Princeton

RTG Summer School

in Financial Mathematics (2013)
Travis Craig Johnson1
2013-06-17

Dept. of Eng. Sci. and App. Math., Northwestern University, Evanston, IL 60208, USA. email: traviscj@traviscj.com

Chapter 1

Administrivia

1.1

Lecturers

1. Rene
Carmona
(Princeton
University)
http://www.princeton.edu/ rcarmona/
2. Rama Cont (Imperial College London)
3. Michael Coulon (Princeton University)
4. Jean-Pierre Fouque (UC Santa Barbara)
5. Johannes Muhle-Karbe (ETH Zurich)
6. Alexander Schied (University of Mannheim)
7. Ronnie Sircar (Princeton University)
8. Glen Swindle (Scoville Risk Partners LLC & NYU)
1.2
1.
2.
3.
4.
5.
6.
7.
8.

Times

9-9:50 - Lecture
10-10:50 - Lecture
11-11:30 - Break
11:30-12:20 - Lecture
12:30-14:00 - Lunch
?? 13:30-14:00 - Special Q&A Session ??
14:00-14:50 - Lecture
15:00-15:50 - Guest Lecture

1.3

Locations

1. Week 1: Friend Center room 006

2. Week 2: Computer Science Building room 104
1.4

Links

1. https://orfe.princeton.edu/rtg/fmsummer/reading
2. https://github.com/traviscj/fmsummer

Part I

Systemic Risk

Chapter 2

Systemic Risk 1 - Fouque

2.1

History

4. If we consider
0 < t1 < ... < tn T

1. 60s and 70s:

(a) Problem
of
Portfolio
Allocation
Variance/Markovitz/)
(b) Option Pricing(1973) (Black-Scholes/etc)

(2.5)

then it gives rise to many differences:

(Mean-

(Wt1 Wt0 , Wt2 Wt1 , ...Wtn Wtn1 )

(2.6)

D(wt ws) = N (0, t s)

(2.7)

such that

2. 90s:
(a) Local Volatility
(b) Stochastic Volatility

A bit of Brownian Motion History:

1. Brown
3. 2000s:
2. Buchelier (1900)
(a) Credit - Taking into account the possibility of default(of 3. Albert Einstein (1905): Heat Equation/Brownian Motion tie-in
company, counterparty, etc)
4. Weiner (1930s): Constructed Brownian Motion: Construct
(b) Credit Basket - structuring risk(Mortgages)
a measure over all continuous trajectories
(c) Credit Default Swap
5. Ito (1940s): Figures out chain rule for brownian motion
(d) Collatorized Default Options - Huge Huge Market.
6. Samuelson (1960s): Generalized Brownian Motion
(e) Financial Crisis - Mortgages were completely mispriced.
If we have a bounded function f(t) where
4. 2008-2010:
T
X
(a) NIF: National Institute of Finance (for doing research
(f(ti+1) f(ti)) <
(2.8)
on the banking system)
0
(b) Handbook on Systemic Risk: Cambridge
(c) Dodd-Frank: Created Office of Financial Research (under If we have a brownian motion instead,
the treasury department)
T
X


2.2 Systemic Risk
Wti+1 Wti 
(2.9)
0
Can consider from many views: mathematics, statistics, etc.
Two main approaches:
Book Reference: Carson & Tree???
1. Coupled Diffusions: Continuous time
2.3 Hitting Times
2. Networks
Draw a plot:
The starting point is Brownian Motion. Suppose we start with
1.
time on horizontal axis
an asset:
2.
y = brownian motion on vertical axis
dSt = Stdt
(2.1)
3. line y = .
which we can solve by
We are interested in the time , defined by
St = S0et

a = inf{t > 0 : Wt a}

(2.2)

Okay, but usually we dont know the , and there is some noise! (We are really looking for Wt = a)
We can instead think
How can we add noise? Since is the return, maybe by adding
some white noise
{a t} = { max Ws a}
0st

dSt = St(dt + Noise)

(2.11)

(2.3)

Then
t = Wtt + (2a Wt)t>
W
where the Noise term is given by Brownian Motion, which has
the form
is a Brownian Motion. RP.
dWt
(2.4)
Now suppose we want to find the probability:
with > 0, and (Wt)t0 for t T . The properties of Brownian
Motion:
1. W0 = 0
2. Wt is continuous
3. Independent, increments

(2.10)

P( t, |Wt a| > b)
=P( t, Wt > a + b) + P( t, Wt < a b)
=2P( t, Wt > a + b)
=2N(0,t)(a b)
4

(2.12)

(2.13)
(2.14)
(2.15)
(2.16)

2.3. HITTING TIMES

Figure 2.1: caption

But we also know that
P( t) = 2N(0,t)(a)

(2.17)

A few remarks:
1. is predictable. We say that is announced by a sequence
n, the hitting time of a n1 .
2. Large Deviations: Consider the probability of a very large
deviation (a )
a2

P(a t) e 2t

where we mean this in term of the log.

(2.18)

Chapter 3

Systemic Risk 2 - Fouque

3.1

Joint Distributions

3.2

Geometric Brownian Motion

Consider the joint distribution of (, Wt). We can think of this as Now consider
dSt = St(dt + dWt)
(3.16)


P(Wt = v)
if v > a
g(v)
00
P( t, Wt = v) =
=
which is a stochastic differential equation. We should think of this
P(Wt = 2a v) if v < a
g(2a v) as a convient notation for something more complicated. Basically,
(3.1)
Z t
For nonstandard Brownian Motions,
StdWt
SadWa
(3.17)
0
Xt = x + t + Wt
(3.2)
We will use Itos lemma which lets us do the chain rule with
And then
= inf{t > 0, xt a}
(3.3) a brownian motion:
1
So then,
dg(Wt) = g0(Wt)dWt + g00(Wt)dt
(3.18)
2
xt a (x + t + Wt a)
(3.4)
t + Wt a x
(3.5) where g(t, Wt) is once differentiable in t and twice differentiable
in Wt. Then if we have
ax

(3.6)
t + wt
dXt = tdt + tdWt
(3.19)

Define = , and a
=
Then we have that

ax
.

then

Wt + t

1
dg(Xt) = g0(Xt)dXt + g00(Xt)dhxit
2
(3.7)
Then an SDE like

Wt

Then Girsanovs Thm is

1 2

Mt = eWt 2
Then

E(Mt ) = 1

dXt = b(t, Xt)dt + (t, Xt)dWt

(3.8)

(3.20)

(3.21)

has a solution like

St = S0e(

(3.9)

2
2

)t

+ Wt

(3.22)

and

3.3 Passage Time

dP
|fx = Mt
(3.10) Now consider
dP
r
We can show that Wt + t Wt is a brownian motion under
dSt = St(dt + dWt) = St(rdt + (dWt +
dt)) (3.23)

P . So next,
E (x/Fs) =

1
E(XMt /Fs)
Ms

(3.11)

well then, P: ertSt is a martingale.

3.4

Defaultable Bond

Now we want to have a defaultable bond. The bond starts at

Then we will be able to show
S0 and has a default value of D(if it touches D before maturity


2

E eiu(Wt Ws ) = e 2 (ts)
(3.12) time T then we have no payout, otherwise get 1.) So then the
price of the bond is
Why do we care about all of that? Well, we can do the following:

PD
(3.24)
(0,T ) = E (>T ) = P ( > T ).

= inf{t > 0, Wt a
}
(3.13)
So now we consider
Then
2


{ inf S0e(r 2 )t+Wt > 0}
(3.25)
dP
0<tT

P( t) = E(t) = E t
(3.14)
dP
By taking the logarithm, we get a nonstandard brownian motion,
which we can evaluate. By some computation, we can get
We can use the joint distribution of (, Wt ).
"
#


 1 2r2
2(ax))
a x t
a x + t

S0
+

rT

= N(
) + e 2 N(
) (3.15)
P (0, T ) = e
N(d2 )
N(d2 )
(3.26)
t
t
D
6

3.6. SYSTEMIC RISK

with

If we take = 0 for a second, then



2
log SD0 + r 2 t

d
(3.27)
x
T W
t
2 =
t
N
The main ingredients here were reflection principle and change
of measure. Alternatively, one could do this completely with
partial differential equations.
3.5 Yield

We can represent the yield by

y(0, t) =
3.5.1

1
PD (0, T )
log
T
P(0, T )

(3.28)

Example

Consider some bond B1 with a 10% default rate. This is too

risky for index funds, and not risky enough for hedge fund.
A financial engineer will do this: Buy two bonds, B1 and B2
(both with 10% default rate), then stack them. Now
1. pays if 2 defaults: 0.99
2. pays if 1 default: .8
Whats the problem? We assumed independence.
In credit, the main risk was correlation of default. This is not
really computable.
(1)
(2)
Consider two processes Wt , Wt and the hitting times (1)
and (2) try and figure the probability P(z(1) > T ; z(2) > T ),
well, its hard to quantify. If
3.6 Systemic Risk
Usually we will consider the log-capitalization of Banks, and in
particular multiple banks:
dXti = itdt + idWti.
3.6.1

(3.29)

Toy Model

We can consider the drift of capitalization, which includes some

coupling of the systems.
p
a X j
(Xt Xti)dt + i(dWt + 1 2dWti) (3.30)
dXti =
N
j6=i

where the Wti are independent. What is the intuition here?

That borrowing and lending goes on between the banks. This
is related to flocking and swarming models.
What other mathematical model behaves this way(where we
have randomness but also attraction)? ornstein uhlenbeck process:
dYt = a(m yt)dt + dWt,

(3.31)

which we know how to solve:

Yt = m + (y m)eat + eat

easdWp

(3.32)

0
2

And we know that this is N (m, 2a ).

Next step is:
p
N
N
1 X i
1 2 X
0
d(
Xt ) = 0dt + dWt +
dWti. (3.33)
N i=1
N
i=1
Now consider if X0i = xi0 = 0, which gives
p
N
N
1 X i
1 2 X i
0
x
t =
Xt = Wt +
Wt
N i=1
N
i=1

(3.34)

(3.35)

Chapter 4

Systemic Risk Day 3

Another one: N with = 0. We must be very careful

heretaking N completely misses the systemic risk. That
is because Xt 0. But we could consider N but then
reconsider with N large but finite later. In this case, the (Xti)
(4.1) become OUs, which are independent. Then

We will continue considering the equation governing the

log-capitalization of the banks
dXti =

N
p
a X j
(xt xit)dt + (dWt0 + 1 2dWti)
N j=1

dXti = aXti + dWti

where W 0, ..., W N are independent brownian motions and a

represents the speed of trading. We can further consider the
mean over all these,
N
X
t = 1
X
Xi
N i=1 t

and we need to find the hitting time of an OU(which is annoying,

but possible).
4.1

(4.2)

Which is equivalent to [ed: is this true?]

r
1 2
= 2
dBt
N

Garnier - Papanicolaou - Wang

Here, we replace the hitting times with a double well potential.

That is, we create a double-well potential: V (x) = ax4 + bx2
(with a > 0, b < 0). Then the equation becomes

(4.3)

where the mean is governed by

p
N
1 2 X
0

dXt = dWt +
dWti
N
i=1

Other models

4.1.1

which lets us write as

t Xti)dt + Noise
dXti = a(X

(4.11)

dXti = hV 0(Xti)dt + a

N
X
(Xti Xtj )dt + dWti

(4.12)

j=1

(4.4)
4.2

Games

Work done with Carmona and Sun. We want the bank to be

able to do something. To start, consider
(4.5)

dXti =

N
a X j
(X Xti)dt + ti dt + Noise
N j=1 t

(4.13)
This gives rise to a flocking behaviorall the banks will follow
the meanat least as long as the mean is large.
Now to prevent the bank from borrowing infinite money, we
Consider the case where = 0 which gives
must introduce a cost:

dBt.
(4.6)
T
X
N
(it)2
Ji =
dt
(4.14)
2
We will count a bank as defaulting when it hits some default
0
amount D < 0. We can consider the event


where we would of course minimize J i. But now of course we

min Xt < D
(4.7) must include an incentiveotherwise, we would just never borrow.
0tT
So we modify to
And in particular, we can calculate the probability of default,

D2 N
D N

P( min Xt < D) = P( t) = 2N ( ) e 2 T (4.8)

0tT
T

Ji =

and

(4.9)
Ji =

D
P(
t) = 2N ( )
T

But this actually isnt too interesting.

(4.15)

But we would also like a convex optimization problem, so we

change to

Another interesting regime: = 0. Then

t Wt0
X

T
X
(t)2
t Xti))dt
( i qti (X
2
0

(4.10)
8

T
X
(t)2
t Xti) +  (X
t Xti)2)dt
( i qti (X
2
2
0

(4.16)

(We can think of this as a regulator imposing some extra cost if

we get too far from the mean.) Finally, we would like to include

4.2. GAMES

some other cost at the very end, and also take the expectation,
Lets solve it:
which gives:
( q2)(exp((T t)) 1) C(+ exp((T t)) )
( T
t = )
X (t)2
( exp((T t)) +) C(1 N12 )(exp((T t)) 1)
t X i) +  (X
t X i)2)dt + c (X
T X i )2
Ji = E
( i qti (X
t
t
T
(4.28)
2
2
2
0
where
(4.17)
We can not take  too small, we must have  q2.
=+
(4.29)

This looks like a Mean Field Game(MFG), pioneered

(4.30)
= (a + q) R
Lions-Lasry for the N case. This is interestingwe pass
to the limit to get an easier problem. Another approach is from
1
R =(a + q)2 + (1 2 )( q2) > 0
(4.31)
Carmona-Delarue. But it turns out that we can solve this game
N
for N finite, via a dynamic programming approach.
Where are we at here? First, we have
We start with ti = i(t, Xt). The dynamic programming
approach means also introducing the quantities
1
Xti)
!

ti = (q + (1 + )t)(X
(4.32)
Z T
N
i
V (t, x) = min Ex
as above + (terminal cond)|Ft
and also
t
(4.18)
N
a X j
This is governed by the Hamilton-Jacobi-Bellman (?) equation:
(X Xti)dt + (above...)dt + Noise
(4.33)
dXti =
N j=1 t

Z
N

X

(i)2
2 X X 2

i
i
T V +
(a(
x xj ) + i)xj V i +
+ (1 )2[ed]
qhere.
(
x xi) + (
x xi)2 = 0
j,k I+missed
one

2 j
2
2
j=1
k
Could do T because the terminal time is annoying. One
(4.19) option is to take c = 0, and optimizing T1 JTi . Taking T
[Ed: there was a random xj ,xk V i floating around here.. I think mean the effective rate of borrowing and lending is
it goes after the jk ) part?] Also jk is the Kronecker delta.
( q2)
Because we want the infimum, we just take the gradient. And
(4.34)
i
we want it with respect to so

xi V i+iq(
xi) = 0 =
i = xi V i+q(
xi),

i = 1, ..,We
N could also take N , but this will be on Friday.
(4.20) Tomorrow will be a different approach and compare the two.
But this assumes that we know V i, which we do not! So next
we consider
N
X



1
1
2 X X  2
tV i+
(a + q)(
x xj ) xj V j xj V i+
+ (1 2)jk xj xk V i+ (q2)(
xxi)2+ (xi V i)2
2
2
2
j=1
j
k

(4.21)
Well try an Ansatz here:
t
(
x xi)2 + t
2

(4.22)

V i(T, x) T = C, T = 0

(4.23)

V i(t, x) =
and impose that

which gives
xj V i = t(
and
xj ,xk Vti = t(

1
ij )(
x xi)
N

1
1
ij )( ik )
N
N

(4.24)

(4.25)

which gives
0t =2(a + q)t + (1
0t = 2(1 2)(1

1
)2 ( q2),
N2 t
1
)t,
N

T = 0

which are x-independent terms! Hooray!

T = C
(4.26)
(4.27)

Chapter 5

Systemic Risk Day 4

Consider

So we will consider

Z T
=
(t)2

inf E
qt(mt Xt) + (mt Xt)2 dt+(otherstuff)
2
2
0
for i = 1, .., N. One way to handle this is the open-loop
(5.11)
feedback approach, which introduces
Then
(Z
) the Hamiltonian is
T
X
i
c
2

( ) qi(
Ji = E
xt xit) +
(
xt xit)2)dt + (
xT xt)2
H
=
[]
y
+
q(mt x) + (t x)2
(5.12)
2
2
0
2
2
(5.2)
Another way is the dynamic programming approach where and again well set up the hamiltonians
we introduce V i(t, x). A more different way is the closed
H
loop approach due to Pontugigin (spelling?). This is also a
dt + (noise)
X0 = x0
(5.13)
dXt =
Forward-Backward DE approach. It works by introducing a
y
Hamiltonian for each player i:
H
dt + (noise)
YT = (0orT C)
(5.14)
dYt =
x
N
X
1

(5.15)
H i(~x,~yi, ~) =
((
xxk )+k )yik + (i)2qi(
xxi)+ (
xxi)2
2
2
k=1
(5.3) Now then, because E(mt Xt) = 0, we have E(Yt) = 0. Using
k
k
Then imposing that for k =
6 i, we have = (t, x). All of this all of that, we can find E(Xt) = mt (no surprise) Then we get
some equation like
becomes the stochastic differential equations
t = tXtdt
(5.16)
H i
dXi =
dt
(5.4) (but this equation is wrong, there is a q somewhere)
y
Two references:
1
H
i,j
i,j
i
tY = C(
1. Carmona-Delarue
xT xT )( ij )
dYt = j dt + ZtdW
T
x
N
2. Mean Field Games: Lions-Cassry Cassy
(5.5)
dXti

a(
xt xit)dt + (dWt0

p
+ 1 2dWti) + idt (5.1)

So now we must calculate these Hamiltonian parts:


N 
1
1
1
H i X
=
(a
+
q)(

)
yik qi( ij )+(
xxi)( ij )
kj
j
x
N
N
N
k=1
(5.6)
and
"N
#
X
X
1
1
ij
ik
i 1
i
dYt =
(a + y)( kj )Yt q ( ij ) +
(
xt xt)( ij ) dt+ZtdWt
N
N
N
k=1
(5.7)
To solve, we take the anstaz
Ytij = t(

1
ij )(
xt xit)
N

(5.8)

Plug that in, then we get some expression for dYtij . Eventually,
we get the differential equation from
t0 = 2(a + q)t + (1

1 2
) ( q2),
n2 t

T = C

(5.9)

Okay, lets switch gears a bit and consider N . Then

t mt is related to E(Xti/w0) Then the one-player model
X
ends up being
p
dXt = a(mt Xt)dt + tdt + (dWt0 + 1 2dWt) (5.10)
10

Chapter 6

Systemic Risk Day 5

???

6.1

Diversification vs Systemic Risk

Now consider a bound [, M]. We are interested in the hitting

time. We can caluclate
Z t
p
u(Yt ) = u(y) +
u0(ys) 2YsdWs
(6.11)

This will be an endogenous approach.

Consider two stocks S 1, S 2 which are geometric brownian
motions. Suppose they have mean growth and random part
0
and independent W 1, W 2. Well consider two banks with initial
wealth X01 = X02 = 1. Consider the equations:
The variance is given by


Z t
dXt1 =Xt1dt + Xt1 (1 )dWt1 + dWt2
(6.1)
2
2


E((u(y))
)
=
u(y)
+
E
u0(Ys)2 22Ysds
dXt2 =Xt2dt + Xt2 dWt1 + (1 )dWt2
(6.2)
0
Now consider T > 0 with 1, 2 and default levels D1, D2. Then we have that
Then
E() < + = < almost surely
U (t, x1, x2) = P (1 > t, 2 > t)
(6.3)
Then we can do
Then the probability of a systemic event is
P (Systemic Event) = 1 + u u1, u2,.

u(y) = E(u( )) = u()P () + u(M)P (M)

(6.4)

(6.5)

n p
n
X
X
dSt = (
i)dt + 2
XtdWti

Then it is clear that there is some critical value c.

i=1
i=1
Switching gears a bit, it is clear that u(t, x1, x2) satisfies a

PDE:
Equation: D: 2 StdBt.
tu = 0 + T.C. + B.C.
(6.6)
Then we have
!
n
n q
(Insert plot: horizontal: (0, T ) , x1 from some value to infinity, etc.)
X
X
xitdWti
dS
=

dt
+
0
+
2
t
i
6.2 T. Ichibu Paper
i=1

This is known as Fellers Diffusion (CIR) model:

dXti = idt +

Which is equivalent to

n
X

q
(xjt xit)pij (Xt)dt + 2 XtidWti

(6.13)

(6.14)

Now u() and P () 0 as n .

Next, if i > 0 and Pij bounded, regular enoguh, then there
exists a solution
Pn (According to Bass-Perkins 2003).
We St = i=1 Xti, and Pij = Pji, and we have the equation

There is a joint distribution of 1, 2. Then we can find a

plot: Put P (systemic) vs (the diversification parameter) The
probability looks something like
(1 u1)(1 u2)

(6.12)

(6.7)

j=1

(6.15)

(6.16)

i=1

StdBt. Also, define

=

n
X

(6.17)

i=1

with Xt0 is the capitalization.

This gives rise to a square Bessel process: 2; then

P ( < +) = 0. If = 2 still P ( < +) = 0 but also
P (lim supt St = ) = 1 and P (inf t St = 0) = 1. Next, consider
If we have
p
dY t = (m Yt)dt + 2YtdW t
(6.8) 0 < < 2; Then P ( < +) = 1 via a reflecting principle.
Finally, if = 0, then 0 is absorbing.
then if Y0 = 0 and if 2 m (with m > 0) then Yt > 0t.
(That is, the force of the process outweighs the volatility.) The 6.3 Multiple Defaults
nice part of this is that it never hits zero.
If for (`1, ..., `k ) {1, 2, ..., n},
We will need to solve an equation Lu = 0 of the form
!
h
X
1
0
2
00
(m y)u + yu = 0
(6.9)
sup |x`i xj | p`i ,j (x) < 2c0 =
2
`i (6.18)
h(h 1)
x
i=1
which has a solution of the form
then we have that
Z y
6.2.1

Quick review

e 2 z 2 dz

u(y) =
1

(6.10)
11

P (t < +s.t.Xt`1 = ... = Xt`h = 0) = 1

(6.19)

12

CHAPTER 6. SYSTEMIC RISK DAY 5

Can also show the converse.

We can show Pij = n1 . We can then consider the mean field
model:
q


i
i
dXt = i + m xt dt + 2 XtidWti
(6.20)
where

1X i
X0
n n
i=1
lim

(6.21)

Next, we could consider the health of a group of banks:

n X
k
X

Xti Xtk pk,j (Xt)

(6.22)

j=1 i=1

We could also create networks: Create a link between i and

j if there is a certain amount of flow. Now we have a graph by
looking at these quantities. Then we can start asking questions
about the size of the system, path lengths between banks,
statistics of the system, etc.

???

Chapter 7

Rama Cont

Bank balance sheets:

1. Assets
(a) Liquid Assets
(b) Interbank claims
(c) Other assets
2. Liabilities
(a) Deposits
(b) Interbank liabilities
(c) Short-term debt
(d) Capital
Key parts: Solvency vs Liquidity.
7.1

Microprudential approach

Traditional approach to risk management and bank regulation is to

focus on failure/non-failure solvency, liquidity) of individual banks.
It focuses on the balance sheet structure of the individual banks.
They assume that losses arise due to exogenous random
fluctuations in risk factors. The main tool for stabilization
of system is the capital requirements. But it ignores links or
interactions between market participants.
7.2

Liability Chain

Due to Hellwig. Consider a finite set of banks i = 1..N, where

1. i borrows $100 from i 1 over i years
2. i lends $100 to i + 1 over i + 1 years

13

Chapter 8

Contagion and systemic risk in financial networks

Rama Cont

Exposure networks: Let Eij denote exposure of i to j.

14

Part II

Commodities and Energy

15

Chapter 9

Commodities & Energy Markets 1 - Coulon

What are commodities? Commodities are goods (either

Forward curve behavior(Samuelson effect): Forward contracts
natural resources or processed) with little or no variation in become more volatile as they become closer to expiration.
quality across supply sources.
Cost of carry relationship: For a financial asset with no storage
cost and interest rates r constant, by a simple no-arbitrage
They are widely traded.
argument:
Recent changes:
F (t, T ) = Ster(T t)
(9.1)
1. gas/electricity deregulated
2. number of market participants has rapidly expanded
But since storing commodities is expensive:
3. banks actively trade derivative products
4. investment volume
F (t, T ) = Ste(r+c)(T t)
(9.2)
Main participants in price models: Traditionally:
However, since arbitrage argument only holds in one direction,
1. Producers (farmers, mines, power plants, refineries)
so instead
2. Consumers (large industrials, utilities, airlines)
F (t, T ) Ste(r+c)(T t)
(9.3)
3. Storage/Delivery (gas pipelines, logistics companies)
Why? We cannot actually short physical quantities. (You cant
Recently:
borrow from future harvests, for example.)
1. Financial institutions, speculators, investors, regulators
2. everyone else (if you read the news, drive a car, eat food, etc) 9.0.1 Theory of storage
Introduce
What are the main modeling challenges:
F (t, T ) = Ste(r+c)(T t)
(9.4)
1. Adapt financial math to very different markets
2. Capturing features of price dynamics and fundamen- where is the convenience yield. Very artificial, but still some
intuitions. Basically, we get some dividend from being the holder
tals(inventory levels, weather )
3. Strike balance between supply/demand, calibration, pricing of the asset. Frequently roll the costs into the convenience yield.
derivatives, etc.
9.0.2 Reduced-Form Model
Differences:

Simplest spot price model consistent with cost-of-carry:

1. Exhibit mean reversion, possibly seasonal, very high volatility

F = Ser (T t)
2. Dynamics closely linked to economic factors
3. Spot commodities are not traded assets in the same sense
as stocks or options (due to storage/delivery issues) (mainly, We get from the Generalized Brownian Motion as
cant assume no arbitrage)
dSt = (r )Stdt + StdWt
4. Critically: Same commodity delivered at two slightly different
location or times can behave as separate assets.
under Q, where r, , and are constant. Then
5. Set of liquidly traded products often significantly smaller than
the risks needed to be hedged( specialness )
F (t, T ) = EQ
t [St ]
6. Correlations are very important: companies interested in
hedging multi-commodity exposure.
which gives
Spot power prices: craziest of all.
h
i
1
Spot price history: more regular.
(wT w0 )
F (t, T ) = Ste(r 2)(T t)EQ
t e
Categories of commodities:
2
1.
2.
3.
4.

Storable vs non-storable
continuous vs seasonal production (or demand)
local vs regional vs global
elasticity of supply or demand to price

so

F (t, T ) = Ste(r)(T t)

(9.5)

(9.6)

(9.7)

(9.8)
(9.9)

What are the weaknesses here?

1. We cannot obtain both contango and backwardation. A
Modeling categories:
natural extension is to let t be stochastics. This is good
1. Reduced-form: direct price modeling, traditional financial math
because we are capturing the relationship with inventory
2. econometic: reduced form elements combined with regressions
(inversely), but t is unobservable
to find relationship between price and key factors
2. No mean reversion in spot price, so all forwards have the same
3. structural: capture key features of supply and demand and
volatility. (Can see this by applying Ito to F (t, T ), giving
approximate market mechanisms while retaining tractability
4. full equilibrium: detailed matching of supply and demand, opF
F
1 2F 2
dFt =
dt +
dS +
dS
(9.10)
timization over participants behavior, physical constraints, etc.
t
S
2 S 2
16

17
which implies
dFt
= dWt
Ft

(9.11)

which has no T dependence, which means no Samuelson effect.

An alternative: Schwarz one-factor (97):
dSt = ( log(St))Stdt + StdWt

(9.12)

under Q. If we apply Ito to yt = log St, then

St = eyt

(9.13)

where yt is ornstein uhlenbeck process

dYt = (

2
yt)dt + dWt
2

(9.14)

which is an exponential ornstein uhlenbeck process. It gives

2
(1e2(T t)))
2
(9.15)
We can apply Ito to ytet. So it is easy to find F (t, T ):
log St N((log S0)e(T t)+
(1e(T t)),

1
F (t, T ) = EQ
t [St ] = exp{mean + variance}
2

(9.16)

Then apply Ito to Ft to see

dF (t, T )
= e(T t)dWt
F (t, T )

(9.17)

Weaknesses: Volatility of long forwards is underestimated.

Also all forwards are perfectly correlated.

Chapter 10

Commodities & Energy Markets 2 - Coulon

10.1

Reduced form models

10.2

Two Factor Models

So far, two simple one-factor spot price model. The two main These try to correct some of the weaknesses of the volatility
structures, etc. See Schwortz (97) for two factors. We need short
approaches are
term shocks which are not mean reverting but long term which is.
1. Generalized Brownian Motion (no mean reversion)
dSt =(r t)Stdt + StdWt
(10.5)
dSt
dF (t, T )

= (r )dt + dWt =
= dWt (10.1)
dt =( t)dt +
dWt
(10.6)
St
f(t, T )

dWtdWt =dt
(10.7)
2. Exponential OU (mean reverting)
The advantage is that this is a combination of mean reverting
and non-mean reverting(which is basically just saying short term
dSt
dF (t, T )
= (log St)dt+dWt =
= e(T t)dWt and long term.)
St
F (t, T )
(10.2) 10.2.1 Schwartz & Smith (00)
Here, the coefficients of the stochastic differential equations Instead:
St = exp{Xt + Yt}
(10.8)
for F (t, T ) describe the vol term structure.
The key observation here is that:
where Xt is arithmetic brownian motion and Yt is O.U.
Both models are lognormal and find find
mean reversion in St decaying vol of F (t, T ). (10.3)
F (t, T ) = EtQ[St]
(10.9)
Picture goes here: time is horizontal axis. GBM is a constant y
value, exponential OU shows a decay. Typical data ( either implied We can show the equivalence with Schwartz (97).
or historical) show somewhat slower decay of the forward curves.
10.2.2 Extension: 3 factor
Notes:
1. Why the mean reversion of the spot prices? We can make We can also add a Vasicek model for rt. But this one doesnt
economic arguments about short term shocks vs long term really matter because interest rates have only minuscule amounts
of volatility relative to the volatility of the commodity.
equilibrium (production/consumption levels).
2. Why not mean reversion in F (t, T ) itself? Its a traded
derivative, it must satisfy the martingale condition. (Forwards
are martingales, at least for fixed T )
3. What about F (t, t+) (fixed tau)? For a start, its not a traded
contract. Therefore, it is likely that it could be mean reverting.
And of course, as 0, we should get the spot price.
Any exceptions? In theory, we can have something that mean
reverts under P but not under Q. But this is quite unusual or artificial. Why? Drift under P with a mean reverting process gives a
market price at risk . That is, drift under P to drift under Q gives

10.3

Calibration

The first priority is typically to match the observed forward

curve. (That is, find , , ). But of course, we cant do this.
So we need to let be time dependent.
We take the Hull-White/Ho-Lee approach: Now we try to take
{, , (t)}. We can either take continuous time or some number
of piecewise linear intervals to get the right number of parameters.
One thing to be careful of: Can easily overfit the model
because of the amount of freedom in this calibration. One thing
to look out for is if the model calibrations come out completely
differently each day; we should exercise caution.

Yt)dt

(10.4) 10.4 Forward Curve Models

Instead, we can start with the forward curve directly (especially
where is the market price of risk. We could actually choose if we do not care about St). The general model under Q is:
to be a function of Yt to kill mean reversion, but again, this
N
is very artificial.
dF (t, T ) X
(i)
=
i(t)dWt
(10.10)
P
Related question: How does Et [St] compare to F (t, T )?
F (t, T )
i=1
Clearly, it depends on the sign of ;
( Yt)dt (

1. Normal backwardation: > 0 (this is hedging pressure where W (i) are independent brownian motions for simplicity.
t
from risk producers pushing F (t, T ) downwards). See Now we must choose:
Inconvenience Yield and The Theory of Normal Contango
- Bouchuev, 2012 for an argument that recent change toward 1. The number of factors N
2. Shape of 1(t, T ), ..., N (t, T ).
contango change driven by contango.
2.
Notes on this:
18

10.4. FORWARD CURVE MODELS

19

1. All of the previous St models are special cases of this. IE:if

N = 1, 1(t, T ) = e(T t) gives the Schwarz 1-factor
2. Calibration to F (0, T ) T is immediate, because it is just the
initial condition for the stochastic differential equationthere
is no calibration.
3. Can show that F (t, T ) is lognormal. Apply Ito to log F (t, T ),
which implies

Z t
Z
N 
X
1 t 2
i
i(u, t)dWu }
(u, T )du +
F (t, T ) = F (0, T ) exp{

2 0 i
0
i=1
(10.11)
So we went from the St dynamics to the F (t, T ) dynamics.
Can we go from the F (t, T ) dynamics to the St dynamics?
Recall that St = F (t, t), so
log St = log(F (0, t) +

n
X

[...])

(10.12)

i=1

If we apply Ito again to log St and do the calculations, then we

have
# N
"
Z t
N Z
X
i(u, t)
i(u, t)
log F (0, t) X t
dSt
(i)
(i)

{ i(u, t)
du +
dWu }dt +
i(t, t)dWt
=
St
t
t
t
0
0
i=1
i=1
(10.13)
which actually implies that St is non-Markovian! (Bad news!)
Unless you have just the right form that the terms cancel.
How can we estimate the volatility functions? 1) From
historical returns, we do Principal Component Analysis. 2) From
options directly. People usually propose a particular parametric
form. One option (which Glen probably uses):
N

dF (t, T ) X
(i)
=
vi(t)i(T )ei (T t)dWt
F (t, T )
i=1

(10.14)

where the termsvi(t), i(T ), and ei (T t) represent the affect

of the whole curve, maturity specific, and steepness of the vol
term structure. (That is, a Gaussian Exponential Factor Model.)

Chapter 11

Commodities - Day 3

Recall the general forward curve model

11.2

Spread Options

General spread option payoff at time T has the form

dF (t, T ) X
(i)
=
i(t, T )dWt
F (t, T )
i=1

(11.1)

(aXT bYT K)+

(11.10)

where XT and YT are different commodity prices (spot or

forward):
Most commodity markets options traded on futures, not spot,
and maturity just before the futures maturity. That gives three 1. Input/Output (e.g. dark if between electricity/coal, spark)
2. Input/Output (e.g. crack if between refined product, crude)
critical times:
3. Calendar (e.g. Dec13 Forward vs Jun13 forward) (storage)
1. t: current time
4. Locational (e.g. Henry Hub vs NorthEast gas) (transport)
2. T1: option maturity
Spread options are critically important, due to the strong link
3. T2: forward maturity
with physical assetsthey are useful as hedging and valuation
The price we want is the
tools. There are some tricky aspects here, because these reflect

optimal(unconstrained) operation, which is not usually available.
Q
+
Vt = exp(r(T1 t))Et (F (T1, T2) K)
(11.2)
11.2.1 Classical Spread Option Pricing
We can use Blacks Formula on any lognormal E [(ex K)+] Margrabe proposed:
and x N(x, x2 ). Many people are familiar doing something
like
(1)
(1)
(1)
X
dSt =rSt dt + 1St dWt
(11.11)
1 2
1
=
(ex +x x K) e 2 x dx
(11.3)
(2)
(2)
(2)
2
dSt =rSt dt + 2St dWt
(11.12)
log Kx
11.1

Option Pricing

(1)

(2)

dSt dSt =dt

(11.13)
ad after a few lines we get




which are correlated geometric brownian motions.
2
x log K
x + x2 log K
x + 12 x
=e
K
(11.4)

11.2.2 Power Plants

x
x
We can approximate a power plant value via a string of spread
But we have that
options
1
X


x = EQ[log(F (T1, T2))] = log(F (t, T2)) x2
(11.5)
Plant Value
exp(rTj )EQ (PTj hg GTj eg ATj K)+
2
1

EtQ [F (T1, T2)] = F (t, T2) = ex + 2 x

jJ

(11.6)

(11.14)
The main challenge is to capture the multi-commodity dependence
structure and the link with demand in a mathematically tractable
r(T1 t)
Vt = e
[F (t, T2)(d+) k(d)]
(11.7) model.
Main approaches:
2
log(F (t,T2 )/K)x
Q
2
where d =
and

=
Var
[log
F
(T
,
T
)].
1.
Reduced-Form: Correlated Lognormals, etc (Carmona &
1 2
x
x
Durrleman)
Fine. Now the general forward model is
2. Full Fundamental: Via production cost optimization problem
N Z
N Z T1 3. Structurally: Embedded into a model for spot power:
X
1 X T1 2
log F (T1, T2) N (log F (t, T2)
i (u, T2)du,
i2(u, T2)du)
2 i=1 t
t
Power = f(gas, coal, carbon, ...)
(11.15)
i=1
(11.8)
Some issues using Margrabe for power valuation:
We just need to integrate the squared forward vol over the life
of the option. For example, the Schwartz 1-factor model gives 1. prices negative
Z T1
 2. not brownian motion
2  2(T2 T1 )
3. etc...
x2 =
(e(T2 u))2du =
e
e2(T2 t)
2
t
(11.9) 11.2.3 Electricity Markets
Plot vs maturity: Exponential decay from T2 T1 until T2 t. Main point: Electricity is not storable, so we require hourly
(Or, alternatively, integrating the exponential increase from t matching of supply and demand for market clearing. The price
to T1 (and avoiding T1 to T2)).
varies widely across different locations and is
20
Hence, we have

11.2. SPREAD OPTIONS

Huge huge huge spikes. The prices might even go negative.
Some pretty weird markets, because of subsidies, etc.
Well focus on the correlation between natural gas and
electricity.
11.2.4

Structural Models for Power

Were shooting for a middle ground which lets us use forwardlooking models and soforth. Also, using historical data is
problematic.
1. Demand Dt: Barlow (2002)
2. Capacity t: Burger et all(2004), Cartea et al (2007)
3. Fuel Prices Gt: Pirrong and Jermakyan(2005) , Aid (2011)
11.2.5

Bid Stack function

Day-ahead generator bids arranged by price to form the bid

stack. The spot price Pt is highest bid needed to match inelastic
demand Dt. The merit order of production costs drives the
dynamics of the stack.
We can model some of these ideas which capture
We can get some closed form solutions via the assumption
of an exponential bid stack.
A better way is to look at the power price vs gas price curves.

21

Chapter 12

Commodities & Energy Markets 4

Inventory and Implications for Valuation : Glen Swindle

12.1

Main Themes

1. Liquidity vs Dimensionality: Liquidity is concentrated in

benchmarks. Risk can be spread over hundreds of locations
and many delivery months. Price decoupling (specialness)
occurs routinely. (Dodd-Frank)
2. Covariance Structures vs Options Markets: The term structure
of volatility and correlation is nontrivial. Liquidity in options
is concentrated in vanilla products with mechanics with limit
information about (limit utility as hedges of)
3. Broad ranges of time scales: Price dynamics exhibits structure
over years (infrared) to hours (ultraviolet)
4. Uncommoditized risks
Carry formalism The forward curves can be viewed as yield
curves. We can do the forward yield via


1
f(t, T + S)
y(t, T, T + S) = log
(12.1)
S
f(T, T )
Key points:
1. All the you can get from market data is q
2. The cost of storage is not exogenous. Storage owners will
charge what the market will bear.
One approach: Fit a fourier series, then look at the residual.
Storage: Types of storage: Aquifer storage, reservoir, salt
caverns, Actually 8.7 tcf of storage, but 4.2 tcf is base gascant
get it back out.
Dynamic Optimization:
Z
V [0, S(0), F (0, )] = sup E{
s()A

d(0, t) [s(t)F (t, t) (s(t), S(t), F (t, t))] dt}

(12.2)
1. d and F are discount factor
2. S is the current inventory level, s = S 0
3. denotes costs associated with injection and withdrawal. Eg:
= k|s(t)|F (t, t)
4. A denotes allowed controls
This is a super hard problem.
Lets figure out a calendar spread option: time spread option.
Consider options with the following payoff


max F (, T + U) + erU F (, T ), 0
(12.3)

22

Chapter 13

Commodities & Energy : Day 5

Non Standard Expiry

Vanilla options markets in commodities always have expiration

dates very near the delivery. Options structures arise in which
expiration varies substantially from the standard market convention. Modeling/estimation is required to infer implied vols over
nonstandard time interval from vanilla implied vols. Examples:
1. CSOs
2. Price holds (one type: Swaptions)
3. Options on ETFs
(a) These involve the vol of the current nearby, which requires
inference from term vols for each contract to the vol pertinent to the time when the contract is active in the index.
(b) Recursive dynamics of Vt is due to the contract rolls:

N(t)
Y F (Tn, Tn)
F (t, Tn+1)
Vt =
(13.1)
F
(T
,
T
)
n
n+1
n=1
Working problem: On 18Feb2009 you are asked by the sales
desk to price WTI Dec11 at-the-money European straddles with
expiration on 17Dec2009. This is the simplest incarnation of
nonstandard expiration. Most occurences are much more complex.
What you know: Dec11 is a liquid contract with standard
options expiration date 16Nov2011. NYMEX options market
are visible on this horizon.
13.0.1

Non-Stationarity

Empirical analysis of price dynamics in commodities constantly

is encumbered by a technical annoyance: contracts expire.
Two basic approaches:
1. Construct rolling nearby series;
2. construct constant-maturity forwards.
Two methods:
1. Nearby contracts refer to a sequence of live contracts (1st
nearby is first, second nearby is next live contract)
2. Constant maturity.
Keypoint: Do analysis on concatenated RETURNS not prices.
13.1

Volatility Backwardation

Volatility backwardates.

23

Part III

Portfolio Optimization,
Transaction Costs, Dynamic Games

24

Chapter 14

Portfolio Optimization, Transaction Costs & Dynamic Games 1 - Muhle-Karbe

1.
2.
3.
4.
5.

Setting
Frictionless Control Heuristics (no transaction costs)
Verification via Convex Duality
Control Heuristics with Transaction Costs
Verification via Shadow Prices

One messiness here is this: The units are goofy. An alternative

is to use a different unit. A better interpretation is: maximize
certainty equivalent:
CE() = U 1(E [U(XT)]) max

(14.8)

Well consider the Black-Merton-Scholes model with two assets:

St0

1. Safe asset normalized to

=1
2. Risky asset following geometric Brownian Motion:
dSt
= dt + dWt
St

How can we determine optimal strategy and utility? Theres a

complex dependence on market parameters, investment horizon,
and preferences! How can we tackle an infinite dimensional
stochastic control problem? Today, well do a heuristic derivation
and tomorrow, well do a rigorous verification.
We first make the problem harder: via a value function like
(14.1)
"
#
Z
T

for a standard Brownian motion (Wt)t0 defined on a filtered

v(t, x) := sup E U(x +
sdSs)|Ft
(14.9)
(s )s[t,T ]
probability space (, F, (Ft)t0, P ).
t
3. Returns have constant expectations > 0, vol > 0
4. simplest benchmark
, which evaluates along any wealth process Xt (a supermartingale
5. no tx costs
with nonpositive drift) and along optimal wealth Xt (martingale
We start with a positive cash endowment x > 0. An investor c with zero drift.) This boils down to: Martingale optimality
hooses number t of risky shares. Corresponding wealth process: principle of stochastic control.
What do we mean by all of this? We start with
Z t
E [v(t, Xt)|Fs]
(14.10)
Xt = x +
sdSs
(14.2)
" "
# #
0
Z T
=E E U(Xt +

sdSs)|Ft |Fs
(14.11)
The safe position implicitly determined by self-financing condition:
t

t0 = x +

"

sdSs tSt

(14.3)

=E

U(Xt

Z
+

sdSs)|Fs

To make everything well-defined, (t)t0 needs to be S-integrable:

"
=E

U(Xs

t 0

(14.4)

"
E

U(Xs

Z
+

sdSs) +
s

ssds <

(14.12)

sdSs)|Fs

(14.13)

sdSs)|Fs

(14.14)

Preferences: Investor maximize expected utility from terminal

wealth at some time horizon T > 0:
E [U(XT)] max

=v(s, x
s)

(14.15)

How can we use the martingale optimality princeiple? The

(14.5) value function v(t, x) generally can depend on many state
variables. Here, the wealth dynamics

over all trading strategies t.

The utility function U should:

dXt = tStdt + tStdWt

=tdt + tdWt

(14.16)
(14.17)

1. be increasing: more is better than less.

2. Concave: risky payoff is worth than its expectation
3. Inada conditions: Extra dollar matters a lot when poor, for some risky position t only depend on the control t. One
guess is a value function only depends on the initial wealth x
irrelevant when rich. (Codify in math by:
and time t. This is an assumption!! Not a result. This needs
to be verified eventually.
0
lim U (x) 0
(14.6)
Suppose that the value function is smooth. Then Itos formula
x
lim U 0(x)
(14.7) yields
x0

25

dv(t, Xt ) = (vt + tvx +

2 2
vxx)dt + (...)dWt
2 t

(14.18)

26 CHAPTER 14. PORTFOLIO OPTIMIZATION, TRANSACTION COSTS & DYNAMIC GAMES 1 - MUHLE-KARBE
This should be a supermartingale for any t: drift 0. It will
also be a martingale for the optimizer t: with drift = 0. Then
maximize drift pointwise in t. Plug in maximizer. Set to zero.
This yields equations for optimal stategry and value functions.
We can consider a special case: Take the exponential
utility function U(x) = eax, and factor the wealth out via
v(t, x) = eaxv(t, 0), with the constant absolute risk-tolerance
1/a. This gives the constant risky position
t =

a2

(14.19)

This is independent of horizon. It is independent of wealth. The

value function and certainty-equivalent are the same.
2
T
2a2
(14.20)
Here, the assets are ranked by the Sharpe ratio /. The same
Sharpe ratio leads to the same payoffs.
An alternative: If we have twice as much money, we are willing
to take twice as much risk. This is debatable, but why not??
v(t, x) = eax exp(

22(t T )

U(x) =

) = CE(
) = x +

x1
1

(14.21)

rederive v(t,x), risk tolerance, proportion, value function, certainty

equivalent.
v(t, x) = x1 v(t, 1)
(14.22)
Constant relative risk-tolerance x/. Constant risky proportion
t := t/Xt = /2, which is independent of horizon.
Investment scales with wealth.
Value function and certainty equivalent:

Chapter 15

Lecture 2

We can verify that all of this actually works.

15.1

15.2

Verification via Convex Duality

Need to rule out doubling strategies. Usual notion: Xt C

(we have a bounded credit line). The value C = 0 works well for
utilities defined on R+; optimal wealth process for power utility
follows geometric Brownian motion:
dXt

(dt + dWt)
=
2
Xt

(15.1)

Markets with Transaction Costs

Before, we bought and sold at the same price St. Now, we

will buy at the ask price St and sell at the bid price (1 )St.
(clearly,  is the width of the relative bid-ask spread.) We only
consider finite variation strategies t = ut dt. These track
the number of shares bought and sold by time t, respectively.
Infinite costs otherwise.
We can do integration by parts on the frictionless self-financing
condition, which then reads as
dt0 = tdSt d(tSt) = Stdt

(15.6)

This is not compatible with utilities on R: optimal wealth pro- The counterpart with transaction costs is
cesses for exponential utility follows brownian motion with drift:
dt0 = Stdut + (1 )Stddt.
(15.7)

dXt =
(dt
+
dW
)
(15.2)
t
Why is all of this important? Non-trivial spreads are common
2
even in the most liquid markets. Constant position/weight
requires infinite turnover, which causes infinite costs, which
How can we verify?
is not feasible. How can we adapt optimal trading strategies?
15.1.1 Convex Duality
How much welfare is lost? What are the endogenous spreads
in equilibrium between market makers and rebalancing investors?
For a concave function U, we can define the conjugate:
(This is a sort of static game which can determine the bid-ask
V (y) = sup{U(x) xy}
(15.3) spread as an output of a model such as: Find a bid-ask spread by
x>0
examining market maker profit from optimal strategy of traders
given a bid-ask spread.) There is also a new difficulty: complex
Then
horizon dependence. Do not trade if horizon is close. Way out?
Two alternatives:
E [U(XT)] E [V (yYT )] + E [XT YT ] = E [V (yYT )] + EQ [XT] y 1. We could look at the expected utility of the final value of our
(15.4)
portfolio:
for y > 0 and density process Yt of an EMM Q.
E[U(XT )]
(15.8)
R
t
Then Xt = x + 0 sdSs is a local Q-martingale.
but this is a bit fishy...
1. Supermartingale if bounded from below for utilities on R+ 2. Another option is to look at
2. Martingale if risky position is sufficiently integrable for utilities
Z
on R (eg, bounded). This implies admissibility.
E[
etU(ct)dt]
(15.9)

0
In either case, E[U(XT )] E[V (yYT )] + xy.
This upper bound is for any trading strategy , any t, and
which is good, except that it is basically intractable.
any y. Candidate is optimal if bound is tight for
So, as before, well consider the exponential utility
1
1. U 0(Xt) = yYt
U(x) = ex or power utility U(x) = x1 . But infinite
2. E[YT (U 0)1(yYT )] = x
horion to reduce stationary problem: For exponential utilities,
x
we
maximize the equivalent annuity:
For exponential utilities U(x) = e
, we get
h
i
V (y) = y log y 1. The second condition gives y =

1
lim sup
log E eXT max!
(15.10)
exp(xE[YT log YT ]). This yields a simplified upper bound:
T
T

E[eXT ] exE[Yt log YT ]

(15.5)

For power utilities, we maximize the equivalent safe rate:



1
lim sup
log E (XT)1 max!
(15.11)
In complete market with unique YT , we can verify equality
T (1 )T
for candidate by direct computation! Sweet!
We can do analogous results for power utilities
This keeps scaling properties in the wealth, and also gets rid
U(x) = x1 /(1 ).
The conjugate comes out to of the horizon dependence.

V (y) = 1
y11/ . Then we look at the expectation, and look
Open questions: What are the control heuristics? How can
we verify it?
at the upper duality bound simplifies.
27

Chapter 16

Lecture 5

[tcj: Missing lectures 3/4 because notes online.]

Idea: Directly tackle asymptotic optimality equations.
General model: General asset prices, costs, preferences. Normalize the safe asset to one. The risky asset will be traded with
proportional costs t = t > 0. The mid price will be given by
q
dSt = bSt dt + cSt dWt
(16.1)

dhit
This is determined by a type of portfolio gamma dhSi
. This
t
implies that Active strategies require a wide buffer and turbulent
markets call for closer tracking.
Note that only the current spread t matters here; future
dynamics are hedged at higher orders, but not considered here.
We can do the following calculation: Given t = (St), then
we have
dT = 0(St)dSt + (...)dt
(16.8)

This gives the general diffusive dynamics. We can include

heteroskedasticity and predictable returns leading to market and then we can consider the quadratic variation which is
timing. We do not need any markovian structure.
dhT it
Why doe we need general asset prices? Transaction costs have
= (0(St))2 = 2(St)
(16.9)
dhSi
small effect in Black-Scholes model, because investors essentially
should buy and hold, and then profit from the market growth.
On the other hand, Mnay strategies are market neutral. They
make a profit by active trading. For example: Buy low, sell high
for a mean-reverting instrument, which might be governed by
dSt = Stdt + dWt

(16.2)

or a momentum strategy, which says: go long in good times,

short in bad times. That gives:
dSt =tdt + dWt,

(16.3)

dt = tdt + dWt

(16.4)

In these cases, it is much harder to avoid the transaction costs.

We will need to find a direct tradeoff between the gains from
trading and costs incurred by trading.
Here, the investor will solve
"Z
#
T
E
U1(t, t)dt + U2(XT (, )) max
(16.5)
0

over all policies (t, t).

Motivations:
1. Individual household: Receive labour income stream, consumes
over lfetime
2. Retirement Fund: Maximize long-run expected utility from
terminal wealth
3. Option trading board: Hedge derivative until maturity.
4. Hedge Fund: time the market on short horizon.
Here, the general model incorporates all these asset prices and
optimization problems.
The optimal policy is to consume at the rate:
t = t + 0t(Xt Xt). We will trade to keep the riskyshares 
in some no-trade region NT t NTt, NT t + NTt . Clearly,
the midpoint is given


NT = t + 0t(Xt Xt)
We can determine the halfwidth to be

 31
3Rt dhit
NTt =
t
2 dhSit

(16.6)

(16.7)
28

Chapter 17

Portfolio Optimization, Transaction Costs, Dynamic Games 6

Sircar

17.1

Announcements

A plot of this in the q2 vs q1 shows two triangles, where the

hypotenuses intersect is the nash equilibrium.

Well hold extra discussion sessions as follows:

1. Tuesday 13:30: Schied
2. Wednesday 13:30: Swindle
3. Thursday 13:30: Cont
4. Thursday 15:00: Sircar
17.2
17.2.1

Definition 1 A point (q1, q2) [] is a nash equilibrium if

1(q1, q2) 1(q1, q2)q1
2(q1, q2) 2(q1, q2)q2

Dynamic Oligopoly Games

We can verify this by noting that

Motivation

1. Try to understand the evolution of energy markets

2. Competition between different fuels a many dimensional
problem.
Concerns: Exhaustibility of fossil fuels(Oil running out); Green Energy (renewables, eg Solar); Exploration and Improved technology;
One approach is to consider the Liquid Financial Markets which implies that
with a large number of small price takers. The completely
opposite extreme is having a small number of relatively large
players, all of whom are competing with eachother, and which
strongly influence prices. Later in the week well return to the
financialization of commodities markets.
17.3

Cournot Market (1838)

This is the first example of a Nash equilibrium. Consider a static,

1 period, two player game. Key points: This model was set up in
the context of producers of mineral water, which is an essentially
inexhaustible resource. We can choose quantities q1 and q2 to
bring to market. The market is governed by pricing (inverse
demand) function P : quantity
7 price, with q1 + q2 = Q. First
characteristic: This is a decreasing function! (Flooding the
market sents price to zero.) Well also assume that the goods are
perfect substitutes, but that the players have different marginal
(ie, per bottle) costs of production 0 a1, a2 1.
Later, well consider these ai to be dynamic (ie, shadow costs
or scarcity costs). For energy, we note that oil is cheap but wind
is more expensive.
To set up the competition, we have
1. Player 1 objective:
max q1(1 q1 q2 a1)
q1 0

2. Player 2 objective:

(17.5)
(17.6)

(17.1)

1
R1(q2) = (1 q2 a1)
2
1
R2(q1) = (1 q1 a2)
2

1
q1 = (1 2a1 + a2)
3
1

q2 = (1 2a1 + a1)
3

(17.7)
(17.8)

(17.9)
(17.10)

Observe that this is decreasing in your costs and increasing

in your opponents cost.
We can calculate an aggregate quantity Q = q1 + q22) =
1
We can calculate the market price
3 (2 a1 a2 ).
P (Q) = 1 Q = 13 (1 + a1 + a2).
We can compare this versus the monopoly case, where the
only player will solve
1
max q(1 q a) = q = (1 a)
q0
2

(17.11)

which implies the monopoly price PM

= 12 (1 + a).
We can also calculate the duopoly case (ie, where N = 2) and
a1 = a2 = a. Then we get PD = 13 (1+2a) < PM = 12 (1+a). Key
point: Competition decreases price, which benefits the consumer.

17.4

Bertrand (1883) Competition

Now, consider where firms set prices, not quantities.

17.5

Exhaustible Resource Problem

Hotelling (1931) did this in the monopoly case, while Dasgupta

(17.2) + Heal (1979).
q2 0
First: Start with two oil producers with finite resources
We can recast these profits as i(q1, q2). We can solve in the (x(t), y(t)). In the Cournot setting, choose the quantities
sense of a Nash equilibrium (NE) which is the intersection of (q1(t), q2(t)) with
the best response:
dx
= q1(t){x>0} + (Noise..)
(17.12)
R1(q2) = arg max 1(q1, q2)
(17.3)
dt
q1
dy
R2(q2) = arg max 2(q1, q2)
(17.4)
= q2(t){y>0} + (Noise..)
(17.13)
q2
dt
29
max q2(1 q1 q2 a2)

30

CHAPTER 17. PORTFOLIO OPTIMIZATION, TRANSACTION COSTS, DYNAMIC GAMES 6

Now, player 1 will want to solve something like

Z
ertq1(t)(1 q1(t) q2(t))dt
v(x, y) = max
q1

(17.14)

(this assumes zero cost of production), while player two considers

Z
w(x, y) = max
ertq2(t)(1 q1(t) q2(t))dt (17.15)
q2

where x(0) = x and y(0) = y.

17.6

Dynamic Programming

We can take the two points with two value functions, and
non-zero sum differential games.
17.6.1

Recipe

Introduce
L = q1

q2
x
y

(17.16)

Then
rv = max Lv + q1(1 q1 q2)

(17.17)

rw = max Lw + q2(1 q1 q2)

(17.18)

q1

and similarly,
q2

Taking this a step ahead, we have



v
v
) q2
rv = max q1(1 q1 q2
q1
x
y

(17.19)

and


w
w
rw = max q2(1 q1 q2
) q1
q2
y
x

(17.20)

If we return to the static two player, and recall the optimal

quantity
1
q1 = (1 2a1 + a2)
(17.21)
3
and profit 1(a1, a2) = q1(1 q1 q2 a1) = (a1 )2. Then the
PDE that we want to deal with is y vs x, on the domain x 0,
y 0, and we want to satisfy the equations
v
= rv
y
w
2(vx, wy ) q1(vx, wy )
= rw
x
1(vx, wy ) q2(vx, wy )

(17.22)
(17.23)

In the case of inexhaustible oil, then v = w =constant. In fact,

1
v = w = 1,2(0, 0)
r

(17.24)

We can interpret vx and wy are shadow costs or scarsity.

Next time: Well try to model what happens on x = 0 and
y = 0. (ie, what happens on the exhaustible case?) Either the
other player gets a monopoly or we must bring in renewables.

SIRCAR

Part IV

High
Frequency Trading and Limit Order Book

31

Chapter 18

High-Frequency Trading & Limit Order Book 1 - Carmona: Limit Book Orders

Black-Scholes theory. The price is given by a single number.

There is infinite liquidity. One can buy or sell any quantity
at this price with no impact on the asset price. This fixes to
account account for liquidity frictions. This doesnt account for
transactions costs, and we add some liquidity frictions, or put the
transaction cost as a proxy for liquidity. This theory will not be
satisfactory for large trades over short periods or high frequency
trading. We need to understand the market microstructure.
Several types of markets:

the question becomes, which exchange should we submit the

order to?)
6. Here, little or no discussion of pools.
18.4

Role of order book

The LOB is crucial in high frequency finance: it explains the

transaction costs. The liquidity providers post trading intentions:
bids and offers. Liquidity takers execute certain orders: Adverse
selection. We can construct a plot by showing the price in the
limit order on horizontal axis and the volume of desired shares
1. Quote driven Markets: Market makers or dealer centralizes on the vertical axis. Because of the FIFO nature we probably
buy/sell orders and provides liquidity by setting bid and ask can not get in at the highest bid or lowest ask.
quotes.
2. Order driven markets: Electronic platforms aggregate all 18.4.1 DELL
available orders in a Limit Book Order. (Eg: NYSE, Can we see a selling panic on the order book? Yes. Data is giant.
NASDAQ, LSE, etc)
Critical point is the bid/ask spread
Here, the same stock is traded on several venues. Price discovery 18.5 Limit order
is difficult due to many instruments being traded off the public
market book. The competition between markets leads to lower A limit order sits in the order book until it is either 1) executed
against a matching market order or 2) it is canceled. A Limit order
fees and smaller tick sizes.
Recently, a huge change has been the creation of dark pools(etc 1. May be executed very quickly if it corresponds to a price near
Lit market). Another issue is the increase in updating frequency
the bid and the ask
of order books.
2. It may take a long time if
18.1 High Frequency Trading
(a) The market price moves away from the requested price.
(b)
The requested price is too far from the bid/ask.
We suspect 60-75% high frequency trading, 10% of which is
predatory. (Amaranth, etc)
3. Can be canceled at any time.
Pros: Smaller tick size; HF traders provide extra liquidity; dark Typically, a limit order waits for a match. The transaction cost
pools reduce trade execution costs from price impact; markets is known, the execution time is uncertain.
more efficient.
Cons: Expensive technological arms race; Dark trading 18.6 Market Order
incentivizes price manipulation, fishing, predatory trading; Little The market order is an order to buy/sell a certain quantity of
or no oversight possible by humans (eg, flash crash) and increased the asset at the best available price in the book. Agents can
systemic risk; HF trading algorithms do not use economic put a market order that, for a buy (sell) order,
fundamentals(e.g. value and profitability of a firm.)
1. The first share will be traded at the ask (bid) price.
18.2 HFT Mishaps
2. The remaining one(s) will be traded some ticks above(below).
Flash crash: Dow Jones IA plunged about 1000 points (recovered in order to fill the order size. The ask (bid) price is then modified
in mintes) biggest one-day point decline.
accordingly.
Other mishaps: AP Twitter Feed/etc.
18.7 Cancellation
18.3 Limit Order Book
We can also cancel orders.
List of all waiting buy and sell orders.
18.8 LOB Dynamics summary
1. The prices are multiples of the tick size.
Agents can put a limit order aand wait for a match.
2. For a given price, orders are arranged FIFO queue
Agents can put a market order that consumes the cheapest
3. At each time t,
limit orders in the book.
(a) The bid price Bt is the price of the highest waiting buy
Agents can put a cancellation.
order.
(b) The ask price At is the price of the lowest waiting sell 18.9 Market impact of large fills
order.
Current mid-price: average. Fill size N = 76015, eg buy. We fill
bid p1; etc. We fill nk shares at
4. The state of the order book is modified by order book events: n1 shares available at best P
Pprice
pk > pk1, such that N = P
ni. The transaction cost is nipi.
Limit orders, market orders, cancelations
1
nipi. The new mid-price is now
5. Consolidated Order book: If the stock is traded in several Then the effective price is N
venues, one aggregates over all visible trading venues. (Then the new average of the order book(with the filled ones removed,
32

18.9. MARKET IMPACT OF LARGE FILLS

naturally.) Note that there is a widening of the bid-ask spread,
as well as a change in the height of the bars (because there is
less volume available now.)

33

Chapter 19

High-Frequency Trading & Limit Order Book 2 - Carmona

19.1

Hidden Liquidity

Some exchanges allow agents to submit hidden orders. They

are made visible to the broader market after being executed.
This is a controversial issue: it is a barrier to the implementation
of a full ytransparent market, and is an impediment to price
discovery and information dissemination.
The results of the first empirical analyses:
1. Encourage fishing
2. After it is revealed that a hidden order was executed: a rash
increase of order placement inside the bid-ask after.
3. HF traders are divided into two groups: Traders trying to
take advantage of the remaining hidden liquidity, and traders
trying to steal execution priority from the fully hidden order.
Distinction between iceberg or fully hidden order.
19.1.1

1. Smith-Farmer-Guillemot-Kirshnamurthy(SFGK) Model
2. Market orders (buys and sells) arrive according to a Poisson
process with rate /2.
3. Cancellation of existing limit orders: outstanding limit orders
die at a rate .
A little bit better one: Cont-Stoikov-Talreja:
1. P = {1, 2, .., n} is a price model
2. LOB at time t is O(t).
3. Admissible state space:
O = {O Zn; 1 k ` n, Op < 0forp k, Op = 0forp = k..`Op
(19.1)
4. Ask price at time t
PA(t) := (n + 1) inf{p; 1 p n, Op(t) > 0}

(19.2)

Partially Hidden Orders: Iceberg Orders

5. Bid price at time t

Dark liquidity posted inside the LOB. Two components: Shown
PB (t) := 0 sup{p; 1 p n, Op(t) < 0}
(19.3)
quantity and the hidden remainder. Order queued with the lit
part of the LIB, only the shown quantity is visible. When the
1

order reaches the front of the queue, only the display quantity 6. Mid Price: P (t) = 2 [PA(t) + PB (t)]
= PA(t) PB (t)
is filled. Then the grade (price and quantity filled) is revealed. 7. Bid-ask spread: S(t)
The hidden part is put at the back of the queue. Sometimes
For the sake of simplicity, we assume changes to the Limit Order
an extra execution fee is charged by the exchange.
Book happen one charse at a time. We review the events causing
the LOB state transitions. One convenient notation here is:
19.1.2 Fully Hidden Order
 p
19.2 Dark Pools
Oj
p1
Oj =
(19.4)
Ojp 1
Dark pools are an electronic engine that matches buy and sell
orders without routing them to lit exchanges. The reason is to
Practical Assumptions:
move large amounts without impacting the price (no need for
iceberg orders). These are run by private brokerages:
1. Limit buy orders arrive at a distance of i ticks from the
opposite best quote at independent, exponential times.
1. Ex: Liquidnet, Pipeline, ITG Posit, Goldman SIGMA X
2. Participatnts submit lists of orders to matching engine
We can summarize all the transition rates with a markov chain.
3. Matched orders are executed at the midpoint of the bid-ask The chain remains in O if it starts there, which is to say that
spread.
PB (t) PA(t), t > 0
(19.5)
4. PROS: trade at midpoint can be better than lit market
5. CONS: may have to wait a long time.
if it is true at time t = 0.
The SEC regulates this in the US as Alternative Trading Systems.
In summary:
They have little to no public disclosure, and little transparency.
1.
This is a descriptive analysis
Supposedly, 32% of trades in 2012 were on dark pools.
2. Uses ideas from queuing theory: first passage times of
19.3 Order book Modeling Objectives
Birth-and -Death processes
Offer a framework to investigate order impact on execution prices. 3. Laplace transform techniques
4. We can compute/estimate probabilities of condition events
1. Optimal mult-period liquidation strategies against a limit
5. But.... its not sufficient for optimal order book strategies.
order book
Optimization problems: The goal of a LOB model is to
2. Detailed but tractable stochastic model of spread and
transaction costs
1. Understand the costs of transactions
3. Benchmark tracking slippage
2. Develop efficient (or optimal) trading strategies
4. Opportunity costs of delayed trading.
Typical challenge: Sell x0 units of an asset and maximize the
sales revenues, using a limited number of market orders only!
19.4 Order Book Models
Roughly speaking, LOB is a set of two histograms (Bids & Asks).
The reduced form model sets it up as a Markov process (Ot)t
on a large state space of order books O. The simplest model:
34

sup
1 ...n <T

N
X
E(U(
PB (i)))
i=1

(19.6)

19.4. ORDER BOOK MODELS

where U is a utility function and E is the expectation over a
model for the dynamics of the LOB Ot.
This is a prohibitively large-dimensional model.

35

Chapter 20

High-Frequency Trading & Limit Order Book 3 - Carmona

Price Impact Models & Optimal Execution

20.1

queries

We already saw that we should split and spread large orders, so:
1.
2.
3.
4.

How can we capture market price impact in a model?

What are the desirable properties of a price impact model?
How can we compute optimal execution trading strategies?
What happens when several execution strategies interact?

20.2

Amlgren-Chriss Price Impact Model

(for example, with the Linear Impact Amlgren-Chriss model:

t).
It = [Xt X0] + X

(20.4)

Our objective is to maximize some form of revnue at time

T , with revenue R(X) from the execution strategy X
Z T
t)Pt)dt
(X
(20.5)
R(X) =
0

Here we assume that the unaffected (fair) price is given by a semimartingale. The mid-price is affected by trading, via two parts: 20.3 Challenges
1. Permanent price impact given by a function g of the trading The first generation considered price impact models: Risk
neutral framework, more complex portfolios (eg, with options), or
speed:
mid
dPt
= g(v(t))dt + dWt
(20.1) robustness and performance constraints(e.g. slippage or tracking
market VWAP).
The second generation uses simplified LOB models, for
2. A temporary price impact given by a function h of the trading
example a simple liquidation problem or performance constraints
speed:
Pttrans = Ptmid + h(v(t))
(20.2) and using both market and limit orders.
20.4 Optimal Execution
The problem is: we want find a deterministic continuous
First we can expand on the definition
transaction path to maximize the mean-variance reward
Z T
1. Closed form solution when permanent and instaneous price
t)Pt)dt
(X
(20.6)
R(X) =
impact functions g and h are linear.
0
2. Efficient frontier: the speed of trading and hence risk/return
Z T
Z T
is controlled by a risk aversion parameter.
tItdt
tP
tdt
X
(20.7)
X
=
This is widely used within the industry.
0
0
Z T
20.2.1 Criticisms
0 +
t C(x)
=x0P
XtdP
(20.8)
mid
0
1. Mid price Pt
is arithmetic brownian motion with drift...
so we can see negative prices, reasonable only for short times, where
maybe that price never actually happens.
C(x) =
(20.9)
2. Are there issues with rate of trading in continuous time?
Can try to maximize expected revenue, but get a boring answer.
3. Price impact is more complex than instantaneous and
A better way is instead to maximize:
permanent.
4.
E [R(X)] var [R(X)]
(20.10)
5. Empirical evidence that it is stochastic.
Here, is a risk aversion parameterlate trades carry some
20.2.2 Optimal Execution
volatility risk.
An execution algorithm has three layers:
For a DETERMINISTIC trading strategy X, we can find the
expectation.
1. Highest: How to slice the order, when to trade, what size,
Instead, we might include risk aversion via a utility function:
how long?
2. Mid: Given a slice, market or limit order? What price level?
3. Low: Given an order, which venue? (we will ignore this!)

E [U(R(XT ))]

All of these models have some shortcomings:

1. they are deterministic
X = (Xt)0tT execution strategy
2. do not react to price changes
Xt: position (the number of shares held) at time t
3. are time-inconsistent
Assume Xt is absolute continuous (to get differentiability)
t the mid-price(unaffected price), Pt transaction price, 4. counter-intuitive
Take P
Furthermore, the computations require
and It is the price impact; that is:
1. solving nonlinear PDEs
t + It
Pt = P
(20.3) 2. singular terminal conditions.
36

Set-up: bf goal: sell x0 > 0 shares by time T > 0.

1.
2.
3.
4.

max

(20.11)

20.5. RECENT DEVELOPMENTS

20.5

Recent Developments

1. Gatheral/Schied(2011)
2. Schied(2012)
3. Almgren-Li (2012): Hedging a large option position. Explicit
solution in some gases)
Modeling the LOB by a shape function:
1. Obizhaeva-Wang (2006)
2. Alfonsi-Fruth-Schied(2010)
3. Alfonsi-Schied-Schulz(2011)
4. Predoiu-Shaikhet-Shreve(2011)

37

Chapter 21

High-Frequency Trading & Limit Order Book 4 - Carmona

Predatory Trading

Large traders facing forced liquidation. Especially if the need

to liquidate is known to other traders. These can be hedge funds
with a nearing margin call, or traders who use portfolio insurance,
stop loss orders, etc. Some institutions/funds can not hold on
to downgrade instruments. Finally, index replication funds at
re-balancing dates(for example, the Russell 3000.)
Forced liquidation can be very costly because of price impact.
Collapses: LTCM vs Amaranth
Reference: Cramer 2002.
Goals:
1. Understand predation
2. Illustrate benefit of stealth trading
3. Illustrate benefit of sunshine trading
The two extremes are:
1. Elastic: temporary impact dominates
2. Plastic: Permanent impact dominates
Optimization problem needs a model of the dynamics of the
order books. We will model it as poisson.

38

Chapter 22

High Frequency Trading Lecture 5

Heterogeneous Beliefs and HF Market Making (Carmona)

Wed like to build a model of agents and have the limit order
book model happen automatically.
Agents:
1. Market Maker: agent that places competitive orders on both
sides of the order book in exchange for privileges. Liquidity
Provider Strategy: adapt pricing and volumes by reading
client flows.
2. Clients: Liquidity Takers, agents who trade with the
market maker. Clients place market orders. Each client has
his/her own information and acts accordingly.
Well assume an ordering in time.
22.1
1.
2.
3.
4.
5.

Theoretical Literature

Early approaches
Inventory models
Informed traders
Zero-Intelligence models
Price impact models.

22.2

Objective

3. The midprice is well defined

4. Marginal costs increase with volume (ie, ct is convex)
5. ct has compact domain.
We can do the Legrendre Transform:
t() :=

sup

(` ct(`))

(22.1)

`supp(ct )

According to Duality:
ctconvex with compact domain t00is a positive finite measure
(22.2)
The distribution t00 represents the order book formed by the
orders of the market maker. If t00 has a density f(x), it is the
shape function we used earlier.
More about the client model: We are NOT trying to implement
an optimal trading strategy. We assume the client is only trying
to predict!
22.3 Client Optimization Problem
Exogeneous state variables
1. Pt is a nonnegative Ito process
2. ct is a random adapted convex function in a fixed domain.
Endogeneous state variables:
 i
dLt = `itdt
(22.3)
dXti = LitdPt ct(`it)dt

Wed like to propose a stochastic agent-based model in which

existence and tractable and realistic properties of the limit order
book appear as a result of the analysis. The client model should
capture the dependence between trades and price dynamics. The
market maker assumes clients are rational and optimizes their
order book choice.
See Carmona, Webster (2012). Mathematics:
where `it is the rate at which the clients trade (control variable),
1. (, F, F = (Ft)t0, P) with W a P-DM that generates F.
k
k
Lit is the volume or total position of the client, and Xti is the
2. F F generated by a P-BM W .
wealth, marked to mid price.
3. Fk such that Pk |Ftk P|Ftk .
Then the objective function is
k
4. Pt is an It
o process adapted to all (F )k=0,..,n.


J i = EPi U i(Xi i , P i )
(22.4)
Note that
1. Each agent has his/her own filtration and probability measure. where U i is the utility function and i is the stopping time. That
2. The filtrations (information structures) are potentially different is, each client maximizes the utility according to his OWN beliefs
3. THe price process is adapted to all of them (i.e., each client about the probability of things happening.
sees the price.)
Theorem 1 Under suitable integrability assumptions on U i and
Trades:
i, the optimal strategy is
1. Midprice Pt is announced by the market at time t.


2. Market maker proposes an order book around Pt
ti := c0t(`it) = EQi P i Pt|Fti
(22.5)
3. The market maker cannot differentiate clients pre-trade
with
4. Client triggers a trade of volume `t
xU i(Xi i , P i )
dQi
5. CLient obtains volume `t and pays cash flow Pt`t +ct(`t) (with


=
(22.6)
ct(`) representing the transaction cost function at time t.)
dPi
EPi xU i(Xi i , P i )
6. The market maker learns the identity of the client post-trade
Note that these are NOT the from the CAPM
(assumption depends upon market, true for FX)
model
The market maker controls transactions cost function ct(`),
Using ti = c0t(`it) or `it = [c0t]1(ti ) = t0 (ti ), we can
and the client i controls trading volumes/speeds `it.
transform
to dual variables:
Hypotheses:

P
dLt = n1 i t0 (ti )dt
1. Marginal costs are defined (ie, ct(`) is differentiable in `)

P
(22.7)
2. Clients may choose not to trade
dXt = LtdPt + n1 i ti t0 (ti ) t(ti ) dt
39

40CHAPTER 22. HIGH FREQUENCY TRADING LECTURE 5HETEROGENEOUS BELIEFS AND HF MARKET MAKING (CARM
Assume the market maker is risk-neutral.
This is super complicated, so well do the natural thing: let
n tend to infinity,
So critically, well need to model the ti ; well have two choices:
1. Microscopic model:
dti = ti dt + dBti + dBt

(22.8)

2. Macroscopic model: Stochastic Partial Differential Equation



1 2
2
dt() = ( + ) + t() + (t()) dtt()dBt
2
(22.9)
What can all of this tell us about Pt? We do not want to
make an explicit model for the price process. Instead, we would
like to infer the price from the client trades.
We can do this via an entropic feedback.
Finally we have a stochastic control problem:
Z
J =
etE [] dt
(22.10)
0

under the constraint 0 het log

We can use the Pontryagin
Can define

t00
t , t idt

m() = ( )

(22.11)

if > 0, and then we have

H()

(22.12)

Chapter 23

HFT Day 6

23.1

Market impact model

These describe the quantitative feedback of a trade execution

strategy on asset prices.
A revenue is given by
Z
RT (X) =

StX dXt

(23.1)

and the liquidation costs are

CT (X) = X0S00 RT (X).

(23.2)

(We may need to add correction terms to these formulas when

X is not continuous.)
We can define regularity in the following ways
1. The model must admit optimal trade execution strategies
for reasonable risk criteria.
2. Optimal strategies ought to be well-behaved.
3. Regularity conditions should be independent of investor
preferences
4. One should distinguish the effects of price impact from
profitable investment strategies that can arise via trend
following. Therefore, we will assume from now on that:
S 0 is a martingale

(23.3)

Definition 2 (Price Manipulation) A round trip is a trade

execution strategy X with X0 = XT = 0. A price manipulation
strategy is a round trip X with strictly positive expected revenues,
E [RT (X)] > 0

(23.4)

41

Part V

Guest Lectures

42

Chapter 24

Risk Measures - Rudloff

Overview:
1. Risk measures: Primal representation, acceptance sets, dual
representation, examples
2. Generalizations: Multivariate risks
Basics: We define a probability space (, F, P ) (the sample
space, -algebra, and probability measure, respectively). Random
variables map the probability space to real numbers.
Risk measures are mathematical models to quantify uncertainy.
It is a functional on the space Lp(, F, P ) with p [0, ] (or
subspaces of random variables):

Lemma 2 Consider a set A Lp. Define

A(X) := inf{t R : X + t A}

(24.3)

It holds
1. inf{t R : t A} = 0 = A(0) = 0
2. A + Lp+ A = A monotone.
3. A satisfies the translation property.
4. A convex = A convex
5. A a cone = A is positively homogeneous
6. A closed = A closed.

: Lp
7 R {+}

(24.1) Lemma 3 There is a one-to-one relationship between lower

semicontinuous risk measures and closed acceptance sets via
The values will be real numbers, in USD or Euros, etc.
1. A = {X Lp : (X) 0} and
Interpretation: The higher (X), the higher the risk.
There are many different risk measures. There is not one 2. (X) = inf{t R : X + t A}.
absolutely objective risk measure. Key questions:
Theorem 2 A function : Lp
7 R {+ inf} is l.s.c. convex
1. Which properties a function should have to be a reasonable risk measure a represetnation of the form
risk measure?
X
2. Which random variables are acceptable for a rm ?
(X) =
{EQ[X] (Q)}
(24.4)
3.
QQ
Some features:
Lp} and (Q) =
where Q := {prob measures Q : dQ
1. R0:Normalization: (0) = 0.
P
P
dQ
Q

p
2. R1:Monotonicity X1, X2 L : X1 X2
=
XA E [X] = ( dP ) is called the penalty function.
(X1) (X2).
then the conjugate function
3. R2:Translation properties (Cash invariance) X Lp, c R,
Aside: If f : X
7 R,
= , (X + c) = (X) c.
f (x) = sup {x(x) f(x)}
(24.5)
Extra features:
xX
1. R3: Convexity in
If f is l.s.c. convex, proper, then
2. R0-R3: Convex Risk Measure
3. R4: Positive Homogeneity (scaling property)
f(x) = sup {x(x) f (x)] = f (x)
(24.6)
4. R0-R4: Coherent Risk Measure
x X
Value-At-Risk is a typical Risk Measure, but it is not convex!
(ie, it is the conjugate of the conjugate.)
24.1 Acceptance Sets
We just need to prove that Q and Q are the same; these fall
We call A := {X Lp : (X) 0 the acceptance set of the out from our original definitions. Start with
risk measure .
(0) = 0
(24.7)
Lemma 1 Consider a function : Lp
7 R {+}. It holds:
which implies
1. monotone = A + Lp+ subset of Ap
inf (y) = 0
(24.8)
2. convex implies A convex
y Lp
3. positively homogeneous implies A is a cone.
which implies that
4. closed implies A is closed.
(y) 0
(24.9)
Reminder: A function F is called closed (or lower semiconNext, since is monotone, we have
tinuous) if
epif := {(v, r) Lp R : f(v) r} Lp R

dom L

(24.2)

is closed in Lp R.
Can we construct a risk measure from a given set of acceptable
positions?

(24.10)

which implies that

(x) = sup {E[Xy] (y)}
43

y L
+

(24.11)

44

CHAPTER 24. RISK MEASURES - RUDLOFF

Finally, from translative, we have that
dom = {y Lp : E[y] = 1

(24.12)

sup {E[xy] }

(24.13)

and

xA

For a l.s.c. coherent risk measure we have that

(X) = sup E Q [(something)]

(24.14)

QQ

A few counterexamples:
1. Variance 2(X) is not monotone or translative!
2. Value at risk V aR(X) at level (0, 1) is not convex!
V aR(X) := inf{x R : P {X + x 0} } = x
(24.15)
Example: Consider two defaultable corporate bonds with face
value $500,000. Payoff X1, X2 (r = 10%, default prob 0.8%, independent). Calculate V aR(X1),V aR(X2), V aR( 12 (X1 + X2)
for = 1%.

50, 000
99.2%
X1 =
(24.16)
500, 000 0.8%

98.4%
50, 000
1
y = (X1 + X2) = 225, 000 1.58

2
500, 000 ??0.006??

(24.17)

This example also shows VaR is not subadditive!

A better one: Conditional Value at Risk:
CV aR(X) :=
or

1
E[X1{Xxa }] + x( P [X x])

(24.18)
Z
1
=
V aR (X)d
(24.19)
0

Also called Average Value at Risk or Expected shortfall at level

(0, 1).
Other ones: Expected loss, worst-case risk measure.
(X) := E[X], max
Consider the exponential utility function u(x) = 1 ex and
the set of acceptable positions whenever the expected utility is
nonnegative.
24.2
24.2.1

Outlook
Multivariate risks

One extension is to now have many random variables instead

of just one that is, X Lpd.
Why would this be important?One application: Important for
dealing with many banks which are interlinked and want capital
requirements for each bank taking interconnected into account.
Another: Markets with transaction costsyou cant sum over
individual stocks because of transaction costs.
For such porfolios, we can look at
R(X) = {u Rd : X + u A}

(24.20)

for some acceptance set A Lpd.

A similar approach (of defining these from their properties)
can be followed here, with similar result.

Chapter 25

Cumulant Moments & Queuing models

45

Bibliography

46