A BELIEF-DRIVEN ORDER BOOK MODEL

RENÉ CARMONA AND KEVIN WEBSTER

1. Introduction
As first pioneered by [19] and treated in detail in [9] one of the crucial roles of
market microstructure is price revelation. Different trading systems have emerged
and evolved, from auctions to limit order books and dark pools. Each of them cor-
respond to a different way of handling the risk of adverse selection, highlighting the
importance played by private beliefs in the trading process. Agents either attempt
to protect their own information, or try to infer that of others. This game of cat
and mouse between liquidity takers and providers, between informed and uniformed
traders highlights the need for models where different beliefs compete on a given
market structure. This paper focuses on the limit order book, a microstructure
used in more than half the world’s exchanges.
1.1. Classification of models. Several surveys of microstructure models exist
([17, 9, 5, 10]) but we will follow the methodology of [9] to situate our model.
”Market microstructure models can be classified along at least four dimensions:
type of orders, sequence of moves, price setting rule, and competitive versus strate-
gic structure.” Our model features limit orders and market orders on both the bid
and the ask. Traders place both types of orders simultaneously. The pricing is
discriminatory: the price of a unit of volume depends on how much has already
been traded. Finally, our model is competitive: the interactions between players
are of mean-field type, and each agent therefore ignores their own impact when the
number of players is large.
The aim of the model is to relate as closely as possible the belief-distribution of
the different agents to the shape of the order book. While time-dependences such as
price impact, order book resilience and decay of information are also important, we
have removed them from the current model, which happens in one period. A direct
generalization in continuous time is possible, but the basic intuitions provided will
be the same, such that these time-dependence question of the order book will be
directly related to time related properties of the agent belief distribution.
This paper was strongly motivated by a series of studies underlining the empirical
properties of the limit order book on electronic exchanges ([6, 22, 8, 7, 25, 23, 16,
11]). In particular, the ’hump shape’ of the order book was successfully reproduced
using reasonable assumptions on the agent belief distribution. The family of models
most closely related are informed trader models ([19, 13]): differing beliefs on the
price play an important role, although our model does not require noise traders to
function. Other approaches focus on inventory risk of the liquidity provider rather

Date: August 24, 2012.

Key words and phrases. Limit-order book, Transaction costs, Equilibrium, Microstructure,
Mean-field interaction.
1
2 RENÉ CARMONA AND KEVIN WEBSTER

than on the adverse selection phenomenon ([14, 4]). Another class of models are so
called ’zero-intelligence’ models ([15, 21, 12, 20, 18]), where strategic considerations
are discarded to see which characteristics could be explained by microstructure only.
Finally, market impact models([2, 3, 26, 1, 24]) essentially directly model the order
book properties without referring to the underlying agents forming it.
1.2. Results. The paper is structured in three parts. First the setting of the
model is introduced. The agents, the probability space and their differing beliefs
are defined. The controls are introduced and their interactions and effects on other
variables made explicit. Transaction costs are defined as well as a duality rela-
tionship with the order book proved. The section ends by writing up each agent’s
optimization problem. The second section explores partial equilibrium. The opti-
mal trading strategy is derived first, followed by the limit order posting strategy.
These highlight the link between agent beliefs and the distribution of trades and
orders. Multiple consequences and several examples conclude the section. The last
section searches for Nash equilibria using a mean field game approach. The exis-
tence of an equilibrium is derived in the infinite player setting and proved to be
an approximate Nash equilibrium in the case of a finite number of agents on the
market.
The main contributions are:
• The duality relationship between order book and transaction costs, which
defines a codebook for trades.
• The optimality equations and the definition of implied alpha, which pro-
vides some economic background to the codebook and links beliefs to trades
and orders.
• The Nash equilibrium result and final shape of the order book.

2. Setup
The aim of this section is to introduce the setting and notation to model agents
with diverse beliefs engaging in trades on a limit order book. The setup of the
model begins with the players and their controls. We introduce the probability
space and the order book they interact on. In particular, the different agents’
actions lead to transaction costs, which we link back to the original order structure
of the market. Next, we describe how these interactions on the order book level
affect the players’ asset positions and total wealth. The derived equations hold
regardless of the number of players, which is reflected in the final notation. Finally,
each agent is given an objective function to maximize. The section ends with a
summary of each agent’s optimization problem.
2.1. Players and the order book. We first start describing the players of our
game, their beliefs and the setting in which they trade.
Assume there are agents indexed on a set I with cardinal n ∈ [2, ∞], each having
a control (li , g i ) in the control space R × M(R). Let (Ω, F, P) be a probability
space and pT a random variable representing the price of an asset. The initial
deterministic price is p0 . Each agent possesses a different probability measure
Pi representing his beliefs and which is dominated by the reference measure P.
The first component li of each agent’s control is a real number representing the
signed volume he wishes to obtain from the market. The second component g i is
a non-negative measure which represents the orders he places on the limit order
 A BELIEF-DRIVEN ORDER BOOK MODEL 3

book1. Assume the exchange restricts the orders to be of size one, making each g i
a probability measure. We will also recenter the measures by shifting p0 to 0. The
limit order book is then just the sum of all the orders, which we will renormalize
with respect to P the total number of agents. If there is a countable number of agents,
n
g = limn→∞ n1 i=1 g i , where the limit is assumed to exist in the weak sense of
distributions. All the orders placed on (0, ∞) are offers, and those on (−∞, 0) are
bids.
The order book will be split so that each agent will face independent copy g
on which they trade simultaneously2. When an agent trades a volume li on the
order book g, he will consume all the orders until he gets filled for the requested
i
volume.
 This
R α meansi he will execute all the orders from 0 to the level f (l ) =
inf α ∈ R : 0 g = l , given that an agent will always prefer cheaper orders to
more expensive ones. The function f is the generalized inverse of the antiderivative
R li
of g. The total amount of money the agent will pay is therefore p0 li + 0 f (l)dl.
Because p0 li is the amount he would have paid in the absence of microstructure,
R li
c(li ) = 0 f (l)dl are the transaction costs associated to the trade and represent
the extra cost incurred by the agent for taking liquidity from the order book. f
is increasing and hence c is a convex function. This chain of computations can be
summarized by the following proposition:
Proposition 2.1 (Duality between order book and transaction costs). Let g be
an order book and c the associated transaction costs. Define γ as the Legendre
transform of c:
γ(α) = sup αl − c(l) (2.1)
l∈supp(c)
then we have the relationship
γ 00 = g (2.2)
where the second derivative is to be understood in the sense of distributions.
Proof. By construction, c is differentiable and the anti-derivative of g is the gener-
alized inverse of c0 .RBy Fenchel’s identity, γ 0 and c0 are generalized inverses of each
·
other. Hence γ 0 = 0 g, which concludes. 
The following properties are standard results in convex analysis:
Corollary 2.2. c is a convex, differentiable function with compact support (in the
sense of convex functions) and
c(l) + γ(α) ≥ αl (2.3)
0 0
for all (α, l) with equality if c (l) = α, or equivalently, l = γ (α).
Proof. g being a non-negative measure with finite mass implies that γ 0 is non-
decreasing and bounded. Its generalized inverse c0 is therefore non-decreasing and
has compact support. This implies that c is convex and has compact support.
The Legendre transform is a bijection for convex functions, making c the Legendre
transform of γ. The result follows by Fenchel’s inequality. 
1In practice, these are Dirac masses placed on a discrete grid and the resulting order book
becomes a histogram of prices.
2This is to avoid having to assign priorities. One can also consider that each agent only trades
once, and that orders are refilled in between each trade until the new price PT is revealed.
4 RENÉ CARMONA AND KEVIN WEBSTER

This allows us to define a codebook for trade volumes. For a given order book,
we define αi = c0 (li ), which also implies the relationship li = γ 0 (αi ). Furthermore,
(1) c(0) = 0. Agents that do not trade incur no transaction costs.
(2) c0 (0) = 0. Transaction costs are defined with respect to the current price
p0 .
(3) γ(0) = 0, γ 0 (0) = 0. These are dual relationships to (1) and (2).
This defines our players, the order book and by duality their incurred transaction
costs.

2.2. Position and wealth. This subsection describes more in detail the interac-
tions between the agents’ control variables and how those affect their portfolio and
wealth.
Each agent simultaneously posts both an order measure γ 00i = g i and a trade
volume li . The order book is then defined as the renormalized sum ofPthe orders
and we likewise define the renormalized measure of trades, µl = n1 i∈I δli , or
equivalently, the measure on the dual variables µ = n1 i∈I δαi . In the countable
P
agent case, limits must be taken and assumed to exist under the Wasserstein metric.
Both are probability measures and we will further assume in the countable case that
µ has a finite first moment. For the sake of convenience, we define µi = δαi . All
orders and trades are considered as simultaneous and the aggregate measures are
public knowledge3.
Let us make more explicit the transfers to agent i triggered by a trade of agent
j. Agent j only perceives the aggregate order book γ 00 and asks for a volume lj ,
or equivalently, executes all order up to the level αj . The volume lj verifies the
duality relationship lj = γ 0 (αj ) = γ 0 , µj . The transfer from player j to player i
however, will be equal to lji = n1 γ 0i (αj ) = γ 0i , µj 4. This corresponds to all the
orders of player i that arePplaced between 0 and αj . The additivity of the order
book ensures that lj = n1 i∈I lji (with a limit in the countable

case).
Similarly, agent j’s cash transfer for trading lj is c lj = αj lj − γ(αj ) by the
duality relationship. Expressing everything in terms of the dual variables leads to
the formula c(lj ) = idγ 0 − γ, µj . The following lemma introduces a convenient
adjoint operator.
Lemma 2.3. For any limit order book γ 00 and any probability measure µ of order
1 with a finite first moment, we have
h(id − α∗ )γ 0 − γ, µi = hγ 00 , Lα∗ µi (2.4)
where we define
Z ∞ 
∗
Lα∗ µ = (id − α ) µ − 1·≤0 (2.5)
·
R∞
Proof. Lα∗ µ is a well defined function because (α − α∗ ) α µ ≤ |α∗ | + R |id|µ for
R

α > 0. A similar result holds when α ≤ 0. Successive integrations by parts and

3The exact rules of what is and is not known depends on the specific market considered.
On NASDAQ, orders can be traced to individual agents, but not trades. In Foreign Exchange,
sometimes both orders and trades can be traced back to market participants.
4Remember that each agent’s order is split equally into one of the n independent copies of the
aggregate order book.
 A BELIEF-DRIVEN ORDER BOOK MODEL 5

simplifications then yield:

 Z ∞   Z ∞ 
00 ∗
γ , (id − α ) µ − 1·≤0 = − γ0, ∗ ∗
µ − 1·≤0 − (id − α )µ − α δ0
· ·
= hγ, µ − δ0 i + hγ 0 (id − α∗ ), µi + α∗ γ 0 (0)
= h(id − α∗ )γ 0 − γ, µi
where the first integration by parts is justified by the fact that γ 0 is bounded, the
Dirac distribution has compact support and the equations
Z ∞
lim µ=0 (2.6)
α→∞ α
Z ∞ Z α
lim µ − 1α≤0 = − lim µ=0 (2.7)
α→−∞ α α→−∞ −∞
(2.8)
and αµ(α) vanishes at the infinities. The linear growth of γ justifies the second
integration by parts. The last line uses the initial conditions of γ and γ 0 . 
This allows us to rewrite the transaction costs as c(lj ) = γ 00 , L0 µj . The transfer
of money from agent j to i is n1 γ 00i , L0 µj .
The position of agent i is therefore
Li = γ 0 , µi − γ 0i , µ (2.9)
While the wealth of agent i at time T is
X i = Li (pT − p0 ) + γ 00i , L0 µ − γ 00 , L0 µi (2.10)
where these formulas are hence well-defined both in the finite and infinite agent
case.
What is important to note is that all the interaction between the agents are of
mean field type. Indeed, each agent interacts with all the others only through his
own control variables and the aggregate variables µ and γ. This is what allows the
notation to extend naturally from the finite to the infinite case. It is also what will
allow us in the last section to use the infinite player limit as an approximation to
the finite player game.
2.3. Objective function. The setup of the game ends with the definition of each
agent’s objective function. We then summarize the game from the beginning: the
players, the control and state variables and the objective

function.
Each agent now has a utility function U i X i , pT which depends on his wealth
and the price5. Assume that each γ 00i is constrained to be a probability measure.
Assume U i to be differentiable and concave in its first variable and the controls
chosen such as to verify the integrability conditions
EPi U i X i , pT < ∞


(2.11)
0 < EPi ∂X U i X i , pT < ∞


(2.12)
Finally, define
J i αi , γ 00i , µ, γ = EPi U i X i , pT



(2.13)
as agent i’s objective function.
5This covers Merton-type problems, indifference pricing and statistical arbitrage.
6 RENÉ CARMONA AND KEVIN WEBSTER

To summarize, each agent therefore faces the optimization problem:

EPi U i X i , pT


max
i 00i
(2.14)
(α ,γ )

subject to
Li = γ 0 , µi − γ 0i , µ

(2.15)
Xi = Li (pT − p0 ) + γ 00i , L0 µ − γ 00 , L0 µi
where µi = δαi and γ 00i must be a probability measure.

3. Partial equilibrium
In this section, we will focus on one particular player i. To maintain the mean-
field structure of our game, we assume the price to be a (random) function only
of the aggregate quantities µ and γ. In particular, any price impact is symmet-
ric amongst players, and in the infinite player limit, each agent’s price impact is
negligeable. In partial equilibrium, we will assume our agent has been given the
price exogenously, which will be strictly true in the infinite player limit and serve
as a stepping stone in the finite player equilibrium. We first derive the optimal
strategy in terms of trades and order placement for our agent. Then we study some
consequences of collective optimal behavior.
3.1. Optimality equations. The agent’s optimization problem is straightforward
to solve. The optimal trading strategy is obtained by differentiating the objective
function in li . The order placing strategy is derived from the affine nature of the
objective function in γ 00i , forcing the optimal solution to be an extremal point of
our permissible set.
The first result gives us a natural way of naming the dual variable αi :
Theorem 3.1 (Implied alpha). For a given µ and γ, the optimal trading strategy
of player i is
αi = EQi [pT − p0 ] (3.1)
i i
i ∂ U (X ,pT)
where dQ
X
dPi = E i ∂X U i (X i ,pT ) .
P

Proof. This result is more easily shown with the primal variable li . Rewriting the
state variables yields Li = li − γ 0i , µ and X i = Li (pT − p0 ) + γ 00i , L0 µ − c(li ).
Note that X i is a concave function of li . 

Because U i is concave and increasing in X, EPi U i X i , pT is a concave func-
tion of li . We can therefore simply equate the derivative to zero. This yields the
equation:
EPi pT − p0 − c0 (li ) ∂X U i X i , pT = 0



(3.2)
which simplifies to
EPi ∂X U i X i , pT c0 (li ) = EPi (pT − p0 ) ∂X U i X i , pT

 

(3.3)
The expectation on the left hand side is positive and finite by assumption, allowing
us to do the announced change of measure. The identity c0 (li ) = αi then concludes.

The quantity pT − p0 corresponds to what practitioners call the realized alpha
of the trade. Motivated by the above result, we name the dual variable αi the
implied alpha of the trade. The theorem states that the implied alpha is equal to
the agent’s expectation of the realized alpha under some new probability measure
 A BELIEF-DRIVEN ORDER BOOK MODEL 7

Qi . Because we have that Qi = Pi in the risk-neutral case, we will name Qi the

risk-neutral probability measure of agent i.
The second result links the agent’s orders to µ, the aggregate distribution of
implied alphas.
Proposition 3.2 (Profitability of an order). For a given µ and γ, player i’s optimal
strategy for orders is
γ 00i = δα̃i (3.4)
where
α̃i = argmax Lαi µ (3.5)
Proof. First we verify that the argmax is well defined. The function m = Lαi µ
is continuous on (−∞, 0] and (0, +∞) and vanishes at the infinities. Furthermore,
both m(0+ ) and m(0) are finite and of opposing signs. Hence one of them is positive
and m reaches its supremum either at a point of continuity,
0 or 0+ .6
Let all the variables but γ 00i be fixed. Denote by X i γ 00i the affine functional

γ 00i −→ γ 00i , LpT −p0 µ − γ 00 , LpT −p0 µi (3.6)

The concavity of U i implies that
∆J := EPi U i X i (γ 00i ), pT − U i X i (δα̃i ), pT



(3.7)
verifies
γ 00i − δα̃i , LpT −pt µ ∂X U i X i (γ 00i ), pT


∆J ≤ EPi
≤ EPi ∂X U i X i (γ 00i ), pT EQi γ 00i − δα̃i , LpT −p0 µ



= EPi ∂X U i X i (γ 00i ), pT

 00i
γ − δα̃i , Lαi µ
The first term is positive by hypothesis. The second term is non-positive as
γ 00i , Lαi µ ≤ ||Lαi µ||∞
= hδα̃i , Lαi µi

where we have used the fact that γ 00i is a probability measure. This concludes. 

The operator Lαi quantifies the expected profitability of an order. Indeed,

Z ∞ 
Lαi µ(α) = α − αi


µ − 1α≤0 (3.8)
| {z } α
exp. spread
| {z }
filling probability

is the expected gain under Qi of an order placed at the level α. The absolute value
of the first term is equal to the spread the liquidity provider expects to gain per
filled order. The second term, up to a sign, is equal to the proportion of agents that
will fill the order. If the agents of our model arrived according to Poisson arrival
times instead of simultaneously, this would be the filling probability of the order.
These two terms are commonly used by practitioners.

6The difference between 0+ and 0− is that one is a buy and the other one a sell order at the
current price p0 .
8 RENÉ CARMONA AND KEVIN WEBSTER

3.2. Corollaries. The objective of this subsection is to explore some consequences

of the above results when adding additional hypotheses. We start with a result
linking the average implied alpha to the market risk premium.
Corollary 3.3 (Market implied price). Assume pT − p0 to be bounded a.s. and all
the agents to be exchangeable7 under P. Then we have
hid, µi = EP [pT − p0 ] (3.9)
and the inequality
hf, µi ≤ EP f (pT − p0 ) (3.10)
for f a convex function such that f (pT − p0 ) is a.s. bounded.
Proof. We have the chain of relationships
n
1X
hf, µi = lim f (EPi [pT − p0 ]) (3.11)
n→∞ n
i=1
n
1X  i 
≤ lim EP Z f (pT − p0 ) (3.12)
n→∞ n
i=1
" n
#
1X i
= EP f (pT − p0 ) lim Z (3.13)
n→∞ n
i=1
= EP f (pT − p0 ) (3.14)
i
where Z i = dQ dP . Fubini’s theorem in (3.13) is justified because f (pT − p0 ) is
bounded, and the Z i are uniformly bounded. (3.12) is an application of Jensen’s
inequality and is an equality for f = id. The exchangeability assumption allows us
to apply the law of large numbers to the Z i which results in (3.14). 

The above result suggests a very intuitive model of the dependence of pT in µ:

pT = p0 + hid, µi + N (3.15)
with N of mean zero. The empirical average of the implied alpha is therefore equal
to the market risk premium. If N is now exogeneous, then pT is uniformly Lipschitz
under the Wasserstein metric, which will matter to find an equilibrium.
The corollary also implies that µ is not quite a pricing measure, as it underes-
timates events far away from the mean. Amusingly, γ 00 does exactly the opposite,
pushing away weight from the average scenario into the periphery.
Corollary 3.4 (The double hump shape of the order book). Assume µ has a
positive, differentiable L2 density and that agents are constrained
R∞ to place orders
on the same side as their trades. Assume furthermore −µ0 α µ < 2µ2 for α > 0
(and the symmetric relationship for α < 0). Then αi and α̃i verify
Z ±∞
µ(α̃i ) α̃i − αi =


µ (3.16)
α̃i

7That is, consider all their intrinsic variables, utility functions and Radon-Nikodym derivatives
dPi
dP
to be assigned in an exchangeable manner at the beginning of the game.
 A BELIEF-DRIVEN ORDER BOOK MODEL 9

R ±∞ R∞ R ±∞ R α̃i
where α̃i µ = α̃i µ if αi > 0 and α̃i µ = − −∞ µ else. In particular, agent
i’s spread is
Z ±∞
1
µ (3.17)
µ(α̃i ) α̃i
Proof. Without loss of generality we can assume αi > 0. The additional constraint
implies that
α̃i = argmaxα>0 m(α) (3.18)
where m = Lαi µ. We have that m(0) < 0, m(αi ) = 0 and m(∞) = 0. Furthermore,
m is differentiable and
Z ∞
m0 (α) = µ − α − αi µ(α)


(3.19)
α
0 +
R∞ i
m (0 ) = 0 µ + α µ(0) > 0 and m(∞) = 0 as µ has finite first moment. Because
µ > 0, the function
Z ∞
1
M (α) = µ − α + αi (3.20)
µ(α) α
has the same sign as m0 (α). Differentiating yields:
R∞
0 µ(α)2 + µ0 (α) α µ
M (α) = − −1 (3.21)
µ(α)2
which is negative by hypothesis. Hence M is decreasing and m0 has a unique zero.
m0 (0+ ) > 0 and hence m has a unique maximum α̃i , which verifies m0 (α̃i ) = 0.
This concludes. 

This provides us with a methodology to derive γ 00 from µ. The inverse problem

is of course harder, but still possible for parametric choices of µ. It is easy to
verify numerically that, for a unimodal ’bell-shaped’ µ probability measure, γ 00 is
a bimodal ’double-hump’ distribution. This coincides with empirical studies of the
limit order book.
Example 3.5. In the case of a Laplace distribution of parameter λ
λ −λ|α|
µ(α) = e (3.22)
2
The Laplace distribution verifies verifies all the hypotheses except differentiability
at 0. The spread is then constant equal to
Z ±∞
1 1
i
µ = (3.23)
µ(α̃ ) α̃i λ
and γ 00 is a Laplace distribution that is shifted λ1 away from 0 on both sides. In
this case there is a well defined best bid and best ask, and the bid-ask spread is equal
to λ2 . This distribution exhibits a peak of liquidity at the best bid and ask, contrary
to what is commonly seen on the market.
Example 3.6. The standard Gaussian distribution verifies all the hypotheses. The
spread does not have an explicit formula, but can be computed numerically. In this
case, the best bid and ask are limited only by the granularity of the market and the
order book has a peak of liquidity behind the best bid and best ask.
10 RENÉ CARMONA AND KEVIN WEBSTER

Figure 1. Simulated double-hump shape of the order book in the

case of a Laplace distribution for µ.

Figure 2. Simulated double-hump shape of the order book in the

case of a Gaussian distribution for µ.

The above optimality equations and their consequences illustrate the usefulness
of the proposed codebook for trades and naturally fits with terms commonly used
in the industry. They can be derived without too much difficulty in a continuous
time framework, even in a setting where agents have differing time horizons and
filtrations.

4. Nash equilibria
The natural notion of equilibrium to pursue in our game is of Nash type. The
mean field nature of the interactions between players allowed us to derive simple
optimality equations in partial equilibrium. Unfortunately, these only hold as long
 A BELIEF-DRIVEN ORDER BOOK MODEL 11

as each agent ignores his own price impact. This will be true in the limit where the
number of players is infinite, but will only approximately hold for a finite but large
number of players. The first subsection therefore explores the infinite player case.
The fixed point theorem proved requires some strong assumptions: in particular, the
agents are assumed to be risk-neutral. In the second subsection the infinite player
strategy is given to players in a finite setting and proved to be an approximate Nash
equilibrium. In both cases, only questions of existence are explored.

4.1. The infinite player limit. By taking the limit as the number of players goes
to infinity, the problem gains in tractability by making everything happen ’in a
continuum’. It also guarantees the absence of price impact for any single agent,
although the players possess, in aggregate, impact over the market.
Without loss of generality, we can assume that p0 = 0 and replace each agent’s
second variable γ 00i by α̃i with the relationship γ 00i = δα̃i . Furthermore assume that
pT lives in the compact set [−K, K] and that
dPi
= Z(θi ) (4.1)
dP
i
with Z uniformly continuous
1
Pn i 1
Pθn such that, as n → ∞ the empirical dis-
in θ and the
tributions νn = n j=1 δθj and νn = n j6=i δθj converge to a probability measure
ν. Assume ν has a continuous density that we abusively also denote by ν. Define
the functional
h  i
J(α, α̃, µ, γ 00 , θ) = E Z(θ) L∗pT (µ) γ 00 (α) − LpT (µ) µ(α̃)
= L∗p(θ,µ) γ 00 (α) − Lp(θ,µ) µ(α̃)
with p(θ, µ) = E [Z(θ)p(µ)]. Assume p is continuous in the second argument with
respect to the Wasserstein metric, uniformly over the first argument. Then the
objective function for player i in the finite agent case is equal to:
J i ( αj j=1..n , (α̃)j=1..n ) = J(αi , α̃i , µn , γn00 , θi )

1 1
= J(αi , α̃i , µin + δ i , γ 00i + δα̃i , θi )
n α n n
Pn Pn
where µn = n1 i=1 δαj and µin = n1 j6=i δαj . γn00 and γn00i are defined similarly.
The idea of the mean field game methodology is to take the limit as n → ∞,
making the dependence of J in the player’s control disappear from the third and
fourth variable. The following theorem shows the existence of a solution to the
limiting problem.
Theorem 4.1 (Infinite player problem).
R 2 Define the Hilbert space L2 (ν) as the
functions f : R → R that verify f ν < ∞. Then there exists a pair of functions
αν and α̃ν in L2 (ν) such that
(αν , α̃ν ) (θ) = argmaxa,ã J(a, ã, µ, γ 00 , θ) ν-a.s. (4.2)
and
µ = ν ◦ αν−1 ; γ 00 = ν ◦ α̃ν−1 (4.3)
Proof. By (3.1), we need to find a fixed point to the map
M α(θ) = p(θ, ν ◦ α−1 ) (4.4)
12 RENÉ CARMONA AND KEVIN WEBSTER

as the optimality condition for α̃ just reads

α̃(θ) = argmax Lα(θ) µ (4.5)
which does not require a fixed point once α is solved for. By assumption, ||M (α)||∞ ≤
K. The ball of radius K is therefore stable and weakly compact with respect to
the considered Hilbert norm. Hence, all we need to show is the weak continuity of
M to conclude by Schauder’s fixed point theorem. First, note that if αn converges
weakly to α, then
Z Z
|f (αn (θ)) − f (α(θ))| ν(θ)dθ ≤ |αn (θ) − α(θ)| ν(θ)dθ → 0 (4.6)

for any 1-Lipschitz function f . Hence ν ◦ αn−1 → ν ◦ α−1 under the Wasserstein
metric. This concludes. 
i i


Corollary 4.2 (Nash equilibrium). The strategy α(θ ), α̃(θ ) i∈I is a Nash equi-
librium.


Proof. By construction, α(θi ), δα̃i verifies the optimality equations for player i
when given the aggregate distributions µ = ν ◦ α−1 and γ 00 = ν ◦ α̃−1 . Hence we
have that for all (β, β̃)
J(β, β̃, ν ◦ α−1 , ν ◦ α̃−1 , θi ) ≤ J(α(θi ), α̃(θi ), ν ◦ α−1 , ν ◦ α̃−1 , θi ) (4.7)

Because we used Schauder’s fixed point theorem, uniqueness of the equilibrium
remains elusive. Furthermore, the absence of continuity of the second optimality
equation makes it hard to treat the case of non-risk neutral agents. On a positive
note, however, the methodology can be extended to the continuous time case using
the mean field game paradigm, as long as the price impact function remains of
mean field type.
4.2. Approximate Nash equilibrium. Next, we want to use the previous solu-
tion as a strategy in the finite player case. The following lemma will help us show
that it leads to an approximate Nash equilibrium.
Lemma 4.3. For given αi , α̃i , θi , νn and ν let
1 1
en = J(αi , α̃i , νni ◦ α−1 + δαi , νni ◦ α̃−1 + δα̃i , θi ) − J(αi , α̃i , ν ◦ α−1 , ν ◦ α̃−1 , θi )
n n
(4.8)
Then there exists n ≥ 2|en | such that n → 0 as n → ∞,uniformly over αi , α̃i
and θi .
Proof. By definition, the first term is equal to
   
1 1
Lp(θi ,νni ◦α−1 + n1 δαi ) δαi , νni ◦ α̃−1 + δα̃i − L∗p(θi ,ν i ◦α−1 + 1 δ i ) δα̃i , νni ◦ α−1 + δαi
n n n α n
D E D E
= Lp(θi ,νni ◦α−1 + n1 δαi ) δαi , νni ◦ α̃−1 − L∗p(θi ,ν i ◦α−1 + 1 δ i ) δα̃i , νni ◦ α−1
n n α

By the uniform continuity of p in its second variable, Lp(θi ,νni ◦α−1 + n1 δαi ) δαi con-
verges weakly to the function Lp(θi ,ν◦α−1 ) δαi . That the convergence is uniform over
θi is obvious and the uniformity over αi can be derived from the explicit formula of
Lp δαi (a). By the portmanteau theorem, as ν is continuous, α ◦ Lp(θi ,ν◦α−1 ) δαi , νni
converges to α ◦ Lp(θi ,ν◦α−1 ) δαi , ν . This concludes. 
 A BELIEF-DRIVEN ORDER BOOK MODEL 13

Finally,


Theorem 4.4 (Approximate Nash equilibrium). The strategy α(θi ), α̃(θi ) i=1..n
is an n -Nash equilibrium. That is, for all (β, β̃) and all i ∈ {1...n} we have that
1 1
J(β, β̃, νni ◦α−1 + δβ , νni ◦ α̃−1 + δβ̃ , θi ) ≤ J(α(θi ), α̃(θi ), νn ◦α−1 , νn ◦ α̃−1 , θi )+n
n n
(4.9)
Proof. By the previous lemma, we have the chain of inequalities
1 1 1
J(β, β̃, νni ◦ α−1 + δβ , νni ◦ α̃−1 + δβ̃ , θi ) ≤ J(β, β̃, ν ◦ α−1 , ν ◦ α̃−1 , θi ) + n
n n 2
1
≤ J(α(θi ), α̃(θi ), ν ◦ α−1 , ν ◦ α̃−1 , θi ) + n
2
≤ J(α(θi ), α̃(θi ), νn ◦ α−1 , νn ◦ α̃−1 , θi ) + n


5. Conclusion
This paper presents a one time-step model designed to link the shape of the order
book to the heterogenous beliefs of the large number of agents trading on it. In
particular, it proposes a codebook for trades that, in equilibrium, summarizes the
agents’ beliefs on the price. The new variable introduced, dubbed ’implied alpha’
to echo the regular notion of alpha, is readily obtained from transaction costs using
a duality relationship. The framework is then used to shed some light on coherent
price impact, price formation and order book shaping mechanisms. Using ideas
borrowed from mean field games, the question of existence of Nash equilibria is
answered positively in the infinite player limit. Existence of approximate Nash
equilibria in the finite player case is also guaranteed.
A natural extension of the framework would be to answer similar questions in
continuous time. Some work has been made to solve the partial equilibrium in a
Gaussian filtration setting with heterogeneous filtrations and probability measures
on a given price process. Although all these additional details do not fundamentally
change the proof method, they do enrich the model significantly by introducing
more trade opportunities into the market. Differing time horizons for the agents
can also add some realism -and additional trading opportunities- without too much
hassle. The question of Nash equilibrium requires more thought as the mean field
nature of the game becomes a dominant feature of the continuous time model.

References
[1] A. Alfonsi, A. Fruth, and A. Schied. Optimal execution strategies in limit order books with
general shape functions. Quantitative Finance, 10(2):143–157, 2010.
[2] R. Almgren and N. Chriss. Optimal execution of portfolio transactions. Journal of Risk,
3(2):5–39, 2000.
[3] R. F. Almgren. Optimal execution with nonlinear impact functions and trading-enhanced
risk. Applied Mathematical Finance, 10(1):1–18, 2003.
[4] Y. Amihud and H. Mendelson. Asset pricing and the bid-ask spread. Journal of Financial
Economics, 17(2):223–249, 1986.
[5] B. Biais, L. Glosten, and C. Spatt. Market microstructure: A survey of microfoundations,
empirical results, and policy implications. Journal of Financial Markets, 8(2):217–264, 2005.
[6] B. Biais, P. Hillion, and C. Spatt. An empirical analysis of the limit order book and the order
flow in the paris bourse. Journal of Finance, 50:1655–1689, 1995.
14 RENÉ CARMONA AND KEVIN WEBSTER

[7] J. P. Bouchaud, Y. Gefen, M. Potters, and M. Wyart. Fluctuations and response in financial
markets: The subtle nature of ’random’ price changes. Quantitative Finance, 4(2):176–190,
2004.
[8] J. P. Bouchaud, M. Mzard, and M. Potters. Statistical properties of stock order books:
Empirical results and models. Quantitative Finance, 2(4):251–256, 2002.
[9] M. Brunnermeier. Asset pricing under asymmetric information. Oxford University Press,
2001.
[10] A. Chakraborti, I.M. Toke, M. Patriarca, and F. Abergel. Econophysics: Agent-based models.
Quantitative Finance, 11(1):1013–1041, 2011.
[11] A. Chakraborti, I.M. Toke, M. Patriarca, and F. Abergel. Econophysics: Empirical facts.
Quantitative Finance, 11(1):991–1012, 2011.
[12] R. Cont, S. Stoikov, and R. Talreja. A stochastic model for order book dynamics. Operations
research, 58(3):549–563, 2010.
[13] D. Easley and M. O’Hara. Price, trade size, and information in securities markets. Journal
of Financial Economics, 19(1):69–90, 1987.
[14] M. B. Garman. Market microstructure. Journal of Financial Economics, 3(3):257–275, 1976.
[15] D. K. Gode and S. Sunder. What makes markets allocationally efficient? Quarterly Journal
of Economics, 112(2):603–630, 1997.
[16] G. F. Gu, W. Chen, and W. X. Zhou. Empirical shape function of limit-order books in the
chinese stock market. Physica A: Statistical Mechanics and its Applications, 387(21):5182–
5188, 2008.
[17] J. Hasbrouck. Empirical market microstructure. Oxford University Press, 2007.
[18] H. Huang and A. N. Kercheval. A generalized birth-death stochastic model for high-frequency
order book dynamics. Quantitative Finance, 12(4), 2012.
[19] A. S. Kyle. Continuous auctions and insider trading. Econometrica, 53(6):1315–1335, 1985.
[20] D. Ladley and K. R. Schenk-Hopp. Do stylised facts of order book markets need strategic
behaviour? Journal of Economic Dynamics and Control, 33(4):817–831, 2009.
[21] S. Maslov. Simple model of a limit order-driven market. Physica A: Statistical Mechanics
and its Applications, 278(3):571–578, 2000.
[22] S. Maslov and M. Mills. Price fluctuations from the order book perspective - empirical facts
and a simple model. Physica A: Statistical Mechanics and its Applications, 299(1-2):234–246,
2001.
[23] R. Naes and J. A. Skjeltorp. Order book characteristics and the volume-volatility relation:
Empirical evidence from a limit order market. Journal of Financial Markets, 9(4):408–432,
2006.
[24] A. Obizhaeva and J. Wang. Optimal trading strategy and supply/demand dynamics. Preprint,
2005.
[25] P. Weber and B. Rosenow. Order book approach to price impact. Quantitative Finance,
5(4):357–364, 2005.
[26] M. Wyart, J. P. Bouchaud, J. Kockelkoren, M. Potters, and M. Vettorazzo. Relation between
bid-ask spread, impact and volatility in order-driven markets. Quantitative Finance, 8(1):41–
57, 2008.