APPLICATIONS OF A NEW SELF-FINANCING EQUATION

RENÉ CARMONA AND KEVIN WEBSTER

arXiv:1905.04137v1 [q-fin.TR] 8 May 2019

Abstract. The goal of this note is to illustrate the impact of a self-financing

condition recently introduced by the authors. We present the analyses of two
specific applications usually considered in more traditional models in financial
mathematics. They include hedging European options with limit orders and
the optimal behavior of market makers.

MSC 2010: 91G99, 91G80, 91G20 JEL: C6

1. Introduction: the self-financing equation

In quantitative finance, the standard self financing equation is a cornerstone of
the theory of frictionless markets. It plays a crucial role in many fundamental re-
sults. Mathematically, speaking it is a simple equation which constrains the wealth
process of an investor to live in a certain sub-space. This sub-space is therefore
viewed as a space of admissible portfolios. New-comers to the mathematical the-
ories of financial markets often gripe with the self-financing condition and how it
relates to the real world. While it can be postulated as a mathematical definition, it
can also be derived from a limiting procedure starting from an accurate description
of the microstructure of trades in the trade clock. This approach is at the core of
our strategy, and to implement it, we shall have to overcome the idiosyncrasies of
the passage from discrete to continuous time.
“The sad fact is that the self-financing condition is considerably
more subtle in continuous time than it is in discrete time.”1
When discussing market models at the macroscopic level, we assume that the
mid-price p and the inventory L are given by Itô processes:
(
dpt = µt dt + σt dWt
(1.1)
dLt = bt dt + lt dWt′
for two Wiener processes W and W ′ with unspecified dependence structure. In the
simplest case, we also consider an adapted process st acting as a proxy for what
remains in the continuous time limit of the bid-ask spread measured in tick size.
The standard self-financing condition of continuous time finance can be stated as
a constraint:
dXt = Lt dpt (1.2)
between the price p of the underlying interest, the inventory L, and the wealth X
of the agent. In most classical financial models, Merton’s portfolio theory is a case

Date: May 13, 2019.

1
J. Michael Steele, Stochastic Calculus and Financial Applications, section 14.5 ’Self-financing
and self-doubt’.
1
2 RENÉ CARMONA AND KEVIN WEBSTER

in point, the price p is exogenously given, the inventory L is the agent’s input, and
her wealth X appears as the output of equation (1.2).
The objective of this paper is twofold. We first formalize the order book in a
mathematical fashion and derive the associated transaction costs and trade equa-
tions. The second goal is to generalize the self-financing portfolio condition (1.2) to
incorporate known pecularities of the high frequency markets including transaction
costs, price impact and price recovery. Among other things, we want this general-
ization to be able to quantify the differences between trading via limit orders and
market orders. Finally, we want to warn the reader that the equations proposed in
this paper are only necessary, and that quantifying limit order fill rates, priorities
and price recovery are beyond the scope of the present paper.

1.1. The order book. We introduce the order book first as a pair of positive
measures (b, a) on the price grid. Under the assumption that the mid-price is well-
defined, an equivalent definition in terms of an order book shape function γ is
introduced. Formally, γ ′′ = a + b where γ ′′ is the second derivative of the function
γ in the sense of Schwartz distributions.
Transaction costs are shown to be given by the Legendre transform c of the
order book shape function γ. This leads to explicit formulas for all mechanical
transactions on the order book, such as the instantaneous price impact, the traded
volume, etc. In particular, the discrete time equation for the wealth associated to
a self-financing portfolio will be shown to be:
∆X = L∆p ± c(∓∆L) + ∆p∆L (1.3)
where ± is + when trading with limit orders and − when trading with market
orders.

1.2. Proposed self-financing equation. In continuous time, the corresponding

self-financing equation takes the form:
Z
dXt = Lt dpt ± c(y)φσt2 (y)dydt + d[L, p]t (1.4)
R

where as before ± is + when trading with limit orders and − when trading with
market orders, and φσ2 is the density function of the Gaussian distribution with
mean 0 and variance σ 2 . We show in Section ?? below that, when time is measured
in the trade clock, the discrete time analog of formula (1.4) can be derived rigorously
from a specific limit order book feature. It also matches real wealth data (see [9]
and the appendix at the end of the paper for empirical evidences). We shall also
impose the constraint
d[L, p] < 0 (1.5)
whenever trading is done with limit orders. The interpretation for this constraint
is price impact. It has also been thoroughly tested on high frequency data in [9].
See also the appendix at the end of the paper for evidence from market data from
a different exchange.
We now explain how our condition (1.4) and the adverse selection constraint (1.5)
relate to the conditions most often found in the literature. Later on in Section
??, the latter will be derived as continuous time limits of discrete self-financing
equations under different scaling assumptions.
 APPLICATIONS OF A NEW SELF-FINANCING EQUATION 3

1.3. The Almgren-Chriss model. The seminal work of Almgren and Chriss [5]
addresses a closely related question. These authors propose a macroscopic model
for the price impact and the change of wealth after a liquidity taker’s decision.
The model leads to a very tractable framework which was, and still is, used in
many optimal execution studies (see [2, 20] for example). This framework can be
summarized by the system:

dpt = f (lt )dt + σt dWt

dLt = lt dt (1.6)

dXt = Lt dpt − c(lt )dt


where f and c are positive functions2.

The main advantage of this model is that price impact appears in a tractable
fashion. Indeed, it comes through the drift f (lt ) of the price process, which creates a
positive correlation between traded volumes and the price. However, it constrains
L to be a differentiable function of time, and as a result, the model parameters
cannot be calibrated to market data directly, making the model difficult to test
empirically. As per the empirical analysis of NASDAQ data reported in [9], there is
ample evidence supporting nondifferentiable inventories (see also the appendix at
the end of the paper). Moreover, certain trading strategies, such as delta-hedging,
latency arbitrage, and statistical arbitrage, naturally lead to inventory models with
infinite variation. Finally, the use of limit orders is not covered by the Almgren-
Chriss approach.

1.4. Transaction cost literature. The branch of classical mathematical finance

most related to our paper is portfolio selection under transaction costs ([12, 19, 21]
or the recent review [18]). Most of these works start from an equation for the wealth
of a liquidity taker which generalizes the classical self-financing equation to a setting
with transaction costs. In general however, these papers do not underscore the
derivation of the model, but instead, they emphasize the study of its consequences.
We hope to appeal to this part of the community by providing more accurate
equations for self-financing portfolios while keeping reasonable tractability, leading
the way to problems related to liquidity provision, such as market making. An
interesting feature of such applications is that the agent does not directly control
her portfolio, adding an additional modeling challenge. For the record we note that
the standard equation used in this branch of the literature is:
st
dXt = Lt dpt − |dL|t (1.7)
2
where st is the bid-ask spread,
Rt and again, the inventory process L is assumed to
have finite variation, i.e. 0 |dL|s < ∞ for all finite t.
Strengths of this model are its simplicity, relative tractability, and straightfor-
ward calibration to the market. However, we see a shortcoming in the fact that the
process L can only have finite variation3. Moreover, price impact, limit orders and
other microstructure considerations are absent from the model.

2c should be understood as a transaction cost function. For this reason, it is often assumed to
be convex.
3See [9] and the appendix for why this is problematic.
4 RENÉ CARMONA AND KEVIN WEBSTER

1.5. Methodology. Rather than postulating the definition of a self-financing port-

folio directly in the continuous limit, our objective is to derive the self-financing
equation from a more fundamental perspective. To obtain our results, we therefore
propose the following novel strategy:

(1) Define a discrete time representation of the market where all the relevant
primary quantities (e.g. price) are rigorously defined on a trade-by-trade
basis. This is what we call the ’microscopic’ scale.
(2) Deduce or define the relevant equations for the derived quantities (e.g.
wealth) and trading constraints (e.g. trading via limit orders) on the same
scale.
(3) Assume that the primary quantities of interest are samples from continuous
time diffusion processes, with a sampling frequency that goes to infinity. As
a result, each microscopic model is embedded in a sequence of microscopic
models that approximate a continuous time model. We call the continuous
time limit the ’macroscopic’ limit of the model.
(4) Using appropriate limit theorems, derive the continuous time analogs of the
derived quantities and trading constraints.

The main quantity of interest to us, the wealth of a self-financing portfolio, does not
require a stochastic model but rather a precise description of the rules underpinning
trades on a limit order book. These rules create structural relationships that we
exploit to express, at the microscopic scale, wealth as a non-linear functional on the
path of our primary quantities. A functional law of large numbers then controls the
limiting argument used to derive our continuous time equivalent of the self-financing
portfolio equation.
The four-step approach described above can be used in a variety of financial
markets. Indeed, it should be possible to derive a self-financing equation for each
form of market microstructure. Controlling a continuous time limit for the input
variables could then lead to a tractable summarizing equation. While it can hold
in other markets, the particularly simple form it takes, and the ease with which
it can be tested on empirical data, make high frequency markets an ideal test bed
for our theory. Our presentation was influenced by the high frequency markets for
three reasons: 1) they are the most relevant among the electronic markets; 2) we
can illustrate and test many specific features on empirical data; 3) we believe that
our price impact constraint is a particularly strong feature of these markets.
Our research also casts new light on the difference between trading via limit
orders and market orders by modeling price impact. This can be done directly
in the continuous time limit, or via the same strategy as for the self-financing
equation. In that case, a model for the price move after each trade is proposed on
the microscopic scale, and a functional central limit theorem is used to derive the
corresponding continuous time equations.
We believe in the potential of our modeling approach, and we hope that other
idiosyncrasies of the financial markets will be incorporated into continuous time
models through this micro-to-macro approach. Notable papers with a similar mi-
croscopic approach are [7, 11, 23] and [15, 16]. The former three papers study
microstructure to derive optimal bid-ask spread policies. The latter two propose
microscopic models of the limit order book. Finally, [14] derives the diffusion limit
of one of those microscopic order book models.
 APPLICATIONS OF A NEW SELF-FINANCING EQUATION 5

1.6. Basic Results of [10]. The first part of the paper formalizes trading on a limit
order book, and provides a microscopic description of trades. A duality relationship
with transaction costs is introduced and the self-financing condition is derived as a
plain accounting relationship for individual trades.
• We start with a model of the limit order book given by two positive (finite)
measures with non-overlapping supports. They represent the distributions
of the limit buy and sell orders.
• From there, we derive the optimal behavior of of a risk neutral liquidity
taker: typically he should place an order at the expected value of the future
values of the price.
• We then argue that in discrete time given by the trade clock, the changes
in the three fundamental quantities should be linked by the fundamental
accounting relationship
∆X = L∆p + c(−∆L) + ∆p∆L (1.8)

which we understand as a single trade self financing equation. Here p rep-

resents a quoted price, typically the mid-price averaging the best bid and
the best ask, L is the inventory (number of shares) and K the amount of
cash held by the trader, andX is the wealth (as marked to the quoted price
P ) defined by X = pL + K. As usual we use the notation ∆ for the change
of a quantity after a trade takes place.
Then, continuous time versions of the self-financing equation are derived as lim-
its of the discrete self-financing equation under different scaling regimes. We review
briefly the various steps of these limiting procedures and we identify one of these
limiting models as being more relevant than those currently used in the literature.
The back-and-forth approach starts from a continuous time model in which both
the quoted price pt and the inventory Lt are Itô processes. See (1.9) below. Since
the assumption that the inventory process has a non-trivial quadratic variation
component is not widely used in the existing financial mathematics literature, [10]
provides empirical evidence in two appendices based on the analyses of the NAS-
DAQ and the TSX high frequency trading data.
Continuous time versions of the single trade self-financing conditions are obtained
in the following way. We start from a filtered probability space (Ω, F , F, P) sup-
porting two F-Wiener processes W and W ′ with unspecified correlation structure
and over a fixed time interval, say [0, 1], we give ourselves two F-adapted processes
for the price and inventory of a liquidity provider:
( Rt Rt
pt = p0 + 0 µu du + 0 σu dWu
Rt Rt (1.9)
Lt = L0 + 0 bu du + 0 lu dWu′

where p0 and L0 are square integrable F0 -measurable random variables, and µ, σ, b

and l are F-adapted and càdlàg processes. We assume that the order book at time t
is given by a convex shape function γt which is a continuous F-adapted process and
satisfies γt (0) = 0, and we denote by ct its Legendre transform giving the transaction
costs associated to the transactions on the order book at time t. The major insight
of [10] was to prove that if at a scale given by an integer N ≥ 1 we assume that
the mid-price evolves according to the discretization pN n = pn/N for n = 0, 1, · · · , N
of the continuous time pt (and similarly for the inventory LN n = Ln/N , then if the
6 RENÉ CARMONA AND KEVIN WEBSTER

shape of the order book is give at the scale N by:

1 √ 
γnN (x) = γn/N Nx . (1.10)
N
then if the single trade self-financing condition holds at scale N , then it is proven
in [10] that the continuous time relationship between the liquidity provider wealth
X, inventory L, the price p and the transaction cost function c is:
dXt = Lt dpt + Φlt (ct )dt + d[L, p]t (1.11)
N
where Xt = limN →∞ X⌊N t⌋ u.c.p. In this formula, Φσ2 is the cumulative distribu-
tion function of the Gaussian distribution with mean 0 and variance σ 2 .
The goal of this note is to present a couple of applications already touted in
the literature, but which get a nw lease on life in light of the new self-financing
condition.
Remark 1.1. Before we proceed, we note that it is possible to recover the classical
Almgren-Chriss price impact model, as well as the standard proportional transaction
cost model, by choosing different renormalizations of the order book shape function
different from (1.10). We refer the interested reader to Section 3 of [10] for a
detailed discussion of this important remark.

2. Price Impact and Models

The above self-financing equations can be considered as bare bones descriptions
of the market. They provide for an accountant’s perspective on the market. Given
a trader’s inventory and the limit order book he or she trades on, the accountant
can track his or her wealth perfectly. The number of hypotheses made are minimal
in order to obtain the result which tracks perfectly wealth once a trading strategy
is given.
However, the above framework does not shed much light on which trading strate-
gies can lead to the inventory processes satisfying the self-financing conditions we
derived. Clearly, if any strategy is permissible in some reasonable sense, trading
with limit orders is always preferable to trading with market orders as it is obviously
preferable to capture the transaction costs rather than pay them. But in practice
there is a trade-off between using limit and market orders. This trade-off has been
captured by key words such as adverse selection, price impact and market response
function. These three terms are related, at least on an informal level, and corre-
spond to attempts by different communities (the economics, mathematical finance
and econophysics communities respectively) to model the hidden cost of placing
limit orders. We propose our own approach within the framework of the previous
two sections. For the sake of definiteness, we use the terminology price impact, in
line with the rest of the mathematical finance community.
We would like to stress that we regard Definition 2.1 as the main thrust of this
section, and our main contribution to the literature on price impact. It is motivated
by the empirical fact that, with very high probability, ∆L∆p ≥ 0 when trading with
market orders and ∆L∆p ≤ 0 when trading with limit orders. In plain words, these
inequalities state that the price never moves immediately against a market order.
It is a very simple, yet robust model of price impact.
Definition 2.1. Let p be a price process, c a transaction cost process and L an
inventory process obtained exclusively via the use of limit orders. We say that the
 APPLICATIONS OF A NEW SELF-FINANCING EQUATION 7

triplet (p, c, L) is consistent if the sample paths of [p, L]t are strictly decreasing a.s.
.
We shall sometimes say that the couple (p, L) is consistent when the transaction
cost process c corresponding to the order book on which the limit orders are placed
is understood from the context.
Next, we illustrate the significance of this definition on a specific model intro-
duced to justify efforts such as [3] and [20] to model price impact as a deterministic
mean-reversion of the limit order book between two trade times. The advantage of
choosing such a model for the purpose of illustration is that we deal with bona fide
equations rather than mere inequalities, leading to stronger results, albeit under
stronger assumptions.
2.1. More explicit and rigid price impact model. The aim of this section is
to show that a model for limit order fill rates with exact price recovery can lead to
models where the price is a function of trade volumes, and vice-versa. This provides
a model of supply and demand in high frequency markets, and closes the loop of our
excursion in modeling. While this structural model is more rigid than our previous
reduced form models, we believe that it illustrates some important market features,
and provides yet another example of our micro-to-macro transition. By exact price
recovery, we mean that in this model, one can compute the exact price move taking
place when a limit order depletes liquidity on the limit order book, and otherwise,
assume that the price recovers to a deterministic value between the previous price
and this new price resulting from the move.
2.1.1. Microscopic assumptions. Let (Ω, F , P) be a probability space and p and L
be two discrete time processes representing the market price and the inventory
of a liquidity provider respectively. Let γ be a C 3 -function valued discrete time
process representing our provider’s shape function and c its associated transaction
cost process. Our fundamental assumption is that
∆p = λc′ (−∆L) (2.1)
or, equivalently
∆L = −γ ′ (λ−1 ∆p) (2.2)
where λ ∈ (0, 1] is a real number that encapsulates price recovery. The bigger λ,
the smaller the price recovery.
2.1.2. Tools. Equation (2.1) allows a liquidity provider to derive the price from
trade volumes and the order book, while equation (2.2) derives the trade volumes
from the prices and the order book. Both lead to the same consistency relationships
between p, L and γ in the continuous limit.
Our analysis is based on a result from [17] which we state for the sake of com-
pleteness.
Let (Ω, F , F, P) be a filtered probability space supporting an 1-dimensional F-
Wiener process W , and Y a 1-dimensional Itô process of the form:
Z t Z t
Yt = Y0 + bt dt + σt dWt , t ∈ [0, 1]. (2.3)
0 0

Assumption 2.2. ((H)+ (K) from [17]) We assume that bt and σt are progressively
measurable, bt is locally bounded and σt is càdlàg.
8 RENÉ CARMONA AND KEVIN WEBSTER

Now let F : Ω × [0, 1] × R → R be a random, F-adapted function that is C 1 in y

and C 0 in (t,y). We will shorten the notation to y 7→ Ft (y).

Assumption 2.3. ((7.2.1), (10.3.2), (10.3.3), (10.3.4) and (10.3.7) from [17])
We assume that a.s. for all t, Ft is an odd function. Furthermore, we assume
that there exists a function g : R → R with at most polynomial growth, and a real
number β > 1/2 such that, for all ω ∈ Ω, (t, s) ∈ [0, 1]2 and y ∈ R we have:
|Ft (y)| ≤ g(y)
|Ft′ (y)| ≤ g(y)
|Ft (y) − Fs (y)| ≤ g(y)|t − s|β

We will make use of the following result (10.3.2) from [17]:

Theorem 2.4. Under assumptions 2.2 and 2.3, there exists a very good filtered
extension of the original space such that we have the following stable convergence
in law as N → ∞:
⌊N t⌋ √
1 X 
√ Fn/N N (Y(n+1)/N − Yn/N ) → Ut
N n=1

where
Z t Z tp
Ut = bs Φσs (Fs′ ) ds + Φσs ((Fs )2 )dWs′ (2.4)
0 0

where Wt′ is a d-dimensional Wiener process such that

Z t
Φσs id Fsk


′
[W , W ]t = p ds
0 σs Φσs (Fsk )2
and id is the identity function.

2.1.3. Continuous time setup. Let (Ω, F , F, P) be a filtered probability space sup-
porting a F-Wiener process W . We will fix either an Itô process
Z t Z t
pt = p0 + µs ds + σs dWs (2.5)
0 0

for the price or

Z t Z t
Lt = L0 + bs ds + ls dWs (2.6)
0 0
for the inventory. In addition to one of these processes, we also fix an order book
shape process γ = (γt )t≥0 and denote by c = (ct )t≥0 the associated transaction cost
process.
Let us assume that L (respectively p) satisfies Assumption 2.2 and c′ (respectively
′
γ ) satisfies Assumption 2.3. These assumptions basically state that the stochastic
transaction cost function ct (·) is symmetric, twice differentiable in space and has
strong smoothness properties in time.
As before, we define the√discretized
 processes LNn = Ln/N (respectively
√  pN
n =
N 1 N 1
pn/N ) and cn (·) = N cn/N N · (respectively γn (·) = N γn/N N · ).
 APPLICATIONS OF A NEW SELF-FINANCING EQUATION 9

2.1.4. Main result. The main result is a straightforward application of Theorem

2.4. If we are given the inventory L and transaction costs c then we have:
Corollary 2.5. There exists a very good filtered extension of the original space
such that we have the stable convergence in law pN ⌊N t⌋ → pt with
p
dpt = −λbt Φlt (c′′t ) dt + λ Φlt ((c′t )2 )dWt′ (2.7)
where
t
Φ (id c′s )
Z
′
[W , W ]t = − pls ds. (2.8)
0 ls Φls ((c′s )2 )
In particular,
d[p, L]t = −Φlt (id c′t ) dt (2.9)
A completely analogous result is obtained if we start with the price p and an
order book shape function γ:
Corollary 2.6. There exists a very good filtered extension of the original space
such that we have the stable convergence in law LN
⌊N t⌋ → Lt with

dLt = −µt Φσt γt′′ (λ−1 ·) dt + Φσt ((γt′ )2 (λ−1 ·))dWt′

p
(2.10)
where
d[p, L]t = −Φσs id γt′ (λ−1 ·) dt.


(2.11)
2.2. A special case. While quite unrealistic, the model of a flat order book has
been used repeatedly because of its extreme tractability. See for example [2, 20].
It corresponds to a shape function satisfying γt′′ = mt for some real valued adapted
process m = (mt )t≥0 taking only positive values. The corresponding cost function
is hence quadratic, and this quadratic transaction cost model leads to the linear
price impact model: (
dpt = − mλt dLt

l2 (2.12)
dXt = Lt dpt + 12 − λ mtt dt.
Note that the sign of the effective transaction costs is that of 21 − λ. Indeed, in
the self-financing case λ = 12 , and price recovery and price impact perfectly cancel
each other. If λ > 21 , then the price impact of trades is stronger than the collected
spread because of insufficient price recovery. Also, because of the uniform structure
of the order book and perfect fill rate, the inventory of the provider is perfectly
anti-correlated with the price.

Absence of price manipulation strategies. A major concern for any dynamic model
of market microstructure is the possible existence of price manipulation strategies.
These can be defined in multiple ways and have been studied extensively by Schied
et al in [3] among others. In the present context, we define price manipulation in
the following way:
Definition 2.7 (Price manipulation). Let c be a transaction cost function, λ a price
recovery parameter, and A a set of couples of processes (p, L) which are consistent
(with respect to c) in the sense of Definition 2.1. We say that this set A is subject
to price manipulation if there exists a (p, L) ∈ A such that L1 = L0 and
E[X1 ] > E[X0 ]. (2.13)
10 RENÉ CARMONA AND KEVIN WEBSTER

In words, we want to exclude round-trip statistical arbitrages, given that the

liquidity provider does not control the incoming market orders. The tractability
of the flat order book allows us to rule out price manipulation strategies in many
situations.

Proposition 2.8. Consider a market with price recovery parameter λ. A constant

flat order book, γt′′ ≡ c for some c > 0 does not allow for price manipulation if
λ ≥ 1. On the other end, price manipulation strategies do exist when λ < 1.

Proof. Using both equations of (2.12) we get:

λ 1−λ 2
dXt = − d(L2 )t + l dt
2mt 2mt t
so that integrating between t = 0 and t = 1 we get, if L1 = L0 :

1 − λ 1 ls2
Z
X1 = X0 + ds (2.14)
2 0 ms

from which one easily concludes. 

3. Applications
Applications of the proposed relationships depend on models of the inventory
and price processes L and p. In the sequel, when we formulate an optimization
problem, we assume that the inventory can be any Itô process. This is an act of
faith as making it happen typically requires good execution algorithms and limit
order fill rates. In any case, to be consistent with the results of [10] recalled above,
we require that these processes satisfy d [p, L]t < 0 when using limit orders and
d [p, L]t ≥ 0 for market orders.

3.1. Option Hedging. In this section we derive a pricing PDE for a European
option in a local volatility model with transaction costs and price impact as given
by the adverse selection term in our self-financing condition. We highlight the
consequences of trading with market orders only as opposed to trading with limit
orders. This is directly related to the literature on option hedging under gamma
constraint. We recover the same PDE structure and interpretation of the gamma
penalization as a liquidity cost, as in [22] and [13], albeit through a different path.
These papers do not emphasize adverse selection and no discussion of the order
book are given. The first paper starts from the gamma-penalized Partial Differential
Equation (PDE) while the other derives it as a limit of PDE models. Neither paper
derives the non-linear penalty in the PDE from a microstructure model. They are
more focused on what the penalty term implies.
In addition to presenting a microstructure-based approach to the problem, our
solution also answers a very practical-minded question: Should one delta-hedge with
limit orders or market orders? We argue that the answer depends upon the sign
of the gamma of the option.
In what follows, we first treat the fully non-linear case to exhibit the strong
parallel with [13] and [22]. The linear case follows as a corollary. Clearly, it is of
more practical interest as a tractable extension of the Black and Scholes framework.
 APPLICATIONS OF A NEW SELF-FINANCING EQUATION 11

3.1.1. Mathematical setup. Let (Ω, F , F, P) be a filtered probability space and W

be a F-Brownian motion which generates the filtration F. We assume that the
mid-price pt is given exogenously and satisfies the stochastic differential equation
(SDE):
dpt = µ(pt )dt + σ(pt )dWt , (3.1)
where µ and σ are two globally Lipschitz functions. We also assume that the order
book shape function process γ = (γt )t≥0 is continuous and random only through
the price level p in the sense that γt (α) = γ(pt , α) for a deterministic function
(p, α) ֒→ γ(p, α). We single out a trader, and define an inventory process as an
F-adapted Itô process
Z t Z t
Lt = L0 + bu du + lu dWu (3.2)
0 0

where (bt )t≥0 and (lt )t≥0 are F-adapted, càdlàg processes. Note that in this formu-
lation, lt is signed. Since we identified trading with limit or market orders to the
sign of d[p, L]t , we shall impose the constraint lt < 0 when we model trading with
limit orders, and lt ≥ 0 when trading is done with market orders.
Denote by c(p, ·) the Legendre transform of γ(p, ·). We will use the function
g(p, l) = sign(l)Φl (c(p, ·)) (3.3)
Given a real number K0 representing the trader’s initial cash endowment, according
to our self-financing condition, her wealth is given by:
Z t Z t
Xt = L 0 p0 + K 0 + Lu dpu + (σ(pu )lu − g(pu , lu )) du. (3.4)
0 0

This follows from the self-financing equation proved in Proposition ??.

Let f ∈ C 0 be the payoff function of a European option with maturity T . We
define a perfect replication strategy as follows.
Definition 3.1. An initial cash endowment K0 and an inventory process L =
(Lt )t≥0 are said to perfectly replicate the European payoff f (pT ) at maturity T if
XT = f (pT ) (3.5)
and the corresponding replication price is defined as X0 = K0 + p0 L0 .
3.1.2. The result.
Theorem 3.2. Let f ∈ C 0 and T > 0. Assume that v ∈ C 1,3 solves the PDE:
∂2v σ 2 (p) ∂ 2 v
 
∂v
(t, p) + g p, σ(p) 2 (t, p) − (t, p) = 0 (3.6)
∂t ∂p 2 ∂p2
with terminal condition v(T, p) = f (p). Then Lt = ∂v ∂p (t, pt ), and K0 = v(0, p0 ) −
∂v
∂p (0, p0 )p0 form a perfect replication strategy for the payoff f (pT ) at maturity T .
Furthermore, the volatility of the replicating inventory is given by
∂2v
lt = σ(pt ) (t, pt ), (3.7)
∂p2
and the replication price of the option is X0 = v(0, p0 ).
12 RENÉ CARMONA AND KEVIN WEBSTER

∂v
Proof. Choosing Lt = ∂p (t, pt ) leads to
∂2v
lt = σ(pt ) (t, pt ) (3.8)
∂p2
and
∂2v ∂2v σ 2 (pt ) ∂ 3 v
bt = (t, pt ) + µ(pt ) 2 (t, pt ) + (t, pt ). (3.9)
∂t∂p ∂p 2 ∂p3
As v ∈ C 1,3 , this choice of Lt is therefore a bona fide inventory process. Writing
down the dynamics of the wealth leads to:
dXt = (σ(pt )lt − g(pt , lt )) dt + Lt dpt
∂2v ∂2v
 
∂v
= −g(pt , σ(pt ) 2 (t, pt )) + σ 2 (pt ) 2 (t, pt ) dt + (t, pt )dpt
∂p ∂p ∂p
∂2v
 
1 2 ∂v ∂v
= σ (pt ) 2 (t, pt ) + (t, pt ) dt + (t, pt )dpt
2 ∂p ∂t ∂p
= d(v(t, pt )).
Since the initial values match, we have that Xt = v(t, pt ) for all times. This
concludes. 
∂2 v
We recall that an option is said to have positive gamma when ∂p2 (t, p) > 0 for
∂2 v
all t and p. A negative gamma option is one for which ∂p2 (t, p) < 0 for all t and p.
Corollary 3.3. When using only one type of order (say limit or market orders),
positive gamma options can only be hedged (in the sense that their payoffs can be
replicated) with market orders. Negative gamma options can only be hedged with
limit orders.
Proof. The identity
∂2v
(t, pt )
lt = σ(pt ) (3.10)
∂p2
implies that lt and the option gamma must be of the same sign. 
Remark 3.4. We illustrate the above result by commenting on a simple practical
example. If one buys and delta-hedges a call option, then a synthetic negative
gamma position must be created through a dynamic trading strategy. Because the
delta of a call option increases as the price increases, the trader must sell when the
price goes up, and buy when the price goes down. Given that limit orders tend to
buy when the price goes up and sell when the price down because of price impact,
their use to lower the cost of delta-hedging makes sense.
Remark 3.5. The above result leads to an intuitive implementation strategy for
delta-hedging negative gamma positions. By computing the gamma of an option,
one can figure out how much needs to be bought should the price move down, or
sold should the price move up. By placing limit orders at the end of the queue, it
is impossible for the price to move without our hedge getting executed. Since the
model has only one source of random shocks (the Wiener process W = (Wt )t≥0 ),
it is expected to be complete in the sense that perfect replication holds. In practical
situations, the incompleteness of the markets and the fact that the price could move
back introduce some hedging risk. Nevertheless, the negative correlation created by
trading with limit orders lowers the cost of the delta-hedge.
 APPLICATIONS OF A NEW SELF-FINANCING EQUATION 13

The PDE (3.6) is non-linear in the second derivative of v. However, it is linear

when all the orders can be filled at the pre-trade best-bid or best-ask prices, in
which case we recover a Black-Scholes type formula.
Corollary 3.6. Assume that all the orders can be filled at the pre-trade best-bid or
best-ask prices and denote by st = s(pt ) the bid-ask spread. Then the pricing PDE
(3.6) becomes
σ(p)s(p) σ 2 (p) ∂ 2 v
 
∂v
(t, p) + √ − (t, p) = 0. (3.11)
∂t 2π 2 ∂p2
s(p) s(p)
Proof. In the bid-ask spread case, c(p, l) = 2 |l| and hence g(p, l) = √ l.
2π


In particular,
√ we recover the standard, frictionless local volatility model when
s(p) = 2πσ(p). This is because in our self-financing condition, the terms repre-
senting transaction costs and adverse selection (i.e. price impact) exactly cancel
each other out, and the frictionless self-financing equation holds true.
In addition to computing the price and delta-hedging ratios under transaction
costs and instantaneous adverse selection, this theory suggests an execution strategy
by specifying when limit or market orders should be used to hedge an option.

3.2. Market Making. For our second application, we adapt to our framework
the key insight of the model proposed in [7]. The aim is to solve the optimization
problem of a representative market maker controlling the spread and maximizing
her profits. The trade-off she faces, and which is the key ingredient of the model,
is the following: the smaller the spread, the likelier trades are, but the less profit
she makes on each of them.
In a way similar to [7, 23], we model the probability of execution of a limit
order by a decreasing function of the quoted spread. This will first be done at the
microscopic level, to obtain a reasonable model for our inventory process L at the
macroscopic level. A key difference with [7] is that we still impose the price impact
constraint, which further depresses the market maker’s profits because of adverse
selection. In this respect, our model is closer to that proposed in [11], which also
proposes a market making control problem subject to adverse selection.
To make sure that the price impact constraint is satisfied, we use, at the mi-
croscopic level, a modified version of the Almgren and Chriss model [5] to relate
the change in mid-price to the change in the aggregate inventory of the liquidity
providers. We assume that
∆n L = −λn+1 ∆n p (3.12)
for a Fn+1 -measurable, positive random variable λn+1 . This is an non-predictable
form of linear price impact, in the sense that, ex-post, the price increment is a
linear function of the traded volume.
To capture the insight of [7], we model λn+1 in such a way that
2
E[ λn+1 | Fn ] = ρn (sn )fn (sn ) and E[ λ2n+1  Fn ] = (fn (sn ))

(3.13)
where sn is the market maker’s chosen spread, and ρn and fn are continuous,
positive function with fn decreasing and ρn ∈ [0, 1]. The assumption that fn is
decreasing in the spread is inherited from [7], and the fact that ρn must be smaller
than 1 is due to Jensen’s convexity inequality. We assume that λn+1 is independent
14 RENÉ CARMONA AND KEVIN WEBSTER

of ∆n p conditionally on Fn . Computing the predictable quadratic variation of Ln

yields:
n−1
X
fk2 (sk )E ∆k p2  Fk ,
  
(3.14)
k=1
while the predictable quadratic covariation of Ln and pn is given by:
n−1
X
ρk (sk )fk (sk )E ∆k p2  Fk .
  
− (3.15)
k=1

This suggests the use of the following model in the continuum limit:
(
dpt = µt dt + σt dWt
(3.16)
dLt = −ρt (st )ft (st )µt dt + ft (st )σt dWt′
Rt
with d[W, W ′ ]t = − 0 ρu (su )du for some adapted, continuous and positive func-
tions ρt (·) and ft (·) with ρt ≤ 1 and ft decreasing. Note that the equation for Lt
can also be written as:
q
dLt = −ρt (st )ft (st )dpt + ft (st ) 1 − ρ2t (st )σt dWt⊥ (3.17)

with a Wiener process Wt⊥ independent from Wt . We will from now on assume
that pt is adapted to the filtration generated by Wt .
Applying our wealth equation, we obtain:
Z T Z T
1
XT = L T pT − pt dLt + √ σt st ft (st )dt. (3.18)
0 2π 0
For both ft and ρt , a natural assumption is that they are time-independent functions
of the spread rescaled by the volatility:
ft (s) = f (s/σt ); ρt (st ) = ρ(st /σt ) (3.19)
0 0
for some C decreasing function f and C function ρ. We will furthermore assume
that g(x) = xf (x) is a decreasing function for x large enough, that g(x) → 0 as
x → ∞, and that f (x) > 0 for all x ≥ 0.
The problem of a risk-neutral market maker attempting to set the spread opti-
mally is to maximize:
sup EXT . (3.20)
s
This is a classical stochastic control problem which we solve using the Pontryagin
maximum principle. Let us define a few functions first.
Lemma 3.7. For all a > 0, define the function Fa by
x
Fa : x 7→ √ f (x) − aρ(x)f (x) (3.21)
2π
Then the function
M (a) = max Fa (x) (3.22)
x∈[0,∞)
is well defined, continuous, and decreasing in a. Furthermore, there exist a mea-
surable selection
m(a) ∈ argmaxx∈[0,∞) Fa (x) (3.23)
and we have that m(a) > 0.
 APPLICATIONS OF A NEW SELF-FINANCING EQUATION 15

Proof. First, note that for all a > 0,

Fa (0) = −aρ(0)f (0) ≤ 0, Fa (a + 1) ≥ f (a + 1) > 0
Next, if g is decreasing on the interval [x0 , ∞), then we can define the function β(a)
as g −1 ◦ f (a + 1) if f (a + 1) is in g[x0 , ∞), and x0 otherwise. β(a) is continuous
and satisfies f (a + 1) ≥ g(x) for all x ∈ (β(a), ∞).
This proves that the maximum of Fa is attained on the compact interval [a +
1, β(a)]. The continuity of M holds by Berge’s maximum theorem. It is decreasing
by definition of Fa . The measurable selection result follows by Theorem 18.19 of
[4]. 
Proposition 3.8. Any solution of the control problem is of the form
st
= m (αt ) (3.24)
σt
where
µt Zt
αt = E [ pT − pt | Ft ] 2 + , (3.25)
σt σt
Zt being the volatility of the martingale representation of pT
Proof. We apply the necessary part of the stochastic Pontryagin maximum princi-
ple. The generalized Hamiltonian is equal to:
Ht (s, L, Y, Z, Z ⊥) = −ρ(s/σt )f (s/σt ) [(Yt − pt ) µt + σt Z]
σt p
+ √ sf (s/σt ) + σt f (s/σt ) 1 − ρ2 (s/σt )Z ⊥
2π
and the adjoint equation is solved by
Yt = E [ pT | Ft ] (3.26)
which, in particular, implies Zt⊥
= 0. Zt can be computed via the martingale
representation theorem on the Brownian filtration generated by Wt .
The Hamiltonian to maximize therefore becomes
 
s
σt2 Fαt (3.27)
σt
and the previous lemma concludes. 
Beyond the optimal control, one might be interested in the dependence in σt
and αt of the market maker’s expected profit as well as the volatility of her inven-
tory. Note that a low volatility of the inventory means that the market maker has
essentially pulled out of the market.
Corollary 3.9. The market maker’s expected profit is
"Z #
T
E M (αt ) σt2 dt (3.28)
0

where αt is given by (3.25) and the function M is defined by (3.22). The volatility
of her inventory is
σt f (m (αt )). (3.29)
Proof. The expected profit can be computed by integrating the Hamiltonian along
the optimal path. The rest follows from the previous proposition. 
16 RENÉ CARMONA AND KEVIN WEBSTER

Recall that M is a decreasing function. A consequence of the corollary is that

the market maker’s expected profit is a decreasing function of αt and, for αt being
fixed, an increasing function of the volatility.
Two issues need to be resolved if one looks for tractable formulas. First, an
explicit model for pT must be given for which the martingale representation term
Zt can be computed. Second, one has to propose a function g for which the maximal
argument m of F can easily be characterized as a function of αt . We address them
in some specific situations.
3.2.1. The martingale case. Note that the latter problem is easily solved when pt
is assumed to be a martingale. Indeed, if we have
dpt = σt dWt (3.30)
for some adapted, continuous and positive process σt , then αt = 1 and we simply
have
st = m(1)σt (3.31)
circumventing the need for explicit functions ρ and f . This result provides a the-
oretical foundation for the empirical claim made in [24] that the spread is a linear
function of volatility.
Plugging this optimal spread back into the objective function, the market maker’s
expected profit and loss (P&L) is given by:
"Z #
T
M (1)E σt2 dt . (3.32)
0

Given that the dynamics of the market maker’s inventory are:

q
dLt = −ρt (st )ft (st )dpt + ft (st ) 1 − ρ2t (st )σt dWt⊥ , (3.33)
we conclude that the inventory is also a martingale. In option-pricing terms,
−ρt (st )ft (st ) is the Gamma exposure of the market maker: it measures the changes
in inventory due to a change in the price. It is negative. The expected profit and
loss however, exhibits positive Vega as it is an increasing function of volatility.
3.2.2. Explicit cases. Other cases where αt can be computed explicitly are:
• the Black-Scholes model
dpt = µpt dt + σpt dWt (3.34)
in which case we obtain:
E [ pT | Ft ] = pt eµ(T −t) ; Zt = σpt eµ(T −t) , (3.35)
and hence
µ  µ(T −t) 
αt = e − 1 + eµ(T −t) . (3.36)
σ2
• the case of a mean reverting (Ornstein-Uhlenbeck) price process
dpt = ρ(p0 − pt )dt + σdWt (3.37)
in which case:
E [ pT | Ft ] = p0 + e−ρ(T −t) (pt − p0 ); Zt = σe−ρ(T −t) , (3.38)
and hence
ρ 2
 
αt = − 2
(pt − p0 ) e−ρ(T −t) − 1 + e−ρ(T −t) . (3.39)
σ
 APPLICATIONS OF A NEW SELF-FINANCING EQUATION 17

Unlike in the martingale case, it is hard to obtain any tractable formulas without
specifying a functional form for ρ and f . In the case where ρ(x) = 1/(1 + x) and
f (x) = 1/(1 + x)2 , the optimal spread becomes
√
st = σt 1 + 3αt (3.40)
Note that m is an increasing function of αt . To compare with the martingale
case, where αt = 1, we therefore want to compare the ratio of αt to 1 in order
to study the impact of the model assumptions on the market maker’s profit and
inventory volatility.
• For the Black-Scholes model, αt is larger than 1 for µ > 0. For µ < 0, there
exists a critical value depending on T and σ for which this ratio flips sign.
• In the case of an Ornstein-Uhlenbeck process, αt is smaller than 1 iff
σ2
(pt − p0 )2 < (3.41)
ρ
that is, if the current price pt isn’t too far from the long-term average p0 .
In line with intuition, the market maker quotes larger spreads, expects less profit,
and captures less volume in the ’momentum’ Black-Scholes model, as compared to
the martingale case. This is because the spread is an increasing function of αt , the
expected profit a decreasing function of αt and the volatility of the market maker’s
inventory is a decreasing function of the spread and hence of αt . In a mean-reverting
market, unless the price is significantly away from its long-term trend as measured
by the inequality (3.41), the market maker quotes smaller spreads, expects more
profit and captures more volume than in the two other market models.

References
[1] Y. Ait-Sahalia and J. Jacod. High-Frequency Financial Econometrics. Princeton University
Press, 2014.
[2] 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.
[3] A. Alfonsi, A. Schied, and A. Slynko. Order book resilience, price manipulation, and the
positive portfolio problem. SIAM Journal on Financial Mathematics , 3(1):511-533, 2012
[4] C. Aliprantis and K. Border. Infinite Dimensional Analysis. Springer, 2006.
[5] R. Almgren and N. Chriss. Optimal execution of portfolio transactions. Journal of Risk,
3(2):5–39, 2000.
[6] T. Ané and H. Geman. Order flow, transaction clock, and normality of asset returns. The
Journal of Finance, 55(5):2259–2284, 2000.
[7] M. Avellaneda and S. Stoikov. High-frequency trading in a limit order book. Quantitative
Finance, 8(3):217–224, 2007.
[8] J. . 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.
[9] R. Carmona, and K. Webster. Trading frictions in high frequency markets. Princeton Uni-
versity Tech. Rep. , 2014. arxiv, http://arxiv.org/abs/1709.02015
[10] R. Carmona, and K. Webster. The Self-Financing Equation in Limit Order Book Markets.
Finance & Stochastics, 2019 (to appear). arxiv, http://arxiv.org/abs/??????
[11] A. Cartea, and S. Jaimungal. Risk metrics and fine tuning of high frequency trading strategies.
Mathematical Finance, 25(3):576-611, 2015.
[12] T. Chellathurai and T. Draviam. Dynamic portfolio selection with nonlinear transaction
costs. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences,
461(2062):3183–3212, 2005.
[13] U. Cetin, H. Mete Soner, and N. Touzi. Options hedging for small investors under liquidity
costs. Finance Stochastics, 14(3): 317-341, 2010.
18 RENÉ CARMONA AND KEVIN WEBSTER

[14] R. Cont and A. de Larrard. Order book dynamics in liquid markets: limit theorems and
diffusion approximations. Working paper, 2011.
[15] R. Cont and A. de Larrard. Price dynamics in a markovian limit order book market. SIAM
Journal for Financial Mathematics, 4(1):1–25, 2013.
[16] R. Cont, S. Stoikov, and R. Talreja. A stochastic model for order book dynamics. Operations
Research, 58(3):549–563, 2010.
[17] J. Jacod and P. Protter. Discretization of Processes. Springer, 2011.
[18] R. Liu and J. Muhle-Karbe. Portfolio choice with stochastic investment opportunities:
a user’s guide. Proc. 1st Princeton Summer School in Mathematical Finance, 2013.
http://arxiv.org/abs/1311.1715
[19] M. Magill and G. Constantinides. Portfolio selection with transactions costs. Journal of Eco-
nomic Theory, 13(2):245 – 263, 1976.
[20] A. Obizhaeva and J. Wang. Optimal trading strategy and supply/demand dynamics. Journal
of Financial Markets, 16:1–32, 2013.
[21] S.E. Shreve and H.M. Soner. Optimal investment and consumption with transaction costs.
Annals of Applied Probability, 4(3):609–692, 1994.
[22] H. M. Soner, and N. Touzi. Hedging Under Gamma Constraints By Optimal Stopping and
Face-Lifting. Mathematical Finance, 17(1):59–79, 2007.
[23] S. Stoikov and M. Saglam. Option market making under inventory risk. Review of Derivatives
Research, 12(1):55–79, 2009.
[24] M. Wyart, J. . 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.

ORFE, Bendheim Center for Finance, Princeton University