdam), Vincent Millot (Université Paris-Diderot) and
Nizar Touzi (Ecole Polytechnique). The authors also would like to thank two anonymous referees for their
suggestions.

Appendix: Proofs of the results

Proof of Propositions 1 and 2, and Theorem 1:

Proof Let us consider a family (vq )|q|≤Q of positive functions solution of the system of ODEs
γ
introduced in Proposition 1 and let us define u(t, x, q, s) = − exp (−γ (x + qs)) vq (t)− k .
Then:

1 γ v̇q (t) γ 2σ 2 2
∂t u + σ 2 ∂ss
2
u=− u+ q u
2 k vq (t) 2

Now, concerning the hamiltonian parts, we have for the bid part (q  = Q):
 
sup λb (δ b ) u(t, x − s + δ b , q + 1, s) − u(t, x, q, s)
δb
#  − γ $
−kδ b vq+1 (t) k
= sup Ae u(t, x, q, s) exp(−γ δ )
b
−1
δb vq (t)

The first order condition of this problem corresponds to a maximum (because u is negative)
and writes:
 − γ
vq+1 (t) k
(k + γ ) exp(−γ δ ) b∗
=k
vq (t)

Hence:
 
1 vq (t) 1  γ
δ b∗ = ln + ln 1 +
k vq+1 (t) γ k

123
Math Finan Econ (2013) 7:477–507 497

and
 
sup λb (δ b ) u(t, x − s + δ b , q + 1, s) − u(t, x, q, s)
δb
γ
=− A exp(−kδ b∗ )u(t, x, q, s)
k+γ
γA  γ − γk vq+1 (t)
=− 1+ u(t, x, q, s)
k+γ k vq (t)

Similarly, the maximizer for the ask part (for q  = −Q) is:
 
1 vq (t) 1  γ
δ a∗
= ln + ln 1 +
k vq−1 (t) γ k

and
 
sup λa (δ a ) u(t, x + s + δ a , q − 1, s) − u(t, x, q, s)
δa
γ
=− A exp(−kδ a∗ )u(t, x, q, s)
k+γ
γA  γ − γk vq−1 (t)
=− 1+ u(t, x, q, s)
k+γ k vq (t)

Hence, putting the terms altogether we get for |q| < Q:

1
∂t u(t, x, q, s) + σ 2 ∂ss2
u(t, x, q, s)
2 
+ sup λb (δ b ) u(t, x − s + δ b , q + 1, s) − u(t, x, q, s)
δb
 
+ sup λa (δ a ) u(t, x + s + δ a , q − 1, s) − u(t, x, q, s)
δa
 
γ v̇q (t) γ 2σ 2 2 γA  γ  γk vq+1 (t) vq−1 (t)
=− u+ q u− 1+ + u
k vq (t) 2 k+γ k vq (t) vq (t)
#   $
γ u kγ σ 2 2  γ − 1+ γk
=− v̇q (t) − q vq (t) + A 1 + (vq+1 (t) + vq−1 (t)) = 0
k vq (t) 2 k

For q = −Q we have:

1
∂t u(t, x, q, s) + σ 2 ∂ss2
u(t, x, q, s)
2 
+ sup λb (δ b ) u(t, x − s + δ b , q + 1, s) − u(t, x, q, s)
δb
γ v̇q (t) γ 2σ 2 2 γA  γ  γk vq+1 (t)
=− u+ q u− 1+ u
k vq (t) 2 k+γ k vq (t)
#   $
γ u kγ σ 2 2  γ − 1+ γk
=− v̇q (t) − q vq (t) + A 1 + vq+1 (t) = 0
k vq (t) 2 k

123
498 Math Finan Econ (2013) 7:477–507

Similarly, for q = Q we have:

1
∂t u(t, x, q, s) + σ 2 ∂ss
2
u(t, x, q, s)
2 
+ sup λ (δ ) u(t, x − s + δ a , q + 1, s) − u(t, x, q, s)
a a
δa
γ v̇q (t) γ 2σ 2 2 γA  γ  γk vq−1 (t)
=− u+ q u− 1+ u
k vq (t) 2 k+γ k vq (t)
#   $
γ u kγ σ 2 2  γ − 1+ γk
=− v̇q (t) − q vq (t) + A 1 + vq−1 (t) = 0
k vq (t) 2 k

Now, noticing that the terminal condition for vq is consistent with the terminal condition
for u, we get that u verifies (HJB) and this proves Proposition 1.
The positivity of the functions (vq )|q|≤Q was essential in the definition of u. Hence we
need to prove that the solution to the above linear system of ordinary differential equations,
namely v(t) = exp(−M(T − t)) × (1, . . . , 1) (where M is given in Proposition 2), defines
a family (vq )|q|≤Q of positive functions.
In fact, we are going to prove that:

∀t ∈ [0, T ], ∀q ∈ {−Q, . . . , Q}, vq (t) ≥ e−(α Q

2 −η)(T −t)

If this was not true then there would exist  > 0 such that:
 
e−2η(T −t) vq (t) − e−(α Q −η)(T −t) + (T − t) < 0
2
min
t∈[0,T ],|q|≤Q

But this minimum is achieved at some point (t ∗ , q ∗ ) with t ∗ < T and hence:
d −2η(T −t)  %
%
vq ∗ (t) − e−(α Q −η)(T −t) % ∗ ≥ 
2
e
dt t=t

This gives:
∗)
 ∗

2ηe−2η(T −t vq ∗ (t ∗ ) − e−(α Q −η)(T −t )
2

∗
 ∗

+e−2η(T −t ) vq ∗ (t ∗ ) − (α Q 2 − η)e−(α Q −η)(T −t ) ≥ 
2

Hence:
2 −η)(T −t ∗ ) ∗)
2ηvq ∗ (t ∗ ) + vq ∗ (t ∗ ) − (η + α Q 2 )e−(α Q ≥ e2η(T −t

Now, if |q ∗ | < Q, this gives:

αq ∗ 2 vq ∗ (t ∗ ) − η(vq ∗ +1 (t ∗ ) − 2vq ∗ (t ∗ ) + vq ∗ −1 (t ∗ ))
2 −η)(T −t ∗ ) ∗)
−(η + α Q 2 )e−(α Q ≥ e2η(T −t

Thus:
 ∗

αq ∗ 2 vq ∗ (t ∗ ) − e−(α Q −η)(T −t ) − η(vq ∗ +1 (t ∗ ) − 2vq ∗ (t ∗ ) + vq ∗ −1 (t ∗ ))
2

2 −η)(T −t ∗ ) ∗)
−(η + α(Q 2 − q ∗ 2 ))e−(α Q ≥ e2η(T −t

All the terms on the left hand side are nonpositive by definition of (t ∗ , q ∗ ) and this gives
a contradiction.

123
Math Finan Econ (2013) 7:477–507 499

If q ∗ = Q, we have:
(α Q 2 + η)v Q (t ∗ ) − η(v Q−1 (t ∗ ) − v Q (t ∗ ))
2 −η)(T −t ∗ ) ∗)
−(η + α Q 2 )e−(α Q ≥ e2η(T −t
Thus:
 ∗
 ∗
−η(v Q−1 (t ∗ ) − v Q (t ∗ )) + (η + α Q 2 ) v Q (t ∗ ) − e−(α Q −η)(T −t ) ≥ e2η(T −t )
2

All the terms on the left hand side are nonpositive by definition of (t ∗ , q ∗ ) = (t ∗ , Q) and
this gives a contradiction.
Similarly, if q ∗ = −Q, we have:
(α Q 2 + η)v−Q (t ∗ ) − η(v−Q+1 (t ∗ ) − v Q (t ∗ ))
2 −η)(T −t ∗ ) ∗)
−(η + α Q 2 )e−(α Q ≥ e2η(T −t
 ∗
 ∗
−η(v−Q+1 (t ∗ ) − v−Q (t ∗ )) + (η + α Q 2 ) v−Q (t ∗ ) − e−(α Q −η)(T −t ) ≥ e2η(T −t )
2

All the terms on the left hand side are nonpositive by definition of (t ∗ , q ∗ ) = (t ∗ , −Q)
and this gives a contradiction.
As a consequence, vq (t) ≥ e−(α Q −η)(T −t) > 0 and this completes the proof of Proposi-
2

tion 2.
Combining the above results, we see that u, as defined in Theorem 1, is a solution of
(HJB). Then, we are going to use a verification argument to prove that u is the value function
of the optimal control problem under consideration and prove subsequently that the optimal
controls are as given in Theorem 1.
Let us consider processes (ν b ) and (ν a ) ∈ A. Let t ∈ [0, T ) and let us consider the
following processes for τ ∈ [t, T ]:
d Sτt,s = σ dWτ , Stt,s = s
d X τt,x,ν = (Sτ + ντa )d Nτa − (Sτ − ντb )d Nτb , X tt,x,ν = x
t,q,ν
dqτt,q,ν = d Nτb − d Nτa , qt =q

where the point process N b has intensity (λbτ )τ with λbτ = Ae−kντ 1qτ − <Q and where the
b

point process N a has intensity (λaτ )τ with λaτ = Ae−kντ 1qτ − >−Q 13 .
a

Now, since u is smooth, let us write Itô’s formula for u, between t and tn where tn =
T ∧ inf{τ > t, |Sτ − s| ≥ n or |Nτa − Nta | ≥ n or |Nτb − Ntb | ≥ n}(n ∈ N):
t,q,ν
u(tn , X tt,x,ν
n−
, qtn − , Stt,s
n
) = u(t, x, q, s)
tn  
t,q,ν σ2 2 t,q,ν
+ ∂τ u(τ, X τt,x,ν
− , qτ − , Sτ ) +
t,s
∂ss u(τ, X τt,x,ν
− , q τ− , Sτ
t,s
) dτ
2
t
tn  
t,q,ν t,q,ν
+ u(τ, X τt,x,ν t,s a t,s t,x,ν
− + Sτ + ντ , qτ − − 1, Sτ ) − u(τ, X τ − , qτ − , Sτ ) λτ dτ
t,s a

t
tn  
b t,q,ν t,q,ν
+ u(τ, X τt,x,ν t,s t,s t,x,ν
− − Sτ + ντ , qτ − + 1, Sτ ) − u(τ, X τ − , qτ − , Sτ ) λτ dτ
t,s b

13 These intensities are bounded since ν b and ν a are bounded from below

123
500 Math Finan Econ (2013) 7:477–507

tn
t,q,ν
+ σ ∂s u(τ, X τt,x,ν
− , qτ − , Sτ )dWτ
t,s

t
tn  
t,q,ν t,q,ν
+ u(τ, X τt,x,ν t,s a t,s t,x,ν
− + Sτ + ντ , qτ − − 1, Sτ ) − u(τ, X τ − , qτ − , Sτ ) d Mτ
t,s a

t
tn  
b t,q,ν t,q,ν
+ u(τ, X τt,x,ν
− − Sτ
t,s
+ ντ , q τ− + 1, Sτ
t,s
) − u(τ, X t,x,ν
τ− , q τ− , Sτ
t,s
) d Mτb
t

where M b and M a are the compensated processes associated respectively to N b and N a for
the intensity processes (λbτ )τ and (λaτ )τ .
Now, because each vq is continuous and positive on a compact set, it has a positive lower
γ
bound and vqτ (τ )− k is bounded along the trajectory, independently of the trajectory. Also,
because ν b and ν a are bounded from below, and because of the definition of tn , all the terms
in the above stochastic integrals are bounded and, local martingales being in fact martingales,
we have:
 
t,x,ν t,q,ν t,s
E u(tn , X tn − , qtn − , Stn ) = u(t, x, q, s)
⎡ t
n  
σ2 2
+E ⎣
t,q,ν t,q,ν
∂τ u(τ, X τt,x,ν
− , q τ− , Sτ
t,s
) + ∂ u(τ, X t,x,ν
τ− , q τ− , Sτ
t,s
) dτ
2 ss
t
tn  
t,q,ν t,q,ν
+ u(τ, X τt,x,ν t,s a t,s t,x,ν
− + Sτ + ντ , qτ − − 1, Sτ ) − u(τ, X τ − , qτ − , Sτ ) λτ dτ
t,s a

t
⎤
tn  
+ u(τ, X τt,x,ν
t,q,ν t,x,ν t,q,ν
− − Sτ + ντ , qτ − + 1, Sτ ) − u(τ, X τ − , qτ − , Sτ ) λτ dτ
t,s b t,s t,s b ⎦
t
Using the fact that u solves (HJB), we then have that
 
t,x,ν t,q,ν t,s
E u(tn , X tn − , qtn − , Stn ) ≤ u(t, x, q, s)

with equality when the controls are taken equal the maximizers of the hamiltonians (these
controls being in A because v is bounded and has a positive lower bounded).
Now, if we prove that
   
t,q,ν t,q,ν
lim E u(tn , X tt,x,ν
n−
, q tn − , S t,s
tn ) = E u(T, X t,x,ν
T , q T , S t,s
T )
n→∞
we will have that for all controls in A:
    
t,x,ν t,q,ν t,q,ν
E − exp −γ (X T + qT STt,s ) = E u(T, X Tt,x,ν , qT , STt,s ) ≤ u(t, x, q, s)

with equality for νtb = δ b∗ (t, qt− ) and νta = δ a∗ (t, qt− ). Hence:
  
t,x,ν t,q,ν
sup E − exp −γ (X T + qT STt,s ) = u(t, x, q, s)
(νta )t ,(νtb )t ∈A
  
∗ t,q,δ ∗ t,s
= E − exp −γ (X Tt,x,δ + qT ST )

and this will give the result.

123
Math Finan Econ (2013) 7:477–507 501

It remains to prove that

   
t,q,ν t,q,ν
lim E u(tn , X tt,x,ν
n−
, qtn − , Stt,s
n
) = E u(T, X Tt,x,ν , qT , STt,s )
n→∞

t,q,ν
First, we have, almost surely, that u(tn , X tt,x,ν
n−
, qtn − , Stt,s
n
) tends towards u(T, X Tt,x,ν
− ,
t,q,ν t,s
qT − , ST ). Then, in order to prove that the sequence is uniformly integrable we will bound
it in L 2 . However, because of the uniform lower bound on v already used early, it is sufficient
t,q,ν
to bound exp(−γ (X tt,x,ν
n−
+ qtn − Stt,s
n
)) in L 2 .
But,
tn tn tn
t,q,ν
X tt,x,ν
n−
+ qtn − Stt,s
n
= ντa d Nτa + ντb d Nτb +σ qτt,q,ν dWτ
t t t
tn
≥ −ν−
a
∞ N Ta − ν−
b
∞ N Tb + σ qτt,q,ν dWτ
t

Hence
 
t,x,ν t,q,ν t,s
E exp(−2γ (X tn − + qtn − Stn ))
⎡ ⎛ ⎞⎤
  tn
 
≤ E ⎣exp 2γ ν− a
∞ N Ta exp 2γ ν−
a
∞ N Tb exp ⎝−2γ σ qτt,q,ν dWτ ⎠⎦
t
   1    1
3
≤ E exp 6γ ν−
a
∞ N T a 3 E exp 6γ ν−
b
∞ N Tb
⎡ ⎛ ⎞⎤ 13
tn
×E ⎣exp ⎝−6γ σ qτt,q,ν dWτ ⎠⎦
t

Now, since the intensity of each point process is bounded, the point processes have a
Laplace transform and the first two terms of the product are finite (and independent of n).
t,q,ν
Concerning the third term, because |qτ | is bounded by Q, we know (for instance applying
Girsanov’s theorem) that:
⎡ ⎛ ⎞⎤ 13
tn
   1
E ⎣exp ⎝−6γ σ qτt,q,ν dWτ ⎠⎦ ≤ E exp 3γ 2 σ 2 (tn − t)Q 2 3
t
 
≤ exp γ 2 σ 2 Q 2 T

Hence, the sequence is bounded in L 2 , then uniformly integrable and we have:

   
t,q,ν t,q,ν
lim E u(tn , X tt,x,ν
n − , qtn − , Stt,s
n
) = E u(T, X Tt,x,ν
− , qT − , STt,s )
n→∞
 
t,q,ν
= E u(T, X Tt,x,ν , qT , STt,s )



We have proved that u is the value function and that δ b∗ and δ a∗ are optimal controls.

123
502 Math Finan Econ (2013) 7:477–507

Proof of Theorem 2: Let us first consider the matrix M + 2ηI . This matrix is a symmetric
matrix and it is therefore diagonalizable. Its smallest eigenvalue λ is characterized by:
x  (M + 2ηI )x
λ= inf
x∈R2Q+1 \{0} xx
and the associated eigenvectors x  = 0 are characterized by:
x  (M + 2ηI )x
λ=
xx
It is straightforward to see that:

Q 
Q−1
x  (M + 2ηI )x = αq 2 xq 2 + η (xq+1 − xq )2 + ηx Q 2 + ηx−Q 2
q=−Q q=−Q

Hence, if x is an eigenvector of M + 2ηI associated to λ:

|x| (M + 2ηI )|x|
λ≤
|x| |x|
⎡ ⎤
1 ⎣ Q 
Q−1
= αq 2 |xq |2 + η (|xq+1 | − |xq |)2 + η|x Q |2 + η|x−Q |2 ⎦
|x| |x|
q=−Q q=−Q
⎡ ⎤
1 ⎣  
Q Q−1
≤ αq 2 |xq |2 + η (xq+1 − xq )2 + η|x Q |2 + η|x−Q |2 ⎦ = λ
|x| |x|
q=−Q q=−Q

This proves that |x| is also an eigenvector and that necessarily xq+1 and xq are of the same
sign (i.e. xq xq+1 ≥ 0).
Now, let x ≥ 0 be an eigenvector of M + 2ηI associated to λ.
If for some q with |q| < Q we have xq = 0 then:
0 = λxq = αq 2 xq − η(xq+1 − 2xq + xq−1 ) = −η(xq+1 + xq−1 ) ≤ 0
Hence, because x ≥ 0, both xq+1 and xq−1 are equal to 0. By immediate induction x = 0
and this is a contradiction.
Now, if x Q = 0, then 0 = λx Q = α Q 2 x Q − η(−2x Q + x Q−1 ) = −ηx Q−1 ≤ 0 and hence
x Q−1 = 0. Then, by the preceding reasoning we obtain a contradiction.
Similarly if x−Q = 0, then 0 = λx−Q = α Q 2 x−Q −η(x−Q+1 −2x−Q ) = −ηx−Q+1 ≤ 0
and hence x−Q+1 = 0. Then, as above, we obtain a contradiction.
This proves that any eigenvector x ≥ 0 of M + 2ηI associated to λ verifies in fact x > 0.
Now, if the eigenvalue λ was not simple, there would exist two eigenvectors x and y of
M + 2ηI associated to λ such that |x| y = 0. Hence, y must have positive coordinates and
negative coordinates and since yq yq+1 ≥ 0, we know that there must exist q such that yq = 0.
However, this contradicts our preceding point since |y| ≥ 0 should also be an eigenvector of
M + 2ηI associated to λ and it cannot have therefore coordinates equal to 0.
As a conclusion, the eigenspace of M + 2ηI associated to λ is spanned by a vector f 0 > 0
and we scaled its R2Q+1 -norm to 1.
Now, because M is a symmetric matrix, we can write v(0) = exp(−M T ) × (1, . . . , 1)
as:

2Q
vq (0) = exp(−λi T )g i , (1, . . . , 1) gqi , ∀q ∈ {−Q, . . . , Q}
i=0

123
Math Finan Econ (2013) 7:477–507 503

where λ0 ≤ λ1 ≤ . . . ≤ λ2Q are the eigenvalues of M (in increasing order and repeated if
necessary) and (g i )i an associated orthonormal basis of eigenvectors. Clearly, we can take
g 0 = f 0 . Then, both f q0 and  f 0 , (1, . . . , 1)  are positive and hence different from zero. As
a consequence:
vq (0) ∼T →+∞ exp(−λ0 T ) f 0 , (1, . . . , 1)  f q0 , ∀q ∈ {−Q, . . . , Q}
Then, using the expressions for the optimal quotes, we get:
 
1  γ 1 f q0
lim δ (0, q) = ln 1 +
b∗
+ ln
T →+∞ γ k k 0
f q+1
 
1  γ 1 f q0
lim δ a∗ (0, q) = ln 1 + + ln
T →+∞ γ k k 0
f q−1

Turning to the characterization of f 0 stated in Theorem 2, we just need to write the

Rayleigh ratio associated to the smallest eigenvalue of M + 2ηI :
f0 ∈ argmin f  (M + 2ηI ) f
f ∈R2Q+1 , f 2 =1

Equivalently:


Q 
Q−1
f0 ∈ argmin αq 2 f q 2 + η ( f q+1 − f q )2 + η f Q 2 + η f −Q 2
f ∈R2Q+1 , f 2 =1 q=−Q q=−Q




Proof of Proposition 3: Let us first introduce H = {u ∈ L loc

1 (R)/x  → xu(x) ∈ L 2 (R) and

u ∈ L (R)}.
2
*  
H equipped with the norm u H = αx 2 u(x)2 + ηu  (x)2 d x is an Hilbert space.
R

Step 1: H ⊂ L 2 (R) with continuous injection.

Let us consider u ∈ H and  > 0.
We have:
 
1
u(x)2 d x ≤ 2 x 2 u(x)2 d x < +∞

R\[−,] R\[−,]

Hence because u  ∈ L 2 (R), we have u ∈ H 1 (R \ [−, ]) with a constant C independent

of u such that u H 1 (R\[−,]) ≤ C u H . In particular u is continuous on R∗ .
*1
Now, if  = 1, ∀x ∈ (0, 1), u(x) = u(1) − u  (t)dt and then |u(x)| ≤ |u(1)| +
√ x
1 − xu   L 2 ((0,1)) .
Because the injection of H 1 ((1, +∞)) in C([1, +∞)) is continuous, we know that there
exists a constant C independent of u such that |u(1)| ≤ Cu H 1 ((1,+∞)) . Hence, there
exists a constant C  such that |u(1)| ≤ C  u H and eventually a constant C  such that
u L ∞ ((0,1)) ≤ C  u H . Similarly, we obtain u L ∞ ((−1,0)) ≤ C  u H .
Combining the above inequalities we obtain a new constant K so that u L 2 (R) ≤
K u H . 


123
504 Math Finan Econ (2013) 7:477–507

A consequence of this first step is that H ⊂ H 1 (R) ⊂ C(R).

Step 2: The injection H → L 2 (R) is compact.

Let us consider a sequence (u n )n of functions in H with supn u n  H < +∞.

Because H ⊂ H 1 (R), ∀m ∈ N∗ , we can extract from (u n )n a sequence that converges
in L 2 ((−m, m)). Using then a diagonal extraction, there exists a subsequence of (u n )n , still
denoted (u n )n , and a function u ∈ L loc
2 (R) such that u (x) → u(x) for almost every x ∈ R
n
and u n → u in the L loc (R) sense.
2

Now, by Fatou’s lemma:

 
supn u n 2H
x 2 u(x)2 d x ≤ lim inf x 2 u n (x)2 d x ≤
n→∞ α
R R

Hence, there exists a constant C such that ∀m ∈ N∗ :

 m 
1
|u(x) − u n (x)| d x ≤2
|u(x) − u n (x)|2 d x + x 2 |u(x) − u n (x)|2 d x
m2
R −m R\[−m,m]
m
C
≤ |u(x) − u n (x)|2 d x +
m2
−m
*
Hence lim supn→∞ |u(x) − u n (x)|2 d x ≤ C
m2
.
R
Sending m to +∞ we get:

lim sup |u(x) − u n (x)|2 d x = 0
n→∞
R

Hence (u n )n converges towards u in the L 2 (R) sense.

Now, we consider the equation −ηu  (x) + αx 2 u(x) = f (x) for f ∈ L 2 (R) and we define
u = L f the weak solution of this equation, i.e.:
 
 2  

∀v ∈ H, αx u(x)v(x) + ηu (x)v (x) d x = f (x)v(x)d x
R R

Step 3: L : L 2 (R) → L 2 (R) is a well defined linear operator, compact, positive and self-
adjoint.
*
For f ∈ L 2 (R), v ∈ H  → f (x)v(x)d x is a continuous linear form on H because the
R
injection H → L 2 (R) is continuous. Hence, by Lax-Milgram or Riesz’s representation theo-
rem, there exists a unique u ∈ H weak solution of the above equation and L is a well defined
linear operator.
Now, L f 2H =  f, L f  ≤  f  L 2 (R) L f  L 2 (R) . Hence, since the injection H → L 2 (R)
is continuous, there exists a constant C such that L f 2H ≤ C f  L 2 (R) L f  H , which in
turn gives L f  H ≤ C f  L 2 (R) . Since the injection H → L 2 (R) is compact, we obtain
that L is a compact operator.
L is a positive operator because  f, L f  = L f 2H ≥ 0.

123
Math Finan Econ (2013) 7:477–507 505

Eventually, L is self-adjoint because ∀ f, g ∈ L 2 (R):


 2 
 f, Lg = αx L f (x)Lg(x) + η(L f ) (x)(Lg) (x) d x
R

 
= αx 2 Lg(x)L f (x) + η(Lg) (x)(L f ) (x) d x = g, L f 
R

Now, using the spectral decomposition of L and classical results on Rayleigh ratios we
know that the eigenfunctions f corresponding to the largest eigenvalue λ0 of L satisfy:
1  f H g H
= = inf
λ0  f  L 2 (R) g∈H \{0} g L 2 (R)

Hence, our problem

 boils
 down
 to proving that the largest eigenvalue of L is simple and
1 α 2
that g : x  → exp − 2 η x is an eigenfunction corresponding to this eigenvalue (it is
straightforward that g ∈ H ).
Step 4: Any positive eigenfunction corresponds to the largest eigenvalue of L.
By definition of  ·  H , ∀ f ∈ H, ||f |f |2H =  f H
 f  L 2 (R) . Hence, if f is an eigenfunction of
L (R)
L corresponding to the eigenvalue then | f | is also an eigenfunction of L corresponding
λ0 ,
to the eigenvalue λ0 . Now, if f˜ is an eigenfunction of L corresponding to an eigenvalue
λ  = λ0 , | f |, f˜ = 0. Therefore f˜ cannot be positive.
Step 5 : g spans the eigenspace corresponding to the largest eigenvalue of L.

Differentiating g twice, we get g  (x) = − αη g(x) + αη x 2 g(x).
√
Hence −ηg  (x) + αx 2 g(x) = αηg(x) and g is a positive eigenfunction, necessarily
associated to the eigenvalue λ0 that is therefore equal to √1αη .
Now, if we look for an eigenfunction f ∈ C ∞ (R) ∩ H —because any eigenfunction of L
is in C ∞ (R)—we can look for f of the form f = gh. This gives:
√
0 = −η f  (x) + αx 2 f (x) − αη f (x)
   √
= −η g (x)h(x) + 2g  (x)h  (x) + g(x)h  (x) + αx 2 g(x)h(x) − αηg(x)h(x)
Hence:

α
0 = 2g  (x)h  (x) + g(x)h  (x) = −2x g(x)h  (x) + g(x)h  (x)
η

α 
⇒ h  (x) = 2x h (x)
η
 
 α 2
⇒ ∃K 1 , h (x) = K 1 exp x
η
x  
α 2
⇒ ∃K 1 , K 2 , h(x) = K 1 exp t dt + K 2
η
0
x  
α 2
⇒ ∃K 1 , K 2 , f (x) = K 1 g(x) exp t dt + K 2 g(x)
η
0

123
506 Math Finan Econ (2013) 7:477–507

Now,
x      x  
α 2 1 α 2 α 2
g(x) exp t dt ≥ exp − x exp t dt
η 2 η η
0 √x
2
 
1
≥ x 1− √
2
Hence, for f to be in H , we must have K 1 = 0. Thus, g spans the eigenspace corresponding
to the largest eigenvalue of L and Proposition 3 is proved.

References

1. Amihud, Y., Mendelson, H.: Dealership market. market-making with inventory. J. Fin. Econ. 8,
31–53 (1980)
2. Amihud, Y., Mendelson, H.: Asset pricing and the bid-ask spread. J. Fin. Econ. 17(2), 223–249 (1986)
3. Avellaneda, M., Stoikov, S.: High-frequency trading in a limit order book. Quant. Fin. 8(3), 217–
224 (2008)
4. Bayraktar, E., Ludkovski, M.: Liquidation in limit order books with controlled intensity. Arxiv preprint
arXiv:1105.0247 (2011)
5. Benston Robert, L., George, J.: Determinants of bid-asked spreads in the over-the-counter market. J. Fin.
Econ. 1(4), 353–364 (1974)
6. Cartea, Á., Jaimungal S.: Modeling asset prices for algorithmic and high frequency trading. Working
Paper (2010)
7. Cartea, Á., Jaimungal S.: Risk measures and fine tuning of high frequency trading strategies. Working
Paper (2012)
8. Cartea, Á., Jaimungal, S., Ricci, J.: Buy low sell high: a high frequency trading perspective. Working
Paper (2011)
9. Cohen, K.J., Maier, S.F., Schwartz, R.A., Whitcomb, D.K.: Market makers and the market spread:
a review of recent literature. J. Fin. Quant. Anal. 14(04), 813–835 (1979)
10. Cohen, K.J., Maier, S.F., Schwartz, R.A., Whitcomb, D.K.: Transaction costs, order placement strategy,
and existence of the bid-ask spread. J. Polit. Econ. 89(2), 287–305 (1981)
11. Garman, M.B.: Market microstructure. J. Fin. Econ. 3(3), 257–275 (1976)
12. Guéant, O., Lehalle, C.A.: Existence and Uniqueness for the Avellaneda–Stoikov PDE. Working Paper.
(2012)
13. Guéant, O., Lehalle, C.A., Fernandez-Tapia, J.: Optimal Execution with Limit Orders. Working Paper.
SIAM J. Fin. Math. (2011) (to appear)
14. Guilbaud, F., Pham, H.: Optimal high frequency trading with limit and market orders. Quant. Fin. (2011)
(to appear)
15. Guilbaud F., Pham H.: Optimal high frequency trading in a pro-rata microstructure with predictive infor-
mation. Arxiv preprint arXiv:1205-3051 (2012)
16. Hendershott, T., Menkveld, A.: Price Pressures. Manuscript, VU University, Amsterdam (2009)
17. Ho, T.S.Y., Macris, R.G.: Dealer bid-ask quotes and transaction prices: An empirical study of some AMEX
options. J. Fin. 39, 23–45 (1984)
18. Ho, T., Stoll, H.R.: Optimal dealer pricing under transactions and return uncertainty. J. Fin. Econ. 9(1),
47–73 (1981)
19. Ho, T.S.Y., Stoll, H.R.: The dynamics of dealer markets under competition. J. Fin. 38(4), 1053–1074 (1983)
20. Lehalle, C.A., Guéant, O., Razafinimanana, J. : High Frequency Simulations of an Order Book: a Two-
Scales Approach. In: Abergel, F., Chakrabarti, B.K., Chakraborti, A., Mitra, M. (eds.) Econophysics of
Order-Driven Markets, New Economic Windows, Springer, New York (2010)
21. Madhavan, A., Smidt, S.: An analysis of changes in specialist inventories and quotations.
J. Fin. 48(5), 1595–1628 (1993)
22. Menkveld, A.J.: High Frequency Trading and The New-Market Makers. Social Science Research Network
Working Paper Series (2010)
23. Mildenstein, E., Schleef, H.: The optimal pricing policy of a monopolistic marketmaker in the equity
market. J. Fin. 38, 218–231, (1983)

123
Math Finan Econ (2013) 7:477–507 507

24. O’Hara, M., Oldfield, G.S.: The microeconomics of market making. J. Fin. Quant. Anal. 21(04),
361–376 (1986)
25. Roll, R.: A simple implicit measure of the effective bid-ask spread in an efficient market.
J. Fin. 39(4), 1127–1139 (1984)

123