Slides Market Microstructure

Market Making and Inventory Management
Models: Where do we Stand?
Olivier Guéant
Market Microstructure: The CFM-Imperial Workshop – 11/12/2023
1
A journey through market
making: from solving equations
to helping practitionners
Introduction to market making
2
Market makers
2
Market makers
• Activity: providing bid and ask/offer prices to other market
participants.
2
Market makers
participants.
• The way they make money: capturing part of the bid-ask spreads.
2
Market makers
participants.
• Risks:
2
Market makers
participants.
• Risks:
• Holding inventory
2
Market makers
participants.
• Risks:
→ face the risk that prices move adversely without them being able
to unwind their position.
2
Market makers
participants.
• Risks:
• Information asymmetry / adverse selection by informed traders
2
Market makers
participants.
• Risks:
→ face the risk of buying / selling when the price will go down / up.
2
Market makers
participants.
• Risks:
→ face the risk of buying / selling when the price will go down / up.
Remark: The talk is mainly about market making in OTC markets. Not
market making in limit order books (no tick size, no queue, no priority).
2
The early literature
3
Models regarding inventory cost / management
• Stoll (1978) • Amihud and Mendelson (1980)

• Ho and Stoll (1981, 1983) • O’Hara and Oldfield (1986)
3

Models regarding adverse selection
• Copeland and Galai (1983)

• Glosten and Milgrom (1985)
• Easley and O’Hara (1987)
3


An economic literature about the determinants of bid-ask spreads in the

1980s and 1990s: Hasbrouck, Huang and Stoll, MRR, etc.
3


An economic literature about the determinants of bid-ask spreads in the

1980s and 1990s: Hasbrouck, Huang and Stoll, MRR, etc.
The financial mathematics community only got interested in

market making from 2008 following the paper by Avellaneda and
Stoikov.
3
Overview of my journey
4
The roots
4
The roots
• Post-PhD inspiration (2010): met C.-A. Lehalle at Crédit Agricole
Cheuvreux who introduced me to Avellaneda-Stoikov (A-S) paper.
4
The roots
• Initial focus and mathematical results: Solved the PDEs of A-S (one
asset) + provided closed-form approximations of optimal quotes –
initially as a theoretical / mathematical exercise.
4
The roots
Multiple interactions with the industry
4
The roots

• OTC trading (neglected area of academic research): Contacted by
bond dealers in London and NYC, by FX dealers (Sasha here in
London), for adapting models to match real-world trading
environments.
4
The roots

• OTC trading (neglected area of academic research): Contacted by
bond dealers in London and NYC, by FX dealers (Sasha here in
London), for adapting models to match real-world trading
environments.
• DeFi: Recently contacted by decentralized finance players to build
new Automated Market Makers.
4
The Avellaneda-Stoikov
framework (revisited with more
recent results)
Setup of the model (I)
One asset: reference price process (“mid”-price) (St )t
dSt = σdWt .
→ Can be the CBBT / CP+ for corporate bonds or a homemade

reference price.
→ Can be EBS / Refinitiv mid price or a homemade composite.
5
Setup of the model (II)
6
• Bid and ask prices of the MM denoted respectively by
Stb = St − δtb and Sta = St + δta .
6
• Point processes N b and N a for the transactions (size ∆).

→ Inventory (qt )t :
dqt = ∆dNtb − ∆dNta .
6
• Point processes N b and N a for the transactions (size ∆).

→ Inventory (qt )t :
dqt = ∆dNtb − ∆dNta .
• Competition and demand is modeled indirecty through the

probability / intensity of jumps.
6
Setup of the model (III)
• The intensities (λbt )t and (λat )t of N b and N a depend on the

distance to the reference price:
λbt = Λb (δtb )1qt− <Q and λat = Λa (δta )1qt− >−Q .

Λb , Λa decreasing (of course!)
7


• Cash process (Xt )t :
dXt = ∆Sta dNta − ∆Stb dNtb = −St dqt + δta ∆dNta + δtb ∆dNtb .
7


• Cash process (Xt )t :
dXt = ∆Sta dNta − ∆Stb dNtb = −St dqt + δta ∆dNta + δtb ∆dNtb .
Three state variables: X (cash), q (inventory), and S (price).
7
PnL and objective function
The PnL at time T of a market maker in the above model is:
PnLT = XT + qT ST = X0 + q0 S0
Z T
+ δ a ∆dNta + δtb ∆dNtb + σqt dWt
0 |t {z } | {z }
spread capture inventory+price risk
8
The PnL at time T of a market maker in the above model is:
PnLT = XT + qT ST = X0 + q0 S0
Z T
+ δ a ∆dNta + δtb ∆dNtb + σqt dWt
0 |t {z } | {z }
spread capture inventory+price risk
The original Avellaneda-Stoikov’s model considers a CARA utility:
CARA objective function (Model A)
sup E [− exp (−γ(XT + qT ST ))] ,

(δta )t ,(δtb )t ∈A
where γ is the absolute risk aversion parameter, and A the set of

predictable processes bounded from below.
8
9
Variant (Cartea, Jaimungal et al.) with a running penalty – proxy of

variance:
Expected PnL minus a quadratic running penalty (Model B –
practionners’ preferred )
" #
γ 2 T 2
Z
sup E XT + qT ST − σ qt dt
(δta )t ,(δtb )t ∈A 2 0
i.e.
Z T γ 2 2

sup E ∆δta Λa (δta )1qt− >−Q + ∆δtb Λb (δtb )1qt− <Q − σ qt dt
(δta )t ,(δtb )t ∈A 0 2
where γ is a kind of absolute risk aversion parameter.
9
Variant (Cartea, Jaimungal et al.) with a running penalty – proxy of

variance:
Expected PnL minus a quadratic running penalty (Model B –
practionners’ preferred )
" #
γ 2 T 2
Z
sup E XT + qT ST − σ qt dt
(δta )t ,(δtb )t ∈A 2 0
i.e.
Z T γ 2 2

sup E ∆δta Λa (δta )1qt− >−Q + ∆δtb Λb (δtb )1qt− <Q − σ qt dt
(δta )t ,(δtb )t ∈A 0 2
where γ is a kind of absolute risk aversion parameter.
→ Optimal control on a very simple finite graph (truncated ∆Z)
9
Value function θ (Model B)
10
Hamilton-Jacobi equation (Model B)

1 2 2 b θ(t, q) − θ(t, q + ∆)
0 = ∂t θ(t, q) − γσ q + 1q<Q ∆H
2 ∆

θ(t, q) − θ(t, q − ∆)
+1q>−Q ∆H a
∆
with final condition θ(T , q) = 0.
10
Hamilton-Jacobi equation (Model B)

1 2 2 b θ(t, q) − θ(t, q + ∆)
0 = ∂t θ(t, q) − γσ q + 1q<Q ∆H
2 ∆

θ(t, q) − θ(t, q − ∆)
+1q>−Q ∆H a
∆
with final condition θ(T , q) = 0.
A Legendre-like transform
H b (p) = sup Λb (δ)(δ − p)

δ
H (p) = sup Λa (δ)(δ − p)

a
δ
10
Comments on the equations
11
• The same family of equations appears with Model A after a change

of variables (H transforms are just more complicated).
11

• A system of 2Q/∆ + 1 non-linear ODEs.
11

• Global existence obtained with maximum principles under light
assumptions of the intensity functions:
11

• Global existence obtained with maximum principles under light
assumptions of the intensity functions:
1. Λb/a is C 2 .
0
2. Λb/a < 0.
3. limδ→+∞ Λb/a (δ) = 0.
4. The intensity functions Λb/a satisfy:
00
Λb/a (δ)Λb/a (δ)
sup 2 < 2.
δ 0
Λb/a (δ)
11
Solution of the initial problem – B (verification argument)
By using a verification argument, we find the optimal quotes

Optimal quotes

θ(t, qt− ) − θ(t, qt− + ∆)
δtb∗ = δ̃ b∗
∆

θ(t, qt− ) − θ(t, qt− − ∆)
δta∗ = δ̃ a∗
∆
where
−1 0

δ̃ b/a∗ (p) = argmaxδ Λb/a (δ)(δ − p) = Λb/a −H b/a (p) .
12
Solution of the initial problem – B (verification argument)
By using a verification argument, we find the optimal quotes

Optimal quotes

θ(t, qt− ) − θ(t, qt− + ∆)
δtb∗ = δ̃ b∗
∆

θ(t, qt− ) − θ(t, qt− − ∆)
δta∗ = δ̃ a∗
∆
where
−1 0

δ̃ b/a∗ (p) = argmaxδ Λb/a (δ)(δ − p) = Λb/a −H b/a (p) .
Easy to obtain numerically: a monotone implicit scheme for θ and then

standard optimization techniques to get the optimal quotes.
12
Closed-form approximations – Model B
13
• Optimal control problem with an ergodic limit (T → ∞) often

reached very rapidly in practice.
13

• Continuous approximation of the eigenvalue problem leads to
closed-form approximations of the quotes in Model A and B (here
we focus on B) when Λb (δ) = Λa (δ) = Ae −kδ :
13

Closed-form approximations (Model B)
r r
b∗ 1 2q + ∆ γσ 2 e a∗ 1 2q − ∆ γσ 2 e
δ∞ (q) ' + δ∞ (q) ' −
k 2 2kA∆ k 2 2kA∆
13

Closed-form approximations (Model B)
r r
b∗ 1 2q + ∆ γσ 2 e a∗ 1 2q − ∆ γσ 2 e
δ∞ (q) ' + δ∞ (q) ' −
k 2 2kA∆ k 2 2kA∆
Closed-form approximations spread / skew (Model B)

r r
2 γσ 2 e γσ 2 e
spread ' + ∆ skew(q) ' 2q
k 2kA∆ 2kA∆
13
Extensions I
Extensions I
14
Extensions I
Basic ideas regarding prices
14
Extensions I

• Going from arithmetic to geometric Brownian motion (easy: just
model size in $ and not in number of securities).
14
Extensions I

• Including a drift and / or price jumps (easy).
14
Extensions I

• Including stoch. vol. models (easy but lead to a system of parabolic
PDEs in dimension depending of the number of factors).
14
Extensions I

Natural ideas regarding price dynamics for today’s audience
14
Extensions I

Natural ideas regarding price dynamics for today’s audience

• Modeling price by microstructural models / point processes:
→ lead to a system of PDEs with nonlocal terms.
14
Extensions I
15
Extensions I
Important practical considerations
15
Extensions I

• Multiple sizes of transactions (one quote per size): does not change
the class of equations.
15
Extensions I

• Multiple tiers of clients (quotes per tier): does not change the class
of equations
→ can also be useful to get signal/drift.
15
Extensions I

of equations
• Doubt on the reference price: perturbation analysis. Not obvious to
filter and be time consistent.
15
Extensions I

of equations
• Asymmetric long/short penalties (important for bonds).
15
Extensions I

of equations
• Asymmetric long/short penalties (important for bonds).
• D2C vs. D2D (internalization vs. externalization) + market impact
on the D2D segment.
15
A first model for the FX market
Organization of the FX market (Schrimpf-Sushko)
16
RFSs and RFQs (D2C) and access to multiple platforms (D2D and
all-to-all)
16
RFSs and RFQs (D2C) and access to multiple platforms (D2D and
all-to-all) → dealers can internalize or externalize the flow.
16
The internalization-externalization dilemma (S.-S.)
17
The internalization-externalization dilemma (S.-S.)
The internalization vs. externalization dilemma has to be taken into

account to build efficient algorithms.
17
A model with client tiering and multiple sizes
• In practice, the id of clients is known → client tiering (N tiers).

• Sizes are not constant and dealers propose price ladders (size is
denoted by z).
State variables: Equations with internalization only
dSt = σdWt
XN Z
dqt = zJ b,n (ds, dz) − zJ a,n (ds, dz)
n=1 R∗
+
| {z }
client flow from tier n
N Z
X
dXt = −St dqt + zδ a,n (t, z)J a,n (dt, dz) + zδ b,n (t, z)J b,n (dt, dz)
n=1 R∗
+
| {z }
spread capture
18
The model
State variables: Equations with externalization (like in
Almgren-Chriss)
dSt = σdWt +kvt dt

XN Z
dqt = zJ b,n (ds, dz) − zJ a,n (ds, dz)+vt dt
n=1 R∗
+
N Z
X
n=1 R∗
+
−vt St dt − L(vt ) dt can also be done with isolated trades

| {z }
execution costs
19
The model
State variables: Equations with externalization (like in
Almgren-Chriss)
dSt = σdWt +kvt dt

XN Z
dqt = zJ b,n (ds, dz) − zJ a,n (ds, dz)+vt dt
n=1 R∗
+
N Z
X
n=1 R∗
+
−vt St dt − L(vt ) dt can also be done with isolated trades

| {z }
execution costs
Objective function
" #
Z T
γ
E XT + qT ST − σ 2 qt2 dt .
2 0
19
Estimating intensities: tiering
Tiering
1.0
0.8
25
Normalized Intensity
0.6 20
( bps 1 )
15
0.4
10
1
5
0.2
0
4 3 2 1 0 1 2
1
0.0
1.25 1.00 0.75 0.50 0.25 0.00 0.25 0.50 0.75
(bps)
1
Λb/a,n (δ, z) = λb/a,n (z)f b/a,n (δ) = λb/a,n (z)
1+ exp(αb/a,n + β b/a,n δ)
20
HJB equation
A partial integro-differential equation



R
 γ 2 2 PN b,n θ(t,q)−θ(t,q+z)
0 = ∂ θ(t, q) − σ q + zH λb,n (z)dz

t


 2 n=1 ∗ z

 R R +

N a,n θ(t,q)−θ(t,q−z)
λa,n (z)dz − H (∂q θ(t, q) + kq)
P

 − n=1 zH z
R∗



 +

θ(T , q) = 0,
where
H b/a,n : p ∈ R 7→ sup f b/a,n (δ)(δ − p) and H(p) := sup vp − L(v ).

δ v
21
Solution of the control problem
Optimal controls

b,n∗ b,n∗ θ(t, qt− ) − θ(t, qt− + z)
δ (t, z) = δ̄
z

a,n∗ a,n∗ θ(t, qt− ) − θ(t, qt− − z)
δ (t, z) = δ̄
z
where
−1 0

δ̄ b/a,n∗ (p) = f b/a,n (z, ·) −H b/a,n (p) ,
and
vt∗ = H0 (∂q θ(t, qt− ) + kqt− ) .
22
Optimal quotes and execution rate
Optimal Dealer Strategy

1.00
Pure Flow
Internalization 0.4
0.75
0.50
0.2
Bid and Ask Quotes (bps)
Execution Rate ( M s 1)
0.25 Tier 2 Optimal Ask Ladder
0.00 0.0
0.25
Tier 1 Optimal Bid Ladder Optimal Execution Rate
0.2
0.50
0.75
0.4
1.00
100 75 50 25 0 25 50 75 100
Inventory ( M)
23
1.0
14
0.8 12
Risk Neutralization Time (min)

10
Volume Fraction
0.6
8
0.4 6
4
0.2
2
0.0 0
= 10 4 = 10 3 = 10 2 = 10 1
24
Optimal risk/reward curve
12
10
P&L ($K)
5 10 15 20 25 30
Risk ($K)
25
Extensions II
Extensions II
26
Extensions II
D2D / hedging was regarded above as in Almgren-Chriss – could also be

done with impulse control (trading a quantity at once and paying a
spread).
Detailing the sources of D2D liquidity
26
Extensions II

spread).
• Dark pool on the D2D segment: not easy to model to keep
low-dimensionality.
26
Extensions II

spread).
low-dimensionality.
• LOB on the D2D segment: not easy to keep low-dimensionality.
26
Extensions II

spread).
low-dimensionality.
Other sources of liquidity
26
Extensions II

spread).
low-dimensionality.

• RFQ in addition to RFS: linked to pre-hedging part of your expected
flow (done with Sasha B. and Philippe B.).
26
Extensions II

spread).
low-dimensionality.

• RFQ in addition to RFS: linked to pre-hedging part of your expected
flow (done with Sasha B. and Philippe B.).
• Hedging with cointegrated assets (spot vs. futures): initially done
with Fayçal Drissi → raises the question of a multi-asset setting!
26
Extensions II
27
Extensions II
Other important extensions
27
Extensions II

• Different and asymmetric liquidity regimes
→ Hawkes processes (with exponential kernel to stay Markovian)
→ Markov-modulated Poisson processes (a few liquidity states):
done with P. Bergault.
27
Extensions II

• Better modelling of adverse selection
→ trades convey information depending on id and price (a fair deal
with a standard client may be better than a seemingly good deal
with an informed client).
27
Extensions II

• Better modelling of adverse selection
→ trades convey information depending on id and price (a fair deal
with a standard client may be better than a seemingly good deal
with an informed client).
• Price signalling: need to model the impact of streamed prices.
27
Regarding extensions
28
Many effects can be taken into account in the one-asset case
28

• Important in itself.
28

• Important for decentralized finance
→ In 2021, 3 days before Christmas, I received an email from David
Bouba who co-founded an AMM (Swaap): a whitepaper, interesting
remarks, and an invitation to discuss with him.
28

→ Less than a year later, we released a paper to build AMM based
on price oracles or offchain reference prices (Bergault, Bertucci,
Bouba, Guéant).
28

→ Less than a year later, we released a paper to build AMM based
on price oracles or offchain reference prices (Bergault, Bertucci,
Bouba, Guéant).
• Not really satisfying for FX (correlations + triplets) or corporate
bonds (many securities for one issuer).
→ diversification and liquidity differences must be taken into
account.
28
A model for multi-currency
market making
A model for multi-currency market making
29
I present here the model we wrote with P. Bergault (Risk rising star
award) and A. Barzykin.
29
• d currencies (currency 1, USD, used for accounting)
29

• d stochastic processes denoted by (St1 )t∈[0,T ] , . . . , (Std )t∈[0,T ] that
model the market price of each of the d currencies in terms of the
reference one.
29

reference one.
• N tiers of clients.
29

reference one.
• R∗+ −marked point processes J n,i,j (dt, dz) modelling transactions
with clients from tier n regarding the currency pair (i, j) – z is the
size variable measured in reference currency.
29

reference one.
• R∗+ −marked point processes J n,i,j (dt, dz) modelling transactions
with clients from tier n regarding the currency pair (i, j) – z is the
size variable measured in reference currency.
• Logistic kernels νtn,i,j (dz) = Λn,i,j z, δ n,i,j (t, z) dz where the δs are

now multiplicative (in bps) and Λn,i,j (z, δ) = λn,i,j (z)f n,i,j (z, δ) with
1
f n,i,j (z, δ) = αn,i,j (z)+β n,i,j (z)δ
.
1+e
29
30
• Possibility to externalize on the D2D segment of the market.
30
• Possibility to externalize
on the D2D segment of the market. → for
i,j
i < j, process ξt modeling the amount (expressed in
t∈[0,T ]
reference currency) per unit of time of currency i bought by the
dealer and paid in currency j.
30
i,j
t∈[0,T ]
i,j
• Execution costs: Li,j (ξ) = ψ i,j |ξ| + η i,j |ξ|1+φ .
30
i,j
t∈[0,T ]
i,j
• Execution costs: Li,j (ξ) = ψ i,j |ξ| + η i,j |ξ|1+φ .
Resulting dynamics of the inventories:
N X
d Z
X z
dqti J n,i,j (dt, dz) − J n,j,i (dt, dz)

= i
n=1 j=1 z∈R∗
+
St
j6=i
 
d i,j i−1 j,i
X ξt
X ξ t
+ −  dt.
Si
j=i+1 t j=1
Sti
30
Dynamics of the account where fees and execution costs are accounted
for in reference currency:
N
X X Z
dXt = zδ n,i,j (t, z)J n,i,j (dt, dz)
n=1 1≤i6=j≤d z∈R∗
+
X
− Li,j ξti,j .
1≤i<j≤d
31
Dynamics of the account where fees and execution costs are accounted
for in reference currency:
N
X X Z
dXt = zδ n,i,j (t, z)J n,i,j (dt, dz)
n=1 1≤i6=j≤d z∈R∗
+
X
− Li,j ξti,j .
1≤i<j≤d
Dynamics of market exchange rates (covariance matrix Σ):

 
d
X i−1
X
dSti = µit Sti dt + σ i Sti dWti + k i  ξti,j − ξtj,i  Sti dt.
j=i+1 j=1
31

Writing all inventories in reference currency terms, i.e. Yti t∈[0,T ]
=

qti Sti t∈[0,T ] , we get
 
d
X i−1
X
dYti = µit Yt−
i
dt + σ i Yt−
i
dWti + k i  ξti,j − ξtj,i  Yt−
i
dt
j=i+1 j=1
N X
X d Z
z J n,i,j (dt, dz) − J n,j,i (dt, dz)

+
n=1 j=1 z∈R∗
+
j6=i
 
d
X i−1
X
+ ξti,j − ξtj,i  dt.
j=i+1 j=1
32
Optimization problem
Mathematically, we assume that the dealer wants to maximize

" d
#
Z T
X γ
E XT + YTi − Yt| ΣYt dt
2 0
i=1
33
HJB equation
The HJB equation is

γ 1 2
0 = ∂t θ(t, y ) + y | µ(t) + y | D(µ(t))∇y θ(t, y ) − y | Σy + Tr D(y )ΣD(y )Dyy

θ(t, y )
Z 2 2
N
θ(t, y ) − θ(t, y + ze i − ze j )
X X
+ zH n,i,j z, λn,i,j (z)dz
z
n=1 1≤i6=j≤d
R∗
+
X
+ Hi,j ∂y i θ(t, y ) − ∂y j θ(t, y )
1≤i<j≤d

+k i y i 1 + ∂y i θ(t, y ) − k j y j 1 + ∂y j θ(t, y ) ,
with θ(T , y ) = 0,
34
HJB equation
The HJB equation is

γ 1 2
0 = ∂t θ(t, y ) + y | µ(t) + y | D(µ(t))∇y θ(t, y ) − y | Σy + Tr D(y )ΣD(y )Dyy

θ(t, y )
Z 2 2
N
θ(t, y ) − θ(t, y + ze i − ze j )
X X
+ zH n,i,j z, λn,i,j (z)dz
z
n=1 1≤i6=j≤d
R∗
+
X
+ Hi,j ∂y i θ(t, y ) − ∂y j θ(t, y )
1≤i<j≤d

+k i y i 1 + ∂y i θ(t, y ) − k j y j 1 + ∂y j θ(t, y ) ,
with θ(T , y ) = 0, where
H n,i,j : (z, p) ∈ R∗+ × R 7→ sup f n,i,j (z, δ)(δ − p)

δ
i,j
H : p ∈ R 7→ sup pξ − Li,j (ξ).
ξ
34
Optimal strategy
Optimal pricing
θ(t, Yt− ) − θ(t, Yt− + ze i − ze j )

δ n,i,j∗ (t, z) = δ̄ n,i,j z,
z
with δ̄ n,i,j (z, p) = (f n,i,j )−1 −∂p H n,i,j (z, p) .

35
Optimal strategy
Optimal pricing
θ(t, Yt− ) − θ(t, Yt− + ze i − ze j )

δ n,i,j∗ (t, z) = δ̄ n,i,j z,
z
with δ̄ n,i,j (z, p) = (f n,i,j )−1 −∂p H n,i,j (z, p) .

Optimal hedging (externalization)

0

ξti,j∗ = Hi,j ∂y i θ(t, Yt− ) − ∂y j θ(t, Yt− )

j
+k i Yt−
i
1 + ∂y i θ(t, Yt− ) − k j Yt− 1 + ∂y j θ(t, Yt− )
35
Numerical approximation
36
• Equations are there but no grid in high dimension
36

• Simple structure for Σ: low-dimensional factor methods (not for this
problem).
36

problem).
• (Deep) Reinforcement Learning (slow and the danger of NN).
36

problem).
• Trick in Bergault, Evangelista, Guéant, Vieira:
36

problem).
• HJB equations associated with continuous state space optimal
control problems boil down to Riccati equations when Hamiltonians
are quadratic.
36

problem).
are quadratic.
• For HJ equations on graphs, the easy case is that of exponential
Hamiltonians (hence the special case of exponential intensities).
36

problem).
are quadratic.
• But...
36

problem).
are quadratic.
• But...
• Quadratic approximations of the H functions (makes sense for not
too asymmetric flows)...
36

problem).
are quadratic.
• But...
• Gives an approximate value function θ̃: Polynomial of degree 2 in y
with coefficients solving a Riccati-like equation (no curse of
dimensionality).
36

problem).
are quadratic.
• But...
• Gives an approximate value function θ̃: Polynomial of degree 2 in y
with coefficients solving a Riccati-like equation (no curse of
dimensionality).
• θ̃ is plugged in the above equations to get great pricing and hedging
strategies. 36
Approximation of the Hamiltonians

θ(t, y ) − θ(t, y + z) θ(t, y ) − θ(t, y − z)
Hb + Ha
z z
H b (p) + H a (−p)
H b(p) H a( p) H b(p) + H a( p)
6.0
5 5
5.5
4 4 5.0
4.5
3 3
4.0
2 2 3.5
3.0
1 1
2.5
0 0 2.0
1.00 0.75 0.50 0.25 0.00 0.25 0.50 0.75 1.00 1.00 0.75 0.50 0.25 0.00 0.25 0.50 0.75 1.00 1.00 0.75 0.50 0.25 0.00 0.25 0.50 0.75 1.00
37
Optimal Top of Book Quotes

2.0
GBPUSD
1.5 EURUSD
EURGBP
1.0
Bid and Ask Quotes (bps)
0.5
0.0
0.5
1.0
1.5
2.0
100 75 50 25 0 25 50 75 100
GBP Inventory (M$)
38
Optimal Externalization Strategy

JPYUSD
CHFUSD
0.4 GBPUSD
EURUSD
CHFJPY
GBPJPY
Execution Rate (M$ s 1)
GBPCHF
0.2 EURJPY
EURCHF
EURGBP
0.0
1.0
EUR GBP CHF JPY
0.8
0.2
0.6
0.4
0.2
0.4
0.0
JPY CHF GBP EUR
100 75 50 25 0 25 50 75 100
EUR Inventory (M$)
39
Optimal risk/reward curve
Inventory Risk and Probability Distribution

30
20
10
EUR Inventory (M$)
10
20
30
30 20 10 0 10 20 30
GBP Inventory (M$)
40
Example (video)
Time for a short animation
41
General conclusion
General conclusion
Final remarks
42
General conclusion
Final remarks
• Market making models are important for OTC markets.
42
General conclusion
Final remarks
• Many effects need better models (adverse selection, information
leakage (skew), etc.).
42
General conclusion
Final remarks
• Finding a way to tackle more effects in high dimensions.
42
General conclusion
Final remarks
• Work on mid-prices in OTC markets (filtering).
42
General conclusion
Final remarks
• Work on mid-prices in OTC markets (filtering).
• Research on AMMs is very active (Paris, Berkeley, UK, etc.) →
interesting even beyond cryptos.
42
The End
Thank you. Questions?
43

Slides Market Microstructure

Uploaded by

Copyright:

Available Formats

Slides Market Microstructure

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Slides Market Microstructure

Uploaded by

Copyright:

Available Formats

Market Making and Inventory Management

Models: Where do we Stand?

• Stoll (1978) • Amihud and Mendelson (1980)

• Stoll (1978) • Amihud and Mendelson (1980)

Models regarding adverse selection

• Copeland and Galai (1983)

• Stoll (1978) • Amihud and Mendelson (1980)

Models regarding adverse selection

• Copeland and Galai (1983)

An economic literature about the determinants of bid-ask spreads in the

• Stoll (1978) • Amihud and Mendelson (1980)

Models regarding adverse selection

• Copeland and Galai (1983)

An economic literature about the determinants of bid-ask spreads in the

The financial mathematics community only got interested in

Multiple interactions with the industry

Multiple interactions with the industry

Multiple interactions with the industry

One asset: reference price process (“mid”-price) (St )t

→ Can be the CBBT / CP+ for corporate bonds or a homemade

• Bid and ask prices of the MM denoted respectively by

Stb = St − δtb and Sta = St + δta .

• Bid and ask prices of the MM denoted respectively by

Stb = St − δtb and Sta = St + δta .

• Point processes N b and N a for the transactions (size ∆).

dqt = ∆dNtb − ∆dNta .

• Bid and ask prices of the MM denoted respectively by

Stb = St − δtb and Sta = St + δta .

• Point processes N b and N a for the transactions (size ∆).

dqt = ∆dNtb − ∆dNta .

• Competition and demand is modeled indirecty through the

• The intensities (λbt )t and (λat )t of N b and N a depend on the

λbt = Λb (δtb )1qt− <Q and λat = Λa (δta )1qt− >−Q .

• The intensities (λbt )t and (λat )t of N b and N a depend on the

λbt = Λb (δtb )1qt− <Q and λat = Λa (δta )1qt− >−Q .

• Cash process (Xt )t :

• The intensities (λbt )t and (λat )t of N b and N a depend on the

λbt = Λb (δtb )1qt− <Q and λat = Λa (δta )1qt− >−Q .

• Cash process (Xt )t :

Three state variables: X (cash), q (inventory), and S (price).

The PnL at time T of a market maker in the above model is:

The PnL at time T of a market maker in the above model is:

The original Avellaneda-Stoikov’s model considers a CARA utility:

CARA objective function (Model A)

sup E [− exp (−γ(XT + qT ST ))] ,

where γ is the absolute risk aversion parameter, and A the set of

Variant (Cartea, Jaimungal et al.) with a running penalty – proxy of

where γ is a kind of absolute risk aversion parameter.

Variant (Cartea, Jaimungal et al.) with a running penalty – proxy of

where γ is a kind of absolute risk aversion parameter.

→ Optimal control on a very simple finite graph (truncated ∆Z)

Hamilton-Jacobi equation (Model B)

Hamilton-Jacobi equation (Model B)

H b (p) = sup Λb (δ)(δ − p)

H (p) = sup Λa (δ)(δ − p)

• The same family of equations appears with Model A after a change

• The same family of equations appears with Model A after a change

• The same family of equations appears with Model A after a change

• The same family of equations appears with Model A after a change