Algorithmic Trading

PIMS Summer School 2016

Market Making

Álvaro Cartea & Sebastian Jaimungal

University of Oxford & University of Toronto

July, 2016

(c) Cartea & Jaimungal, 2016 Algo Trading July, 2016 1 / 24

The Limit Order Book
I The limit order book is a record of collective interest to buy or
sell certain quantities of an asset at a certain price.
Buy Orders Sell Orders
Price Volume Price Volume
60.00 80 60.10 75
59.90 100 60.20 75
59.80 90 60.30 50

I Graphical representation of the limit order book:

Volume

Price

(c) Cartea & Jaimungal, 2016 Algo Trading July, 2016 2 / 24

Market Orders
I An incoming market order lifts limit orders from the book.

Volume

Price

(c) Cartea & Jaimungal, 2016 Algo Trading July, 2016 3 / 24

Market Orders
I An incoming market order lifts limit orders from the book.

Volume

Price

(c) Cartea & Jaimungal, 2016 Algo Trading July, 2016 4 / 24

Market Orders
I An incoming market order lifts limit orders from the book.

Volume

Price

(c) Cartea & Jaimungal, 2016 Algo Trading July, 2016 5 / 24

Agent’s Goal
I Optimally place limit orders in the limit order book (LOB)

Volume

Price

(c) Cartea & Jaimungal, 2016 Algo Trading July, 2016 6 / 24

Market Making

I The market maker’s problem is to find prices at which to

post limit buy/sell orders to profit from round-trip trades

I The benchmark models: Ho & Stoll (81), Avellanda &

Stoikov (08) Cartea & Jaimungal (12), and others
I Need to account for
I market order arrival rate

I Probability that market maker is filled at a given level

I midprice dynamics

(c) Cartea & Jaimungal, 2016 Algo Trading July, 2016 7 / 24

Market Making

I Midprice dSt = σ dWt , σ > 0,

I δ ± depth at which the agent posts LOs:
I Sell LOs are posted at a price of St +δt+

I Buy LOs at St −δt−

I M ± Poisson arrival of other participants’ buy (+) and sell (−)

MOs which arrive at with intensities λ± ,

I N δ,± counting processes for the agent’s filled sell (+) and buy
(−) LOs,

I Conditional on an MO arrival, the LO is filled with probability

± ±
e −κ δt with κ± ≥ 0

(c) Cartea & Jaimungal, 2016 Algo Trading July, 2016 8 / 24

Market Maker’s Control Problem
The MM’s performance criteria is
 Z T 
H δ (t, x, S, q) = Et,x,q,S XTδ + QTδ (ST − α QTδ )−φ (Quδ )2 du ,
t

α ≥ 0, and φ ≥ 0. Value function is

H(t, x, S, q) = sup H δ (t, x, S, q) ,

δ ± ∈A

and inventory capped: above by q > 0 and below by q < 0.

I X δ cash process

dXtδ = St +δt+ dNδ,+ − St −δt− dNδ,−

t t .

I Q δ inventory process and satisfies

δ δ
Qtδ = Nt ,− − Nt ,+ .

(c) Cartea & Jaimungal, 2016 Algo Trading July, 2016 9 / 24

DPE

A DPP holds and the value function satisfies the DPE

0 = ∂t H + 1
σ 2 ∂SS H − φ q2
|2 {z } |{z}
midprice diffusion running inv penalty
( )
+ −κ+ δ + +

+ λ sup e H(t, x + (S + δ ), q − 1, S) − H 1q>q
δ+
| {z }
Prob Filled sell LO
 
 − −

+ λ− sup −κ δ
H t, x − (S − δ − ), q + 1, S − H



e 1q<q ,
δ−  | {z } 
Prob Filled buy LO

where 1 is the indicator function, and with terminal condition

H(T , x, S, q) = x + q (S − α q) .

(c) Cartea & Jaimungal, 2016 Algo Trading July, 2016 10 / 24

Solving HJB
I Make an ansatz for H. In particular, write

H(t, x, q, S) = x + q S + h(t, q) .

I first term is accumulated cash

I second term is the book value of the inventory

marked-to-market

I last term is the added value from following an optimal market

making strategy up to T .

n + +
o
φ q 2 = ∂t h(t, q) +λ+ sup e−κ δ δ ++ h(t, q − 1) − h(t, q) 1q>q
δ+ n
− −
o
+λ sup e−κ δ δ −+ h(t, q + 1) − h(t, q) 1q<q ,
−

δ−

with terminal condition h(T , q) = −α q 2 .

(c) Cartea & Jaimungal, 2016 Algo Trading July, 2016 11 / 24
Optimal Controls

Then the optimal depths in feedback form are given by

1
δ +,∗ (t, q) = − h(t, q − 1) + h(t, q) , q 6= q , (1a)
κ+
1
δ −,∗ (t, q) = − − h(t, q + 1) + h(t, q) , q 6= q , (1b)
κ
and boundaries δ +,∗ (t, q) = +∞ and δ −,∗ (t, q) = +∞.

Substituting the optimal controls into the DPE we obtain

+ +
φ q 2 = ∂t h(t, q)+ λκ+ e −1 e −κ (−h(t,q−1)+h(t,q) 1q>q
− − (2)
+ λκ− e −1 e −κ (−h(t,q+1)+h(t,q)) 1q<q .

(c) Cartea & Jaimungal, 2016 Algo Trading July, 2016 12 / 24

Symmetric fill probability
Analytical solution if κ = κ+ = κ− :
1
h(t, q) = log ω(t, q) ,
κ
and stack ω(t, q) into a vector
 0
ω(t, q) = ω(t, q), ω(t, q − 1), . . . , ω(t, q) .

Now, let A denote the (q − q + 1)-square matrix whose rows are

labeled from q to q and whose entries are given by

 −φκ q 2 , i = q,
 + −1

λ e , i = q − 1,
Ai,q = − −1 (3)

 λ e , i = q + 1,
0, otherwise,


2
with terminal and boundary conditions ω(T , q) = e −ακq .

(c) Cartea & Jaimungal, 2016 Algo Trading July, 2016 13 / 24

 Then,
ω(t) = e A(T −t) z , (4)
where z is a (q − q + 1)-dim vector where each component is
2
zj = e −ακ j , j = q, . . . , q. Inserting the controls (1) into the DPE

equation (2) and writing h(t, q) = κ1 log ω(t, q), after some
straightforward computations, one finds that ω(t, q) satisfy the
coupled system of equations

∂t ω(t) + Aω(t) = 0. (5)

(c) Cartea & Jaimungal, 2016 Algo Trading July, 2016 14 / 24

 Optimal Postings

(c) Cartea & Jaimungal, 2016 Algo Trading July, 2016 15 / 24

Optimal postings φ = 0.001

0.02 0.02

q = −2 q=2

Buy Depth (δ− )

Sell Depth (δ+ )

0.015 0.015
q = −1 q=1
q=0 q=0
0.01 q=1 0.01 q = −1
q=2 q = −2
0.005 0.005
q=3 q = −3

0 0
0 10 20 30 0 10 20 30
Time (secs) Time (secs)
(a) φ = 0.001

Figure : The optimal depths as a function of time for various inventory

levels, T = 30, λ± = 1, κ± = 100, q = −q = 3, α = 0.0001, σ = 0.01,
S0 = 100.

(c) Cartea & Jaimungal, 2016 Algo Trading July, 2016 16 / 24

Optimal postings φ = 0.02

0.02 0.02
q = −2 q=2

Buy Depth (δ− )

Sell Depth (δ+ )

0.015 0.015 q=1

q = −1
q=0 q=0
0.01 0.01
q=1 q = −1

q=2 q = −2
0.005 0.005
q=3 q = −3
0 0
0 10 20 30 0 10 20 30
Time (secs) Time (secs)
(a) φ = 0.02

Figure : The optimal depths as a function of time for various inventory

levels, T = 30, λ± = 1, κ± = 100, q = −q = 3, α = 0.0001, σ = 0.01,
S0 = 100.

(c) Cartea & Jaimungal, 2016 Algo Trading July, 2016 17 / 24

Mean reversion in inventory

Given the pair of optimal strategies δ + (t, q), δ − (t, q), the expected
drift in inventories qt is given by

1
µ(t, q) , lim E [Qs − Qt | Qt − = q] ,
s↓t s −t (6)
− −κ− δ −,∗ (t,q) + −κ+ δ +,∗ (t,q)
=λ e −λ e .

I The drift µ(t, q) depends on time.

I For the same level of inventory the speed depends on how

near of far is the strategy from T

(c) Cartea & Jaimungal, 2016 Algo Trading July, 2016 18 / 24

Inventory

25 30

Inventory Drift (7(q))

◦ Sell
Fill Rate (λ e−κ δ(q) )

20 + Buy 20

10
15
0
10
-10
5 -20

0 -30
−10 −5 0 5 10 -10 -5 0 5 10
Inventory (q) Inventory (q)

Figure : Long-term inventory level. Model parameters are: λ± = 1,

κ± = 100, q = −q = 10, α = 0.0001, σ = 0.01, S0 = 100, and
φ = 2 × 10−3 , 10−3 , 5 × 10−4 .


(c) Cartea & Jaimungal, 2016 Algo Trading July, 2016 19 / 24

 10
100.14
Inventory (Qt )

Midprice (St )
5 100.12

100.1
0

100.08
-5
100.06
-10
0 100 200 300 185 190 195
Time (sec.) Time (sec.)

Figure : Inventory and midprice path. Model parameters are: λ± = 1,

κ± = 100, q = −q = 10, φ = 0.02, α = 0.0001, σ = 0.01, S0 = 100.

(c) Cartea & Jaimungal, 2016 Algo Trading July, 2016 20 / 24

Profit and Loss

6 10
φ = 0.001 φ = 0.001
φ = 0.05 φ = 0.05
φ = 0.1 8 φ = 0.1
φ=1 φ=1
Frequency

Frequency
4
6

4
2
2

0 0
0 0.2 0.4 0.6 0.8 −10 −5 0 5 10
Profit and Loss Lifetime Inventory (Qt )

Figure : P&L and Life Inventory of the optimal strategy for 10,000
simulations, λ± = 1, κ± = 100, q = −q = 10, α = 0.0001, σ = 0.01,
and S0 = 100.

(c) Cartea & Jaimungal, 2016 Algo Trading July, 2016 21 / 24

 Market Making with No Terminal Penalty

(c) Cartea & Jaimungal, 2016 Algo Trading July, 2016 22 / 24

Solving HJB with α = 0

Assume no penalties for liquidating inventories at time T . Thus

the ansatz is
H(t, x, q, S) = x + q S + g(t) . (7)
Thus,
n + +
o n − −
o
0 = gt (t) + λ+ sup e−κ δ δ + + λ− sup e−κ δ δ − ,
δ+ δ−

and the optimal postings are:

1
δ ∗,+ =
κ+
and
1
δ ∗,− = .
κ−

(c) Cartea & Jaimungal, 2016 Algo Trading July, 2016 23 / 24

Solving HJB with α = 0

Alternatively note that

I A risk-neutral MM, who does not penalise inventories, seeks
to maximise the probability of being filled at every instant in
time.

I Thus, the MM chooses δ ± to maximise the expected depth

conditional on a market order hitting or lifting the appropriate
± ±
side of the book: maximises δ ± e −κ δ . The FOC
± δ± ± δ±
e −κ − κ± δ ± e −κ = 0. (8)

Thus, we see that the optimal half spreads are as above.

(c) Cartea & Jaimungal, 2016 Algo Trading July, 2016 24 / 24