Dynamic Games and Applications (2021) 11:556–579

https://doi.org/10.1007/s13235-020-00373-w

Dynamic Equilibrium of Market Making with Price

Competition

Jialiang Luo1 · Harry Zheng1

Accepted: 6 November 2020 / Published online: 10 December 2020

© The Author(s) 2020

Abstract
In this paper, we discuss the dynamic equilibrium of market making with price competition
and incomplete information. The arrival of market sell/buy orders follows a pure jump pro-
cess with intensity depending on bid/ask spreads among market makers and having a looping
countermonotonic structure. We solve the problem with the nonzero-sum stochastic differ-
ential game approach and characterize the equilibrium value function with a coupled system
of Hamilton–Jacobi nonlinear ordinary differential equations. We prove, do not assume a
priori, that the generalized Issac’s condition is satisfied, which ensures the existence and
uniqueness of Nash equilibrium. We also perform some numerical tests that show our model
produces tighter bid/ask spreads than those derived using a benchmark model without price
competition, which indicates the market liquidity would be enhanced in the presence of price
competition of market makers.

Keywords Dynamic equilibrium · Market making · Price competition · Nonzero-sum

stochastic differential game · Generalized Issac’s condition

Mathematics Subject Classification 93E20 · 90C39

JEL Classification C7 · G1

1 Introduction

Market makers play an important role in providing liquidity for other market participants.
They keep quoting bid and ask prices at which they stand ready to buy and sell for a wide
variety of assets simultaneously. One of the key challenges faced by market makers is to
manage inventory risk. Market makers need to decide bid/ask prices which influence both
their profit margins and accumulation of inventory. Many market makers compete for market

B Harry Zheng
h.zheng@imperial.ac.uk
Jialiang Luo
jialiang.luo14@imperial.ac.uk
1 Department of Mathematics, Imperial College, London SW7 2AZ, UK
Dynamic Games and Applications (2021) 11:556–579 557

order flows as their profits come from the bid/ask spread of each transaction. Traders choose
to buy/sell at the most competitive prices offered in the market. Hence, market makers face
a complex optimization problem. In this paper, we model market making for a single asset
with price competition as a nonzero-sum stochastic differential game.
There has been active research on optimal market making in the literature with focus on
inventory risk management. Stochastic control and Hamilton–Jacobi–Bellman (HJB) equa-
tion, a nonlinear partial differential equation (PDE), are used to derive the optimal bid/ask
spread. Ho and Stoll [13] give the first prototype model for the market making problem. Avel-
laneda and Stoikov [2] propose a basic trading model in which the asset mid-price follows a
Brownian motion, market buy/sell order arrivals follow a Poisson process with exponentially
decreasing intensity function of bid/ask spread, and market makers optimally set the bid/ask
spread to maximize the expected utility of the terminal wealth. Guéant et al. [11] discuss
a quote-driven market and include the inventory penalty for terminal utility maximization.
Guéant [10] extends the model in Guéant et al. [11] to a general intensity function and
reduces the dimensionality of the HJB equation for CARA utility. Cartea and Jaimungal [6]
consider the market impact and capture the clustering effect of market order arrivals with a
self-exciting process driven by informative market orders and news events and solve the HJB
equation by an asymptotic method. Cartea et al. [5] study the model uncertainty, similar to
Avellaneda and Stoikov [2], Guéant et al. [11], except for the self-exciting feature of market
order arrivals. Fodra and Pham [8] divide the market orders depending on the size which
may bump up the mid-price that follows a Markov renewal process. Abergel et al. [1] discuss
a pure jump model for optimal market making on the limit order book with the Markov
decision process technique conditioned on the jump time clock.
One common feature in the aforementioned papers is that market order arrivals follow
a Poison process with controlled intensity. The probability that a market maker buys/sells
a security at the bid/ask price she quotes is a function of her own bid/ask spread only.
This setting provides tractability, but ignores the influence of prices offered by other market
makers. The price competition between market makers in practice is an important trading
factor and needs to be integrated in the model. Kyle adopts the game-theoretic approach in
a number of papers [14–16] to study the price competition between market participants of
informed traders, noisy traders and market makers, finds the equilibrium explicitly and shows
its impact on price formation and market liquidity. To the best knowledge of the authors, there
are no known results in the literature on price competition between market makers who keep
trading to profit from bid/ask spread while minimizing inventory risk and improving market
liquidity. The primary motivation of this paper is to fill this gap. Market making with price
competition is the key difference of our model to that of Guéant et al. [11] and others in
the literature. The standard optimal stochastic control is not applicable to our model due to
the looping dependence structure, and the equilibrium control is used instead to solve the
problem.
The main contributions of this paper are the following: Firstly, we discuss price compe-
tition between market makers in a continuous-time setting with inventory constraints and
incomplete market information of competitors’ inventory and extend the classical optimal
market making framework in Avellaneda and Stoikov [2] with the game-theoretic approach.
Secondly, we prove the existence and uniqueness of Nash equilibrium for the game under
linear quadratic payoff and prove the generalized Issac’s condition is satisfied for a system
of nonlinear ordinary differential equations (ODEs), rather than assuming it to hold a priori
or solving it explicitly as in the most literature, see [3,4,12,17]. Thirdly, we perform some
numerical tests to compute the equilibrium value function and equilibrium controls (bid/ask
spreads) and compare results with those from a benchmark model without price competition,
558 Dynamic Games and Applications (2021) 11:556–579

and we find our model reduces the bid/ask spread and improves the asset liquidity in the
market considerably.
The rest of the paper is organized as follows. In Sect. 2, we introduce the model setup
and notations. In Sect. 3, we state the main results on the existence and uniqueness of Nash
equilibrium, the generalized Issac’s condition and the verification theorem for the equilibrium
value function. In Sect. 4, we perform numerical tests to show the impact of price competition
and compare the results with a benchmark model without price competition. In Sect. 5, we
prove the main results (Theorems 3.3 and 3.4). Section 6 concludes.

2 Model

Consider a market in a probability space (, F , P) with homogeneous market makers in a

set mm . Choose one of them as a reference market maker, whose states include time variable
t ∈ [0, T ], asset reference price St , cash position X t and the inventory position qt . St is public
information known to all market makers, whereas X t and qt are each market maker’s private
information. The reference asset price St is assumed to follow a Gaussian process

d St = σ dWt ,

where W is a standard Brownian motion adapted to the filtration {Ft }t∈R+ , generated by
W and augmented with all P-null sets, and σ is a constant representing asset volatility.
The terminal time T is small, normally a day, the probability that St becomes negative is
negligible, and we may assume St is always positive. Market makers do not buy/sell the asset
at the reference price, but at bid and ask prices, and make profit from the bid/ask spread.
Denote by a a buying order and b a selling order. The reference market maker’s bid price Stb
and ask price Sta are given by

Stb = St − δtb , Sta = St + δta ,

where δtb and δta are the bid and ask spreads controlled by the reference market maker.
At time t, other market makers also quote bid and ask prices simultaneously to compete
with the reference market maker. Among their quotes, there exist a lowest ask price and a
highest bid price, which are the most competitive prices other than reference market maker’s
prices. Denote by ka the market maker who provides the lowest ask price Skaa ,t , and kb the
market maker who provides the highest bid price Skbb ,t ; in other words, δkbb ,t and δkaa ,t are the
lowest bid and ask spreads among the reference market maker’s competitors.
Traders tend to sell/buy at the most competitive bid/ask price, but may accept less com-
petitive prices due to other factors such as liquidation of large quantities. From the reference
market maker’s perspective, the arrival of buying/selling orders is unpredictable, but the
intensities depend on both her bid/ask spreads and the most competitive ones. The lower her
bid/ask spreads to the most competitive ones, the more likely they are to be hit by traders.
Hence, the arrival intensity is decreasing in terms of her spread and increasing in the most
competitive spread. The arrival of selling market order Ntb and that of buying market order
Nta are Poisson processes with controlled intensities λbt and λat , defined by

λat = f (δta , δkaa ,t ), λbt = f (δtb , δkbb ,t ),

where f is the intensity function. Denote by f 1 the first-order partial derivative of f to its
 the second-order partial derivative of f to its first variable, etc.
first variable, f 11
Dynamic Games and Applications (2021) 11:556–579 559

Assumption 2.1 Assume f is twice continuously differentiable and for all δ, x, y ∈ R,

f  (δ,δ)
f (δ, x) > 0, f 1 (δ, x) < 0, f 2 (δ, x) ≥ 0, limδ→+∞ − f1(δ,δ) > 0, and

f (δ, x) f 11 (δ, y) − 2 f 1 (δ, x) f 1 (δ, y) + | f 1 (δ, x) f 2 (δ, y) − f 12

(δ, y) f (δ, x)| < 0.
(2.1)
Furthermore, assume there exists a twice continuously differentiable function λ : R → R
such that f (δ, x) ≤ λ(δ) for all x ∈ R, limδ→+∞ λ(δ)δ = 0 and λ(δ)λ (δ) < 2(λ (δ))2 .

Some conditions in Assumption 2.1 are technical and needed in the proof. Many functions
satisfy these conditions, for example, f (δ, x) = λ(δ)g(x), where λ is the one in Assumption
 (δ)
2.1 with negative first-order derivative and limδ→+∞ − λλ(δ) > 0, and g is increasing, positive
and bounded. Here is another example:

e−aδ
f (δ, x) := √ , (2.2)
1 + 3ek(δ−x)
where  is the magnitude of market order arrival rate, a the decay rate, k the dependence rate
of the difference between
√
reference market maker’s price and the most competitive price in
the market with a ≥ 2 k > 0. It is easy to check that f satisfies all conditions in Assumption
2

2.1. Some simple functions may not satisfy Assumption 2.1. For example, a constant function
is excluded; if it were allowed, it would imply the size of bid/ask spread does not affect the
arrival rate for market makers, clearly unrealistic.
We assume there is an inventory position constraint for all market makers. Let Q =
{−Q, . . . , Q} be a finite set of integers with Q and −Q the maximum and minimum positions
a market maker may hold and qt ∈ Q. When qt = Q (or −Q), market maker cannot buy
(or sell) any more. Denote by I b and I a the indicator functions of market maker’s buying or
selling capability:

I b (q) := 1{q+1∈Q} , I a (q) := 1{q−1∈Q} ,

where 1 A is an indicator that equals 1 if A is true and 0 if A is false. When market maker’s
bid price is hit by a market order (Ntb increases by 1), her inventory qt increases by 1 and
she pays Stb for buying the asset. Similarly, when market maker’s ask price is hit by a market
order (Nta increases by 1), her inventory qt decreases by 1 and she receives Sta for selling the
asset. The dynamics of cash X t and inventory qt are given by

d X t = Sta I a (qt )d Nta − Stb I b (qt )d Ntb

dqt = I b (qt )d Ntb − I a (qt )d Nta

with the initial condition (X 0 , q0 ) = (x, q) ∈ R × Q.

The reference market maker does not have complete information on the whole market.
Denote by (xkb , qkb ) and (xka , qka ) the states of market makers kb and ka , respectively. They
are random variables from the reference market maker’s perspective, as her competitors’ states
are not public information. The reference market maker can only deduce the probability
distribution for both (xkb , qkb ) and (xka , qka ) based on available public information. We
assume their probability distributions are known and time-invariant. They are Pb for (xkb , qkb )
and Pa for (xka , qka ). This incomplete information assumption is a reasonable approximation
of real market. We next use a heuristic example to illustrate the incomplete information setting
and Pa and Pb .
560 Dynamic Games and Applications (2021) 11:556–579

Example 2.2 Consider at time t there are 3 market makers quoting in the market including the
reference market maker. Their potential states, corresponding probability and bid/ask spread
are assumed by following table.

x q Probability Bid spread (bps) Ask spread (bps)

0 −1 1 10 50
3
0 0 1 30 30
3
0 1 1 50 10
3

For simplicity, we assume they all have same cash position x = 0 and there are only three
inventory possibilities q = −1, 0, 1. Assume uniform probability on q = −1, 0, 1. When
q = −1, market maker will prefer to buy than sell. Hence, they will quote lower bid spread
10bps and higher ask spread 50bps. For q = 1, it is the opposite. Denote the inventory of
the reference market maker’s two competitors as q1 and q2 . We can calculate Pa as
1
Pa (0, −1) = P(q1 = −1)P(q2 = −1) =
9
1
Pa (0, 0) = P(q1 = −1)P(q2 = 0) + P(q1 = 0)P(q2 = −1) + P(q1 = 0)P(q2 = 0) =
3
5
Pa (0, 1) = 1 − (Pa (0, −1) + Pa (0, 0)) = .
9
Take Pa (0, −1) as an example. It is the probability that market maker among the two that
quotes the lowest ask spread has inventory −1, which implies both market makers have
inventory q1 = q2 = −1 as otherwise a lower ask spread 30 bps or 10 bps would be quoted
if one of them had inventory 0 or 1. Other values for Pa and Pb can be calculated similarly.
We assume market makers take closed-loop feedback strategies that are deterministic
functions of state variables at time t, that is, there exist functions δ a and δ b such that bid/ask
spreads of market maker are given by

δta = δ a (t, S, x, q), δtb = δ b (t, S, x, q).

Denote by Aa and Ab the sets of all δ a and δ b that are lower bounded square integrable
measurable functions, δ := (δ b , δ a ) ∈ Ab × Aa reference market maker’s strategy, δ  :=
{δ m , m ∈ mm } the collection of all market makers’ strategies, so reference market maker’s
strategy δ ∈ δ  . Using the game theory convention, we may label the reference market maker
as 0 and δ −0 the set of strategies of all other market makers in mm except the reference
market maker, i.e., δ −0 := {δ m , m = 0, m ∈ mm }.
As everyone else in mm can be reference market maker’s competitor when a market
order arrives, their strategies influence her expected market order arrival intensity. Reference
market maker’s cash and inventory are determined by her own strategy δ as well as those in
the set δ −0 . Starting at time t ∈ [0, T ] with initial asset price S, cash x and inventory q, the
reference market maker wants to maximize the following payoff function:
  T 
1
J (δ, δ −0 , t, S, x, q) = Et X T + qT ST − l(|qT |) − γ σ 2 (qs )2 ds , (2.3)
2 t
Dynamic Games and Applications (2021) 11:556–579 561

where Et is the conditional expectation operator given St = S, X t = x and qt = q. The

reference market maker wants to maximize the expected value of terminal wealth while
penalizes the holding inventory at terminal time T and throughout the time interval [0, T ]
with γ a positive constant representing the risk adverse level and l an increasing convex
function on R+ with l(0) = 0, denoting the liquidity penalty for holding inventory at T .
Due to the circular dependence nature among market makers and their strategies, we use a
game-theoretic approach to solve the problem. We next define the Nash equilibrium.

Definition 2.3 We call the Nash equilibrium exists for a game G mm if there exists an equi-
librium control profile δ ∗ = {δ ∗m , m ∈ mm }, such that for every reference player 0 in
mm , given her strategy δ ∗ ∈ δ ∗ and other players’ strategy set δ ∗−0 , her payoff satisfies the
following equilibrium condition:
J (δ ∗ , δ ∗−0 , t, S, x, q) = max J (δ, δ ∗−0 , t, S, x, q). (2.4)
δ∈Ab ×Aa

Moreover, the reference market maker’s equilibrium control is δ ∗ and the equilibrium value
function is
V (t, S, x, q) := J (δ ∗ , δ ∗−0 , t, S, x, q). (2.5)

3 Main Results

In this section, we prove the existence and uniqueness of Nash equilibrium for G mm when
price competition is in place. We first reduce the model’s dimension by ansatz, then charac-
terize the equilibrium value function by a system of nonlinear ODEs, prove the verification
theorem and finally show the existence and uniqueness of Nash equilibrium by an equivalent
ODE system.
Writing the integral form of X T and qT in payoff function (2.3) with Ito’s lemma, we can
simplify the equilibrium value function V as

V (t, S, x, q) = x + q S + θq (t), (3.1)

where θq : [0, T ] → R is defined by

 T
1
θq (t) = sup Et [ [δsa f (δsa , δkaa ,s ) + δsb f (δsb , δkbb ,s ) − γ σ 2 qs2 ]ds − l(|qT |)]
δ∈A ×A
b a t 2
(3.2)
with Et being the conditional expectation operator given qt = q. Since process qt takes value
in a finite set Q, it is a Markov chain with M = 2Q + 1 states. Hence, game G mm is reduced
to a continuous-time finite-state stochastic game. Define a function θ : [0, T ] → R M as

θ (t) = (θ−Q (t), . . . , θ Q (t)). (3.3)

The equilibrium bid/ask spreads only depend on state qt at time t. As market makers are
homogeneous, under equilibrium at time t, any two market makers with the same state q
quote the same bid/ask spread, denoted by πqb (t) and πqa (t), respectively. Note that πqb (t)
exists for every q ∈ Q except q = Q when market maker reaches the maximum inventory
and stops quoting bid price. πqa (t) is similarly defined. We can define the equilibrium control
as

π a (t) = (π−Q+1
a
(t), . . . , π Q
a
(t)), π b (t) = (π−Q
b
(t), . . . , π Q−1
b
(t)).
562 Dynamic Games and Applications (2021) 11:556–579

The market maker’s equilibrium control δ ∗ = ((δ a )∗ , (δ b )∗ ) is given by

(δ a )∗ (t, S, x, q) = πqa (t), (δ b )∗ (t, S, x, q) = πqb (t). (3.4)

When market order arrives at time t, the reference market maker expects her most competitive
market maker in bid side to have inventory q with probability Pqb and in ask side Pqa . As
there are only finite number of states, the most competitive market maker’s state probability
is given by:

P a = (P−Q+1
a
, . . . , PQa ), P b = (P−Q
b
, . . . , PQ−1
b
).

Market makers with inventory on boundary values do not quote in the market, so P−Q a =
PQ = 0.
b

We next provide a characterization for the value function θ and the equilibrium controls
π a , π b . Applying the dynamic programming principle, we get the following Hamilton Jacobi
ODE system:
1 2 2
θq (t) = γ σ q − sup ηa (θ (t), δ, π a (t), q)I a (q) − sup ηb (θ (t), δ, π b (t), q)I b (q)
2 δ δ
θq (T ) = −l(|q|)
πqa (t) ∈ argsupηa (θ (t), δ, π a (t), q), ∀q ∈ {−Q + 1, . . . , Q}
δ
πqb (t) ∈ argsupηb (θ (t), δ, π b (t), q), ∀q ∈ {−Q, . . . , Q − 1},
δ
(3.5)
where ηa , ηb : R M ×R×R M−1 ×Q → R are defined by vectors μ = (μ−Q , . . . , μ Q ) ∈ R M ,
ξ a = (ξ−Q+1
a , . . . , ξ Qa ) or ξ b = (ξ−Q
b , . . . , ξb
Q−1 ) as


Q
ηa (μ, δ, ξ a , q) := P ja f (δ, ξ aj )(δ + μq−1 − μq )
j=−Q+1
(3.6)

Q−1
ηb (μ, δ, ξ b , q) := P jb f (δ, ξ bj )(δ + μq+1 − μq ).
j=−Q

Q  Q−1
j=−Q+1 P j f (δ, π j (t)) and j=−Q P j f (δ, π j (t)) are reference market
Note that a a b b

maker’s expected intensity of buying/selling market order arrival when her spread is δ and
other market makers take the equilibrium control. We can now characterize the Nash equi-
librium.

Theorem 3.1 Assume the Nash equilibrium of the game G mm exists. Then, the equilibrium
value function V can be decomposed as (3.1) with function θ . Equilibrium control δ ∗ can be
written as (3.4) with two vectors π a (t) and π b (t). Moreover, θ , π a (t) and π b (t) satisfy the
ODE system in (3.5).

The equilibrium condition for π a (t) and π b (t) in (3.5) leads to the following generalized
Issac’s condition, which is also defined in Cohen and Fedyashov [7] to ensure the existence
of Nash equilibrium for nonzero-sum stochastic differential game and a natural extension of
the standard Issac’s condition in the zero-sum game to the nonzero-sum game.
Dynamic Games and Applications (2021) 11:556–579 563

Definition 3.2 We call the generalized Issac’s condition holds if there exist functions wa , w b :
R M → R M−1 such that for any vector μ ∈ R M ,
ηa (μ, wqa (μ), wa (μ), q) = sup ηa (μ, δ, wa (μ), q), ∀q ∈ {−Q + 1, . . . , Q}
δ
(3.7)
η b
(μ, wqb (μ), w b (μ), q) = sup ηb (μ, δ, w b (μ), q), ∀q ∈ {−Q, . . . , Q − 1},
δ

where wqa , wqb : RM → R and wa , w b are defined by

wa (μ) := (w−Q+1
a
(μ), . . . , waQ (μ)), w b (μ) := (w−Q
b
(μ), . . . , w bQ−1 (μ)).

If the generalized Issac’s condition is satisfied, we can substitute the function wa , w b into
operators ηa , ηb , and system (3.5) is reduced to the following ODE system:
1 2 2
θq (t) = γ σ q − ηa (θ (t), wqa (θ (t)), wa (θ (t)), q)I a (q)
2 (3.8)
− ηb (θ (t), wqb (θ (t)), w b (θ (t)), q)I b (q)θq (T ) = −l(|q|).
We next state the verification theorem.

Theorem 3.3 Assume that f satisfies Assumption 2.1 and that there exist bounded strategies
π a , π b and function θ on [0, T ] satisfying system (3.5). Then, the Nash equilibrium of the
game G mm exists. The equilibrium value function is given by (3.1) and the equilibrium control
by (3.4).

From Theorems 3.1 and 3.3, we know the existence and uniqueness of Nash equilibrium
for game G mm are equivalent to the existence and uniqueness of equilibrium controls π a , π b
and function θ that satisfy the ODE system (3.5). We now state the main result of the paper.

Theorem 3.4 Assume f satisfies Assumption 2.1. Then, there exists a unique Nash equilib-
rium for game G mm . Specifically, there exist unique locally Lipschitz continuous functions
wa , w b that satisfy the generalized Issac’s condition in Definition 3.2, and there exists unique
classical solution θ to the ODE system (3.8), such that the equilibrium value function is given
by (3.1) and the equilibrium controls by
π a (t) = wa (θ (t)), π b (t) = w b (θ (t)), t ∈ [0, T ]. (3.9)

4 Numerical Test

In this section, we numerically find the Nash equilibrium value function and bid/ask spread
when there is price competition with the intensity f defined in (2.2) and compare the
numerical results with those derived using a benchmark model in Guéant [10] without price
competition and with the intensity f˜(δ) := 0.5e−aδ and the liquidity penalty l(q) := 0.1q 2 .
To make two models comparable, we define parameters for f and f˜ in a way that when every
market maker provides the same bid/ask spread, the intensity of market order arrivals is the
same in both cases, which gives 0.5 in the definition of f˜. The parameters of both models
are set as follows:
Here, S is the initial asset value, N the number of time steps in discretization, T the period
of one day, σ the daily volatility, a and  used in intensity functions, γ inventory penalty
coefficient and Q the inventory maximum capacity. Furthermore, probabilities of the most
564 Dynamic Games and Applications (2021) 11:556–579

S σ (daily) γ k a  T (day) N Q

20.0 0.01 1.0 2.0 2.0 60.0 1.0 100 10

Fig. 1 Ask spread strategy profile at time 0.5

competitive market makers’ state P a and P b are assumed to be given by (see Example 2.2
for explanation of P a and P b )
a
P−10 = P10
b
=0
P0a = P0b = 0.2
P1a = P−1
b
= 0.4
P2a = P−2
b
= 0.3
Pqa = 1/170, q = −10, 0, 1, 2
Pqb = 1/170, q = 10, 0, −1, −2.
Figures 1 and 2 plot the equilibrium bid/ask spreads of both models at time 0.5. We note
that higher inventory leads to lower ask spread but higher bid spread, indicating the preference
of market makers to sell rather than to buy in order to remain inventory neutral and that the
equilibrium bid/ask spreads of our model are tighter than those of the benchmark model,
indicating improved market liquidity.
Figure 3 plots the equilibrium ask spreads with different inventory levels on [0, T ]. Market
makers with positive inventory are more willing to sell and clear their positions due to the
liquidity punishment at terminal time T , and this willingness increases as time nears T as
the equilibrium ask spread is decreasing when t tends to T . For market makers with negative
inventory, it is opposite. This explains empirical facts that trading volume increases at the
end of the day.
Figure 4 plots the expected intensity functions in terms of bid/ask spread at time 0.5, which

are given by G b (δ) = f˜(δ) for the benchmark model and G(δ) = Qj=−Q+1 P ja f (δ, π aj (t))
for our model, respectively. The one from our model is derived endogenously from equilib-
rium, while the one assumed by the benchmark model comes from Avellaneda and Stoikov
Dynamic Games and Applications (2021) 11:556–579 565

Fig. 2 Bid spread strategy profile at time 0.5

Fig. 3 Equilibrium ask spread for competition model

[2] in which the distribution of market order size and the statistics of the market impact are
used. When price competition is in place, the market order arrival intensity decays faster,
indicating that when price competition is in place but market maker assumes there were not,
they would tend to overestimate the market order arrival intensity and quote higher bid/ask
spreads.
Figures 5 and 6 plot the equilibrium value function θ near the starting time 0 and the
terminal time T , respectively. We note that θ with price competition takes lower value than
the one without at time 0.1 but performs better at time 0.9, especially when there are still
large inventories to be liquidated, as market makers of the benchmark model overestimate
the arrival intensity, which results in higher spreads and worse performance.
In summary, when price competition between market makers is in place, market maker
tends to quote tighter bid/ask spreads and the market has better liquidity and lower transaction
566 Dynamic Games and Applications (2021) 11:556–579

Fig. 4 Intensity versus ask quote at time 0.5

Fig. 5 Value function θ at time 0.1

cost. However, the profit of market maker is reduced. The value function is lower when there
is competition between market makers.

5 Proofs of Theorems 3.3 and 3.4

5.1 Proof of Theorem 3.3

Proof To verify that (δ  )∗ is the equilibrium control profile and V is the equilibrium value
function, it is sufficient to check that they satisfy the equilibrium condition in (2.4). For any
market maker in mm , given other market makers’ strategies in (δ  )∗ and any admissible
Dynamic Games and Applications (2021) 11:556–579 567

Fig. 6 Value function θ at time 0.9

strategy δ, we should prove:

J (δ, (δ −0 )∗ , t, S, x, q) ≤ J ((δ)∗ , (δ −0 )∗ , t, S, x, q) = V (t, S, x, q).
Assume the reference market maker takes the arbitrary strategy δ, while every other
market maker decides his/her bid/ask spread by (δ a )∗ (t, St , X t , qt ) = πqat (t) and
(δ b )∗ (t, St , X t , qt ) = πqbt (t). Denote reference market maker’s cash position at any time
t as X t∗,δ , while their inventory is qt∗,δ . Then, for any time t ∈ [0, T ], by ansatz (3.1) and
Itô’s lemma with respect to function θ , we get the following:

V (T , ST , X T∗,δ , qT∗,δ ) = X T∗,δ + qT∗,δ ST + θq ∗,δ (T ) = x + q S + θq (t)

T
 T  T  T  T
+ δub I b (qu∗,δ )dNub + δua I a (qu∗,δ )d Nua + qu∗,δ dSu + θ  ∗,δ (u)du
t t t t qu
 T  T
+ (θq ∗,δ +1 (u) − θq ∗,δ (u))I b (qu∗,δ )dNub + (θq ∗,δ −1 (u) − θq ∗,δ (u))I a (qu∗,δ )dNua .
u u u u
t t
(5.1)
As qu∗,δ takes value in finite set Q, and the solution for ODE exists on compact set [0, T ],
we know both θq (u) and θq (u) are uniformly bounded on [0, T ] for all q ∈ Q and:
 T   T 
E (qu∗,δ )2 du < +∞, E (θ  ∗,δ (u))2 du < +∞.
t t qu

Moreover, from assumption that f (δ, x) ≤ λ(δ) for all x, we have admissible control satisfies
(see [10, page 16]):
⎡ ⎤
Q  T
E⎣ Pja
f (δu , π j (t))I (qu )|δu + θq ∗,δ −1 (u) − θq ∗,δ (u)|du ⎦ < +∞
a a a ∗,δ a
u u
j=−Q+1 t
⎡ ⎤

Q−1  T
E⎣ P jb f (δub , π bj (t))I b (qu∗,δ )|δub + θq ∗,δ +1 (u) − θq ∗,δ (u)|du ⎦ < +∞.
u u
j=−Q t
568 Dynamic Games and Applications (2021) 11:556–579

Taking expectation on both sides of (5.1), we have:

 T 
E V (T , ST , X T∗,δ , qT∗,δ ) = V (t, S, x, q) + E θ  ∗,δ (u)du
t qu
 T 
+E ηa (θ (u), δua , π a (u), qu∗,δ )I a (qu∗,δ )du
t
 T 
+E ηb (θ (u), δub , π b (u), qu∗,δ )I b (qu∗,δ )du .
t

where ηa and ηb are defined in (3.6). Hence, we have:

 T 
E V (T , ST , X T∗,δ , qT∗,δ ) ≤ V (t, S, x, q) + E θ  ∗,δ (u)du
t qu
   
T T
+E sup ηa (θ(u), δua , π a (u), qu∗,δ )I a (qu∗,δ )du + E sup ηb (θ(u), δub , π b (u), qu∗,δ )I b (qu∗,δ )du .
t δua t δub
(5.2)
As θ satisfies ODE system (3.5) for every u ∈ [0, T ]. We substitute it into the corresponding
part in (5.2) and have following.
  
∗ ∗,δ ∗,δ 1 2 T ∗,δ 2
J (δ, (δ −0 ) , t, S, x, q) = E V (T , ST , X T , qT ) − γ σ (qu ) du ≤ V (t, S, x, q).
2 t

On the other hand, if the reference market maker also takes equilibrium control, her cash
position and inventory are denoted by X t∗ and qt∗ , respectively. And we have the following:
ηa (θ (t), πqa (t), π a (t), q) = sup ηa (θ (t), δ, π a (t), q),
δ
ηb (θ (t), πqb (t), π b (t), q) = sup ηb (θ (t), δ, π b (t), q).
δ

Substituting the equilibrium control defined in (3.4) to (5.2) can conclude the proof as
following:
  T 
1
J ((δ)∗ , (δ −0 )∗ , t, S, x, q) = E V (T , ST , X T∗ , qT∗ ) − γ σ 2 (qu∗ )2 du
2 t
= V (t, S, x, q) ≥ J (δ, (δ −0 )∗ , t, S, x, q).

5.2 Proof of Theorem 3.4

The proof of Theorem 3.4 is made of three steps:

1. There exist functions wa , w b such that for any vector μ ∈ R M , wa (μ) and w b (μ) satisfy
Equation (3.7).
2. wa and w b are unique and locally Lipschitz continuous, which guarantees RHS of the
ODE system (3.8) is also locally Lipschitz continuous.
3. There exists unique classical solution to ODE system (3.8).
The key step for proving Steps 1 and 2 is to characterize the vectors wa (μ) and w b (μ)
by the first-order condition of Hamiltonian. They are the solutions to some equation system.
Then, we can prove step 1 and 2 by discussing the zero point for the equation system. The key
Dynamic Games and Applications (2021) 11:556–579 569

step for proving Step 3 is to obtain upper bound estimation for θ . It can be done by showing
θ is also a solution to another system of ODE, which admits the comparison principle, and
hence upper bound for its solution. Without confusion of notations, we write wa (μ) and
w b (μ) as,

wa (μ) = wa = (w−Q+1
a
, . . . , waQ ), w b (μ) = w b = (w−Q
b
, . . . , w bQ−1 ).

5.2.1 Proof of Step 1

We first show that wa and w b satisfy the equilibrium condition of the Hamiltonian system.
We provide some preliminary results for the existence and uniqueness of the maximum point
for Hamiltonian G qa (δ) := ηa (μ, δ, w, q) given any vector μ ∈ R M and w ∈ R M−1 . We can
define G qb (δ) and prove the same result similarly.

Lemma 5.1 Assume intensity function f satisfies all the assumptions in Theorem 2.1. Then,
given any vectors w = (w−Q+1 , . . . , w Q ) ∈ R M−1 and μ, the maximum point exists and
is unique for function G qa when q = −Q + 1, . . . , Q. Furthermore, the maximum point of
G qa (δ) satisfies the first-order condition:

dG qa (δ)
= 0.
dδ
Proof Given any vector μ and w, the expected intensity function d is defined by


Q
d(δ) := a
P j2 f (δ, w j ).
j=−Q+1

From Assumption 2.1, we know for any δ, x and y:

 
f (δ, x) f 11 (δ, y) + f (δ, y) f 11 (δ, x) < 4 f 1 (δ, x) f 1 (δ, y). (5.3)

Simple calculation shows

d(δ)d  (δ) < 2(d  (δ))2 ,

which implies δ + μq−1 − μq + d(δ)/d  (δ) is a strictly increasing function of δ. Combining

dG a (δ ∗ )
with d  (δ) < 0, we conclude that there exists a unique δ ∗ such that dδ
q
= 0 and G qa (δ)
is strictly increasing for δ < δ and strictly decreasing for δ > δ , that is, δ ∗ is the unique
∗ ∗

global maximum point of G qa .

Step 1 is equivalent to following theorem, which proves that the generalized Issac’s con-
dition in Definition 3.2 holds for any vector μ ∈ R M . We only focus on wa , as the proof of
w b is similar.

Theorem 5.2 Assume the intensity function f satisfies Assumption 2.1. Then, for any fixed
vector μ = (μ−Q , . . . , μ Q ) ∈ R M , there exists vector wa = (w−Q+1
a , . . . , waQ ) such that
for q = −Q + 1, . . . , Q,

wqa = argmax{ηa (μ, δ, wa , q)}. (5.4)

δ
570 Dynamic Games and Applications (2021) 11:556–579

Define a mapping T : R M−1 → R M−1 as

Tq (w) = argmax{ηa (μ, δ, w, q)}, ∀q ∈ {−Q + 1, . . . , Q}
δ∈R (5.5)
T (w) := (T−Q+1 (w), . . . , TQ (w)),
(5.4) is equivalent to wa = T (wa ), namely, wa is a fixed point of mapping T . We need the
following Schauder fixed-point theorem to prove the existence of wa .

Theorem 5.3 (Schauder) If K is a nonempty convex closed subset of a Hausdorff topological

vector space V and T is a continuous mapping of K into itself such that T (K ) is contained
in a compact subset of K , then T has a fixed point.

To apply Theorem 5.3, we need to show the existence of K and the continuity of T . The
next lemma confirms the first requirement.

Lemma 5.4 Given any vector μ = (μ−Q . . . , μ Q ) ∈ R M and mapping T defined in (5.5),
there exists a nonempty convex compact set K ⊂ R M−1 such that T (K ) ⊂ K .

Proof Firstly, for any vector w ∈ R M−1 , define y = (y−Q+1 , . . . , y Q ) = T (w). There exist
a uniform δmin ∈ R such that for every q,
yq ≥ δmin . (5.6)
We can prove by contradiction. Assume there was no lower bound for yq . Defining
G qa (δ) = ηa (μ, δ, y, q) for q = −Q + 1, . . . , Q, we know

yq = argmax{G qa (δ)}.
δ

Denote the uniform upper bound and lower bound of μq−1 − μq among all q ∈ Q as Md
and m d . We have
yq > −Md .
Otherwise, G qa (yq ) < 0 and contradicts with the fact that δ > −m d , G qa (δ) > 0 and
yq = argmax{G qa (δ)}. Hence, we can conclude that
δ

yq ≥ δmin := −M p .

Secondly, for any vector w ∈ [δmin , +∞) M−1 , define y = (y−Q+1 , . . . , y Q ) = T (w).
There exists a uniform δmax ∈ R such that for every q,
yq ≤ δmax . (5.7)
Define δ0 := −m d + 1. By definition of m d , for every q we have
δ0 + μq−1 − μq ≥ 1 > 0.
Hence, for every q ∈ Q, G qa (δ0 ) > 0. Moreover, as f is increasing to its second argument,
for any vector w ∈ [δmin , +∞) M−1 , we have:


Q
G qa (δ0 ) ≥ P ja f (δ0 , δmin ). (5.8)
j=−Q+1
Dynamic Games and Applications (2021) 11:556–579 571

By assumption limδ→+∞ λ(δ)δ = 0, there exists δmax > δ0 such that


Q 
Q
max{ P ja λ(δmax )(δmax + μq−1 − μq )} < P ja f (δ0 , δmin ). (5.9)
q
j=−Q+1 j=−Q+1

As f (δmax , ·) is bounded by λ(δmax ) uniformly, (5.8) and (5.9) imply that for any vector
w ∈ [δmin , +∞) M−1 ,
max G qa (δmax ) < G qa (δ0 ).
q

Since δmax > δ0 and G qa (δmax ) < G qa (δ0 ), we know that the maximum point δ ∗ of G qa cannot
be in the interval (δmax , ∞) as it would otherwise be a contradiction to G qa (δ) being a strictly
increasing function of δ for δ < δ ∗ . Hence for any q ∈ Q,
yq ∈ [δmin , δmax ],

which shows T (K ) ⊂ K , where K = [δmin , δmax ] M−1 .

To prove T is a continuous mapping, we need the following Berge’s maximum theorem.

Theorem 5.5 (Berge) Let X and be metric spaces, f : X × → R be a function jointly

continuous in its two arguments and C : → X be a compact-valued correspondence. For
x in X and θ in , let
f ∗ (θ ) = max{ f (x, θ )|x ∈ C(θ )},
and
x ∗ (θ ) = arg max{ f (x, θ )|x ∈ C(θ )} = {x ∈ C(θ ) | f (x, θ ) = f ∗ (θ )}.
If C is continuous at some θ , then f ∗ is continuous at θ and x ∗ is nonempty, compact-valued
and upper hemicontinuous at θ , that is, if θn → θ and bn → b as n → ∞ with bn ∈ x ∗ (θn ),
then b ∈ x ∗ (θ ).

The next lemma shows that any single-valued, bounded, upper hemicontinuous mapping
is a continuous function.

Lemma 5.6 Let A, B be two Euclidean spaces, : A → B be a single-valued, bounded and

upper hemicontinuous mapping, then is a continuous function.

Proof For any sequence an → a and bn = (an ) ( is a single-valued mapping), if bn tends

to a limit b, then we must have b = (a) by the hemicontinuity of and we are done.
Assume the sequence bn did not have a limit. Since bn is a bounded sequence, there exist
at least two subsequences bn k and bn k that converge to two different values b and b . Since
an → a, we must have both an k and an k tend to a, the hemicontinuity of would imply
b = (a) and b = (a), a contradiction to the assumption that b = b . Therefore, is
continuous.

We now can prove the mapping T defined in (5.5) is continuous on K .

Lemma 5.7 (Continuous Mapping T in R M ) Given any vector μ = (μ−Q . . . , μ Q ) ∈ R M

and bounded set K defined in Lemma 5.4, mapping T defined in (5.5) is continuous on K .
572 Dynamic Games and Applications (2021) 11:556–579

Proof We prove that given vector μ, each element Tq (w) of mapping T is continuous respect
to each wq . As the maximum point of ηa (μ, ·, w, q) exists and is unique for every q ∈
{−Q + 1, . . . , Q}, Tq is a well-defined single-value mapping. Moreover, ηa (μ, δ, w, q) is
jointly continuous w.r.t. δ and w. By Berge’s maximum theorem, Tq is upper hemicontinuous
function of w on bounded set K . Therefore, by Lemma 5.6, for q ∈ {−Q + 1, . . . , Q}, Tq
is also continuous w.r.t. every wq . We conclude that given vector μ, the mapping T is a
continuous mapping from K → K .
Finally, we can prove Theorem 5.2, which concludes the proof of step 1.
Proof of Theorem 5.2 As the intensity function f satisfies Assumption 2.1, from Lemma 5.1,
the maximum point of G qa (δ) exists and is unique for every q. Fixed vector μ ∈ R M , define
mapping T : R M−1 → R M−1 as in (5.5). wa is the fixed point of mapping T . To show the
existence of fixed point to the mapping, Schauder fixed-point theorem is applied to T by the
following steps.
Firstly, by Lemma 5.4, there exists a bounded closed set K ⊂ R M−1 which is equivalently
a compact set, such that T (K ) ⊂ K . From the proof of Lemma 5.4, the compact set K is
convex.
Secondly, from Lemma 5.1 and 5.7, T is a single value continuous mapping from K to
K . By Theorem 5.3, T has a fixed point for every given μ, denoted by wa , and
wqa = T (wa ) ∈ K . (5.10)
This concludes the proof of Step 1.

5.2.2 Proof of Step 2

We first state a global implicit function theorem in [9, Theorem 4], which is used in the proof.
Theorem 5.8 Assume F : Rn × Rm → Rn is a locally Lipschitz mapping such that
• For every y ∈ Rm , the function φ y : Rn → R, defined by φ y (x) = 2 ||F(x,
1
y)||2 , is
coercive, i.e., lim||x||→∞ φ y (x) = +∞.
• The set ∂x F(x, y) is of maximal rank for all (x, y) ∈ Rn × Rm .
Then, there exists a unique locally Lipschitz function f : Rm → Rn such that equations
F(x, y) = 0 and x = f (y) are equivalent in the set Rn × Rm .
With the help of Theorem 5.8, we can show the local Lipschitz continuity of functions wa
and w b .
Theorem 5.9 Assume the intensity function f satisfies Assumption 2.1. Then, there are single-
valued and locally Lipschitz continuous functions wa , w b : R M → R M−1 , such that they
satisfy the generalized Issac’s condition (3.7) in Definition 3.2 for any given vector μ ∈ R M .
Proof We provide the proof for wa only. The proof for w b is similar.
To begin with, from Assumption 2.1, we have (5.3) for all δ, x and y. From Lemma 5.1, the
maximum point of G qa (δ) = ηa (μ, δ, wa , q) is unique. From Remark 5.1, given any vector
μ, wa that satisfies the generalized Issac’s condition in Definition 3.2 is also the solution to
the following first-order condition for every q,

Q
P ja [ f (wqa , waj ) + f 1 (wqa , waj )(wqa + μq−1 − μq )] = 0.
j=−Q+1
Dynamic Games and Applications (2021) 11:556–579 573

For any vector μ and δ = (δ−Q+1 , . . . , δ Q ), define function Fq : R M−1 × R M → R for

every q ∈ {−Q + 1, · · · Q} as following:
Q
j=−Q+1 P j f (δq , δ j )
a
Fq (δ, μ) := −  Q − δq − (μq−1 − μq ).
a 
j=−Q+1 P j f 1 (δq , δ j )

Define mapping F : R M−1 × R M → R M−1 as

F(δ, μ) := (F−Q+1 (δ, μ), . . . , FQ (δ, μ)).
F is continuously differentiable and wa is determined implicitly by F(wa , μ) = 0. From
the proof of step 1, there exists a function wa : R M → R M−1 such that F(wa (μ), μ) = 0
for any vector μ. If we can verify Theorem 5.8 holds in this case, the function wa satisfying
F(wa (μ), μ) = 0 must be unique and continuously differentiable, which concludes our
proof. Hence, the next step is to verify Theorem 5.8.
Firstly, we prove that the Jacobian matrix of F never vanishes. Denote Jacobian matrix
∂ Fi
of F with respect to δ as ∂δ F, a 2Q × 2Q matrix, and its component at (i, m) is ∂δ m
(δ, μ)
for i, m = −Q + 1, . . . , Q. Denote by, for i ∈ {−Q + 1, . . . , Q},
⎛ ⎞2
Q
Di := ⎝ Pma f 1 (δq , δm )⎠ > 0
m=−Q+1

1 
Q 
Q

Ai = Pma P ja [ f 11 (δi , δm ) f (δi , δ j ) − f 1 (δi , δm ) f 1 (δi , δ j )]
Di
m=−Q+1 j=−Q+1

1 a 
Q

Iim := P P ja [ f (δi , δ j ) f 12 (δi , δm ) − f 1 (δi , δ j ) f 2 (δi , δm )].
Di m
j=−Q+1

For m = i, we have:
∂ Fi
(δ, μ) = −1 + Ai + Iii .
∂δi
From Assumption 2.1, we have (5.3), and simple calculation shows:

1 
Q 
Q

− 1 + Ai = Pma P ja [ f 11 (δi , δm ) f (δi , δ j ) − 2 f 1 (δi , δm ) f 1 (δi , δ j )] < 0.
Di
m=−Q+1 j=−Q+1

Hence,
 
 ∂ Fi   
   
 ∂δ (δ, μ) ≥ 1 − Ai − Ii,i .
i

For i = m, the nondiagonal element of the Jacobian matrix ∂δ F is given by:

∂ Fi
(δ, μ) = Iim .
∂δm
To compare the diagonal element with the sum of nondiagonal elements, we have:
   
 ∂ Fi    ∂ Fi  
Q
 (δ, μ)−  (δ, μ) ≥ 1 − Ai − |Iim |. (5.11)
 ∂δ   ∂δ 
i m =i m m=−Q+1
574 Dynamic Games and Applications (2021) 11:556–579

From the definition of Ai and Iim ,


Q
1 − Ai − |Iim |
m=−Q+1

1 
Q 
Q
= Pma { P ja [2 f 1 (δi , δm ) f 1 (δi , δ j ) − f 11

(δi , δm ) f (δi , δ j )] (5.12)
Di
m=−Q+1 j=−Q


Q

−| P ja [ f (δi , δ j ) f 12 (δi , δm ) − f 1 (δi , δ j ) f 2 (δi , δm )]|}.
j=−Q+1

By the assumption of f in (2.1), we have


Q
P ja [2 f 1 (δi , δm ) f 1 (δi , δ j ) − f 11

(δi , δm ) f (δi , δ j )]
j=−Q+1
⎡ ⎤ (5.13)

Q
 
±⎣ P ja − f 2 (δi , δm ) f 1 (δi , δ j ) + f 12

(δi , δm ) f (δi , δ j ) ⎦ > 0.
j=−Q+1

Therefore, as Di > 0, from (5.11), (5.12) and (5.13), we conclude that

∂ Fi  ∂ Fi
| (δ, μ)| − | (δ, μ)| > 0.
∂δi ∂δm
m =i

The Jacobian matrix ∂δ F(δ, μ) is strictly diagonally dominant and is therefore a nonsingular
matrix.
Secondly, we show that given any fixed vector μ, whenever ||δ|| → ∞, ||F(δ, μ)|| → ∞.
For any vector sequence k , k = 1, 2 · · · , ||k || → ∞. Then, there exists sequence n k ∈
{−Q + 1, . . . , Q}, k = 1, 2 · · · , such that |δnk k | → ∞. δnk k is the n k th element of vector k . In
the case that δnk k → −∞, as we have
Q
m=−Q+1 Pma f (δnk k , δm
k)
L n k ( ) :=  Q
k
< 0.
m=−Q+1 Pma f 1 (δnk k , δm
k)

Hence, we know following when k → +∞:

Fn k (k , μ) = −L n k (k ) − δnk k − (μn k −1 − μn k ) > −δnk k − (μn k −1 − μn k ) → +∞.

It means when δnk k → −∞, ||F(k , μ)|| → ∞.

On the other hand, in the case that δnk k → +∞, we can always assume δnk k =
max{δik }i∈Q,i>−Q . As f 1 < 0, f > 0 and f is increasing function to its second variable, we
have the following estimation on Fn k (k , μ):
Q
m=−Q+1 Pma f (δnk k , δm
k)
Fn k ( , μ) = −  Q
k
− δnk k − (μn k −1 − μn k )
m=−Q+1 Pma f 1 (δnk k , δm
k)
Q
m=−Q+1 Pm f (δn k , δn k )
a k k
≤− − δnk k − (μn k −1 − μn k ).
Pnak f 1 (δnk k , δnk k )
Dynamic Games and Applications (2021) 11:556–579 575

f 1 (δ,δ)
From the assumption that limδ→+∞ − f (δ,δ) > 0, we have:
Q
m=−Q+1 Pm f (δn k , δn k )
a k k
0<− lim  < +∞.
δnk →+∞
k
Pn k f 1 (δn k , δn k )
a k k

Then, by taking δnk k → +∞, we finally have:

lim Fn k (k , μ) = −∞.
δnk →+∞
k

Hence, when fixed μ and → +∞, we also get ||F(k , μ)|| → ∞. Moreover, if
δnk k
is consisted of two subsequences such that one converges to +∞, another to −∞, by
δnk k
combining above, we can still get ||F(k , μ)|| → ∞. We conclude that whenever ||δ|| → ∞,
||F(δ, μ)|| → ∞.
Theorem 5.8 implies that there exists a function wa : R M → R M−1 such that
F(wa (μ), μ) = 0 and wa is unique and locally Lipschitz continuous, which concludes
the proof of Step 2.

5.2.3 Proof of Step 3

We next prove there exists a unique classical solution θ to ODE system (3.8) on [0, T ]. The
proof is divided by two parts. Firstly, we show the solution to ODE system (3.8) is bounded if
it exists. Secondly, we provide the proof for existence and uniqueness of the classical solution
to ODE system (3.8).
Lemma 5.10 Assume the intensity function f satisfies Assumption 2.1. If θ : [0, T ] → R M
is a solution to the ODE system (3.8), then for all q ∈ Q we have
1
− γ σ 2 Q 2 T − l(Q) ≤ θq (t) ≤ 2 sup λ(δ)δT .
2 δ

Proof We first prove the upper bound. From the assumption on f and the proof for steps 1
and 2, the ODE system (3.8) is well defined. Since θ is assumed to be a solution, define twice
continuously differentiable functions d 0 and d 1 as

Q−1
d 0 (t, δ) := P jb f (δ, w bj (θ (t))) ≤ λ(δ)
j=−Q


Q
d 1 (t, δ) := P ja f (δ, waj (θ (t))) ≤ λ(δ).
j=−Q+1

From Assumption 2.1, we have (5.3) for all δ, x and y. Simple calculation shows that d 0 and
d 1 satisfy
∂ 2d ζ ∂d ζ
d ζ (t, δ) ≤ λ(δ), (t, δ)d ζ (t, δ) < 2( (t, δ))2 , ζ = 0, 1.
∂δ 2 ∂δ
On the other hand, θ is also the solution to ODE system for all q ∈ Q:
1 2 2
θq (t) = γ σ q − sup{d 0 (t, δ)(δ + θq+1 (t) − θq (t))}I b (q)
2 δ (5.14)
− sup{d 1 (t, δ)(δ + θq−1 (t) − θq (t))}I a (q)θq (T ) = −l(|q|).
δ
576 Dynamic Games and Applications (2021) 11:556–579

The comparison principle for ODE system (5.14) can be proved easily with similar argument
in the proof of comparison principle in Guéant [10]. Define operator H ζ : [0, T ] × R → R
for both ζ = 0, 1 as
H ζ (t, μ) := sup{d ζ (t, δ)(δ + μ)}.
δ

Then, from Guéant [10], we know H ζ is an increasing and nonnegative function in μ.

max H ζ (t, 0) ≤ sup{λ(δ)δ}.

t∈[0,T ],ζ =0,1 δ

Define θ̄ : [0, T ] → RM as following:

θ̄q (t) = 2 sup λ(δ)δ(T − t).
δ

Substituting θ̄ into ODE system (5.14), we have

1
− θ̄q (t) + γ σ 2 q 2 − H 0 (t, θ̄q+1 (t) − θ̄q (t))I b (q) − H 1 (t, θ̄q−1 (t) − θ̄q (t))I a (q)
2
 1
1
= (sup λ(δ) − H ζ (t, 0)) + γ σ 2 q 2 ≥ 0
δ 2
ζ =0

θ̄q (T ) = 0 ≥ θq (T ) = −l(|q|).
Then, by the comparison principle from Guéant [10], we know for every q ∈ Q,
θq (t) ≤ θ̄q (t) ≤ 2 sup λ(δ)δT .
δ

We next prove the lower bound. Let θ̃ : [0, T ] → R M satisfy the following ODE system
for all q ∈ Q:
1
θ̃q (t) − γ σ 2 q 2 = 0
2 (5.15)
θ̃q (T ) = −l(|q|).
The closed-form solution is given by
1 2 2
θ̃q (t) = γ σ q (t − T ) − l(|q|).
2
Note we have estimation that for every vector μ ∈ R M and every q ∈ Q,
ηa (μ, wqa (μ), wa (μ), q) ≥ 0, ηb (μ, wqb (μ), w b (μ), q) ≥ 0.
 
Since θ̃q (T ) ≤ θq (T ), θ̃q (t) ≥ θq (t), then it can be proved similarly as the proof of the upper
solution that for every q ∈ Q:
1
θq (t) ≥ θ̃q (t) ≥ − γ σ 2 Q 2 T − l(Q).
2

To prove the existence of a classical solution to the coupled ODE system (3.8), we cite
the Picard–Lindelof theorem in ODE theory that provides the existence and uniqueness of
solution.
Dynamic Games and Applications (2021) 11:556–579 577

Theorem 5.11 (Picard–Lindelof theorem) Consider the initial value problem in R M :

y  (t) = F(t, y(t)), y(t0 ) = y0 ,
where F : R × R M → R M is uniformly Lipschitz continuous in y with Lipschitz constant
L (independent of t) and continuous in t. Then, for some value ε > 0, there exists a unique
solution y(t) to the initial value problem on the interval [t0 − ε, t0 + ε].

The next lemma is a direct conclusion from the proof of Theorem 5.11, see Teschl [18]. It
helps us to extend the local existence and uniqueness of solution to the global existence and
uniqueness.

Lemma 5.12 Let Ca,b = [t0 − a, t0 + a] × Bb (y0 ), where Bb (y0 ) is a closed ball in R M with
center y0 and radius b. Define
M= sup F(t, y).
(t,y)∈Ca,b

Then, the solution to the ODE system (3.8) exists and is unique on interval [t0 − , t0 + ],
if  satisfies following:
 
b 1
 < min , ,a .
M L

Theorem 5.13 Consider the terminal-value ODE problem on [0, T ]:

θ  (t) = F(t, θ (t)), θ (T ) = θ0 , (5.16)

where F : [0, T ] × R M → R M is a jointly locally Lipschitz continuous function. Assume

that there exists a constant K such that if solution θ exists on any subinterval of [0, T ],
θ (t) ∈ [−K , K ] M . Then, there exists a unique solution to (5.16) on [0, T ].
√ √
Proof Define A T ,2√ M K := [0, T ] × [−2 M K , 2 M K ] M . F is a continuous function.
Hence, there exists uniform constant C > 0 such that
C := sup F(t, y).
(t,y)∈A T ,2√ M K
(5.17)

Since F is jointly locally Lipschitz continuous, there exists a series of open set Ai such that
F is Lipschitz continuous in Ai with Lipschitz coefficient L i , and A T ,2√ M K ⊂ ∪i Ai . By
Heine–Borel theorem, there are finite set I of i such that A T ,2√ M K ⊂ ∪i∈I Ai . Defining
L := maxi∈I L i , we know F is Lipschitz continuous on the compact set A T ,2√ M K with
uniform Lipschitz coefficient L.
As terminal value θ0 ∈ [−K , K ] M , we define C 0 √ := [0, T ] × B√ M K (θ0 ). Then,
√ T, MK
√
⊂ A T ,2 M K . For  := min{ , L , T }, the solution
θ to ODE system (5.16)
MK 1
C0 √
T, MK C
exists and is unique on [T − , T ]. If  = T , then we are done, otherwise, update the new
terminal time as T̃ := T − . Since θ (T̃ ) ∈ [−K , K ] M by assumption, we can update
a new terminal value θ0 := θ (T̃ ). Define a new C 1 √ := [0, T̃ ] × B√ M K (θ (T̃ )) ⊂
√ T̃ , M K
A T ,2√ M K . For  := min{ MC , L , T̃ }, solution θ to ODE system (5.16) exists and is unique
K 1

on [T̃ − , T̃ ], and hence exists and is unique also on [T̃ − , T ]. Repeat this process, and
we can reach  = T̃ after finite number of steps, in which case we have proved the existence
and uniqueness of solution θ to ODE system (5.16) on the whole time interval [0, T ].
578 Dynamic Games and Applications (2021) 11:556–579

Combining Lemma 5.10, Theorem 5.9, and Theorem 5.13, we can finally proceed to show
that the ODE system (3.8) has a unique classical solution.

Theorem 5.14 There exists unique classical solution θ to ODE system (3.8) on [0, T ].

Proof According to Lemma 5.10, we know if the solution θ exists on any subinterval of
[0, T ], there exists constant K ≥ 0 such that

− K ≤ θq (t) ≤ K .

Define F : [0, T ] × R M → R M as
1 2 2
Fq (t, θ (t)) := γ σ q − ηa (θ (t), wqa (θ (t)), wa (θ (t)), q)I a (q)
2
− ηb (θ (t), wqb (θ (t)), w b (θ (t)), q)I b (q)
F(t, θ (t)) := (F−Q (t, θ (t)), . . . , FQ (t, θ (t))).

As q is finite, the original ODE system (3.8) can be rewritten in a vector form with F as in
(5.16). Then, F is a jointly locally Lipschitz continuous function, and if solution θ exists on
any subinterval of [0, T ], θ (t) ∈ [−K , K ] M . By Theorem 5.13, the ODE system has unique
solution on [0, T ]. This concludes the proof of step 3.

5.2.4 Completion of Proof of Theorem 3.4

From Steps 1, 2 and 3, we know there exist unique locally Lipschitz continuous functions
wa , w b that satisfy the generalized Issac’s condition in Definition 3.2, the ODE system (3.8)
is well defined and equivalent to the ODE system (3.5). There exists a unique classical
solution to ODE system (3.8). Define the equilibrium value function for G mm by (3.1), and
the equilibrium controls by (3.9). As θ is the classical solution to the ODE system (3.8), it
is a continuous function on [0, T ] and hence bounded. Then, both π a (t) = wa (θ (t)) and
π b (t) = w b (θ (t)) are bounded on [0, T ]. θ , π a (t) and π b (t) satisfy the ODE system (3.5).
Hence, from the verification Theorem 3.3, the equilibrium for game G mm exists. On the
other hand, as the solution to ODE system (3.5) is unique, by Theorem 3.1 we know the
equilibrium point is also unique.

6 Conclusions

In this paper, we have modeled the price competition between market makers, proved the
generalized Issac’s condition, which ensures the existence and uniqueness of Nash equilib-
rium for market making with price competition, and derived the equilibrium strategies and the
equilibrium value function. We have also performed numerical tests to compare our model
with a benchmark model in the existing literature without price competition and found that
the introduction of price competition reduces bid/ask spreads and improves market liquidity.
There remain many open questions, for example, the jump processes N a and N b are no longer
of Poisson type but more general (Hawkes processes, more general Markov jump processes),
the set of inventory position constraints is no longer a finite set but may be infinite (eventually
uncountable if considering a whole interval); we leave these and other open questions to our
future research.
Dynamic Games and Applications (2021) 11:556–579 579

Acknowledgements The authors are very grateful to the anonymous reviewers and the editors for their con-
structive comments and suggestions which have helped to improve the paper of the previous version.

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which
permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give
appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence,
and indicate if changes were made. The images or other third party material in this article are included in the
article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is
not included in the article’s Creative Commons licence and your intended use is not permitted by statutory
regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.
To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

References
1. Abergel F, Huré C, Pham H (2020) Algorithmic trading in a microstructural limit order book model.
Quant Finance 20:1–21
2. Avellaneda M, Stoikov S (2008) High-frequency trading in a limit order book. Quant Finance 8:217–224
3. Bensoussan A, Siu C, Yam P, Yang H (2014) A class of non-zero-sum stochastic differential investment
and reinsurance games. Automatica 50:2025–2037
4. Buckdahn R, Cardaliaguet P, Rainer C (2004) Nash equilibrium payoffs for nonzero-sum stochastic
differential games. SIAM J Control Optim 43:624–642
5. Cartea A, Donnelly R, Jaimungal S (2017) Algorithmic trading with model uncertainty. SIAM J Financ
Math 8:635–671
6. Cartea A, Jaimungal S (2015) Risk metrics and fine tuning of high-frequency trading strategies. Math
Finance 25:576–611
7. Cohen S, Fedyashov V (2017) Nash equilibria for nonzero-sum ergodic stochastic differential games. J
Appl Probab 54:977–994
8. Fodra P, Pham H (2015) High frequency trading and asymptotics for small risk aversion in a markov
renewal model. SIAM J Financ Math 6:656–684
9. Galewski M, Rădulescu M (2018) On a global implicit function theorem for locally lipschitz maps via
non-smooth critical point theory. Quaest Math 41:515–528
10. Guéant O (2017) Optimal market making. Appl Math Finance 24:1–43
11. Guéant O, Lehalle C, Fernandez-Tapia J (2013) Dealing with the inventory risk: a solution to the market
making problem. Math Financ Econ 7:1–31
12. Hamadene S, Lepeltier J, Peng S (1997) BSDEs with continuous coefficients and stochastic differential
games. In: El Karoui N, Mazliak L (eds) Backward stochastic differential equations. Chapman and Hall,
London, pp 115–128
13. Ho T, Stoll H (1981) Optimal dealer pricing under transactions and return uncertainty. J Financ Econ
9:47–73
14. Kyle A (1984) Market structure, information, futures markets, and price formation. In: Storey GG, Schmitz
A, Sarris AH (eds) Int Agric Trade. Taylor & Francis, New York, pp 45–64
15. Kyle A (1985) Continuous auctions and insider trading. Econometrica 53:1315–1335
16. Kyle A (1989) Informed speculation with imperfect competition. Rev Econ Stud 56:317–355
17. Lin Q (2015) Nash equilibrium payoffs for stochastic differential games with jumps and coupled nonlinear
cost functionals. Stoch Process Appl 125:4405–4454
18. Teschl G (2012) Ordinary differential equations and dynamical systems. American Mathematical Society,
New York

Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.