You are on page 1of 28

1. Introduction 2. ABM-CARMA Model 3. Emprical Model Performance 4.

Conclusion
Commodity Modelling with CARMA Processes
Marcel Prokopczuk
ICMA Centre, Henley Business School
Practical Quantitative Analysis in Commodities
London, 18 June 2010
Marcel Prokopczuk, ICMA Centre 1/28
1. Introduction 2. ABM-CARMA Model 3. Emprical Model Performance 4. Conclusion
Motivation: Oil Price Development
Dec96 Dec98 Dec00 Dec02 Dec04 Dec06 Dec08
0
50
100
150
U
S
D

p
e
r

B
a
r
r
e
l
Oil Price
Marcel Prokopczuk, ICMA Centre 2/28
1. Introduction 2. ABM-CARMA Model 3. Emprical Model Performance 4. Conclusion
Motivation: Term Structure of Futures Prices
0 5 10 15 20 25
19
20
21
22
23
24
25
26
27
08Jan1997
U
S
D

p
e
r

B
a
r
r
e
l
Maturity [months]
Marcel Prokopczuk, ICMA Centre 3/28
1. Introduction 2. ABM-CARMA Model 3. Emprical Model Performance 4. Conclusion
Motivation: Term Structure of Futures Prices
0 5 10 15 20 25
14.5
15
15.5
16
16.5
17
17.5
18
13May1998
U
S
D

p
e
r

B
a
r
r
e
l
Maturity [months]
Marcel Prokopczuk, ICMA Centre 4/28
1. Introduction 2. ABM-CARMA Model 3. Emprical Model Performance 4. Conclusion
Motivation: Term Structure of Futures Prices
0 5 10 15 20 25
24
25
26
27
28
29
06Jun2001
U
S
D

p
e
r

B
a
r
r
e
l
Maturity [months]
Marcel Prokopczuk, ICMA Centre 5/28
1. Introduction 2. ABM-CARMA Model 3. Emprical Model Performance 4. Conclusion
Motivation: Term Structure of Futures Prices
0 5 10 15 20 25
71
71.5
72
72.5
73
73.5
74
74.5
31May2006
U
S
D

p
e
r

B
a
r
r
e
l
Maturity [months]
Marcel Prokopczuk, ICMA Centre 6/28
1. Introduction 2. ABM-CARMA Model 3. Emprical Model Performance 4. Conclusion
Motivation: Term Structure of Futures Prices
0 5 10 15 20 25
104
106
108
110
112
114
30Apr2008
U
S
D

p
e
r

B
a
r
r
e
l
Maturity [months]
Marcel Prokopczuk, ICMA Centre 7/28
1. Introduction 2. ABM-CARMA Model 3. Emprical Model Performance 4. Conclusion
Pricing of Futures Contracts
Futures contracts can be priced by no-arbitrage arguments:
Financial contracts: Cost-of-carry
F
t
(T) = S
t
e
r (Tt)
With storage costs:
F
t
(T) = S
t
e
(r +s)(Tt)
No explanation for backwardation
Inferior empirical performance
Equity futures can show backwardation due to dividends
F
t
(T) = S
t
e
(r q)(Tt)
Marcel Prokopczuk, ICMA Centre 8/28
1. Introduction 2. ABM-CARMA Model 3. Emprical Model Performance 4. Conclusion
Pricing of Commodity Futures Contracts
Commodities dier from pure nancial assets as they are
hold for consumption or production
Similar to the dividend yield of a stock, the holder of the
commodity receives a convenience yield from holding
stocks of commodities
Kaldor (1939): ... stocks of goods... also have a yield...,
by enabling the producer to lay hands on them the
moment they are wanted, and thus saving the cost and
trouble of ordering frequent deliveries, or waiting for
deliveries.
Marcel Prokopczuk, ICMA Centre 9/28
1. Introduction 2. ABM-CARMA Model 3. Emprical Model Performance 4. Conclusion
Modelling the Convenience Yield
Convenience yield deterministic function of price:
Brennan/Schwartz (1985)
Brennan (1991)
Poor empirical performance
Stochastic convenience yield:
Gibson/Schwartz (1990)
Schwartz (1997)
Schwartz/Smith (2000)
Cassasus/Collin-Dufresne (2005)
All models assume explicitly or implicitly that the
convenience yield follows an Ornstein-Uhlenbeck process
Marcel Prokopczuk, ICMA Centre 10/28
1. Introduction 2. ABM-CARMA Model 3. Emprical Model Performance 4. Conclusion
Modelling the Convenience Yield
The assumed Ornstein-Uhlenbeck process is the
continuous limit of an AR(1) process
An analysis of the approximated (net) convenience yield

t,T1,T
= ln

F(t,T)
F(t,T1)

shows that an AR(1) is not able to capture the


dynamics appropriately
An ARMA(1,1) or higher order AR(q) model yield much
better t to the data
Marcel Prokopczuk, ICMA Centre 11/28
1. Introduction 2. ABM-CARMA Model 3. Emprical Model Performance 4. Conclusion
Outline
Contribution of this work:
1. Formulation of a commodity pricing model in
continuous time allowing for higher order auto-
regression and moving average components:
ABM-CARMA(p,q)
2. Derivation of closed-form solutions for futures and
options prices
3. Application to the crude oil futures market,
demonstrating the models superior empirical
performance
Marcel Prokopczuk, ICMA Centre 12/28
1. Introduction 2. ABM-CARMA Model 3. Emprical Model Performance 4. Conclusion
Model Description: ABM-CARMA(2,1)
Latent factor spot price model in continuous time:
One non-stationary factor Z
t
: long-term equilibrium
modelled by an Arithmetic Brownian Motion
One stationary factor Y
t
: short-term deviations from
the equilibrium modelled by a CARMA(1,0) process
(Schwartz/Smith 2000)
ln S
t
= Z
t
+ Y
t
dZ
t
= dt +
Z
dW
Z
t
dY
t
= kY
t
dt +
Y
dW
Y
t
dW
t
Marcel Prokopczuk, ICMA Centre 13/28
1. Introduction 2. ABM-CARMA Model 3. Emprical Model Performance 4. Conclusion
Model Description: ABM-CARMA(2,1)
Latent factor spot price model in continuous time:
One non-stationary factor Z
t
: long-term equilibrium
modelled by an Arithmetic Brownian Motion
One stationary factor Y
t
: short-term deviations from
the equilibrium modelled by a CARMA(2,0) process
xxxxxxxxxxxxxxxxxx
ln S
t
= Z
t
+ Y
t
dZ
t
= dt +
Z
dW
Z
t
d

Y
t
= k

Y
t
dt +
Y
dW
Y
t
dY
t
=

Y
t
dt
Marcel Prokopczuk, ICMA Centre 14/28
1. Introduction 2. ABM-CARMA Model 3. Emprical Model Performance 4. Conclusion
Model Description: ABM-CARMA(2,1)
Latent factor spot price model in continuous time:
One non-stationary factor Z
t
: long-term equilibrium
modelled by an Arithmetic Brownian Motion
One stationary factor Y
t
: short-term deviations from
the equilibrium modelled by a CARMA(2,1) process
xxxxxxxxxxxxxxxxxx
ln S
t
= Z
t
+ Y
t
dZ
t
= dt +
Z
dW
Z
t
d

Y
t
= k

Y
t
dt +
Y
dW
Y
t
dY
t
=

Y
t
dt
Marcel Prokopczuk, ICMA Centre 15/28
1. Introduction 2. ABM-CARMA Model 3. Emprical Model Performance 4. Conclusion
Model Description: ABM-CARMA(2,1)
Latent factor spot price model in continuous time:
One non-stationary factor Z
t
: long-term equilibrium
modelled by an Arithmetic Brownian Motion
One stationary factor Y
t
: short-term deviations from
the equilibrium modelled by a CARMA(2,1) process
xxxxxxxxxxxxxxxxxx
ln S
t
= Z
t
+ Y
t
+

Y
t
dZ
t
= dt +
Z
dW
Z
t
d

Y
t
= k

Y
t
dt +
Y
dW
Y
t
dY
t
=

Y
t
dt
Marcel Prokopczuk, ICMA Centre 16/28
1. Introduction 2. ABM-CARMA Model 3. Emprical Model Performance 4. Conclusion
Model Discussion
Model is formulated directly under the equivalent
martingale measure
Closed form (ane) solutions for the futures price:
ln F(Y
t
,

Y
t
, Z
t
, t; T) = Z
t
+ A

ABM
+ B

Y
t
+ CY
t
+ D

CARMA
Dierence to the standard Schwartz/Smith 2000 model:
Term structure:
Much more exible, especially at the short end
Volatilities:
Non-monotonous structure and higher curvature
Marcel Prokopczuk, ICMA Centre 17/28
1. Introduction 2. ABM-CARMA Model 3. Emprical Model Performance 4. Conclusion
Factor Loading of

Y
t
B
AR
B
MA
0 2 4 6 8 10
0.0
0.1
0.2
0.3
0.4

B
A
R

0 2 4 6 8 10
0.0
0.2
0.4
0.6
0.8
1.0

B
M
A

Marcel Prokopczuk, ICMA Centre 18/28


1. Introduction 2. ABM-CARMA Model 3. Emprical Model Performance 4. Conclusion
Factor Loading of Y
t
C
AR
C
MA
0 2 4 6 8 10
0.0
0.2
0.4
0.6
0.8
1.0

C
A
R

0 2 4 6 8 10
0.20
0.15
0.10
0.05
0.00

C
M
A

t
,
t

Marcel Prokopczuk, ICMA Centre 19/28


1. Introduction 2. ABM-CARMA Model 3. Emprical Model Performance 4. Conclusion
Entire Futures Curve
0 2 4 6 8 10
2.5
3.0
3.5

Z
t

X
t

Marcel Prokopczuk, ICMA Centre 20/28


1. Introduction 2. ABM-CARMA Model 3. Emprical Model Performance 4. Conclusion
Model Implementation: Data
Data used:
Crude oil futures traded at the New York Mercantile
Exchange (NYMEX)
Sample period: January 1996 to December 2008
Weekly observations (Wednesday)
Maturities 1 to 24 months
Data source: Bloomberg
Panel data set of 676 x 24 observations
Marcel Prokopczuk, ICMA Centre 21/28
1. Introduction 2. ABM-CARMA Model 3. Emprical Model Performance 4. Conclusion
Model Implementation: Kalman Filter Estimation
Implementation of the ABM-CARMA(2,1) model:
Write discretized version in state space form
Dynamics of latent factors: Translation equations
x
t+t
= Gx
t
+c +
t
with
c = t ,
G = e
At
,
E
t
[
t
] = 0 ,
V
t
[
t
] =

t
e
A(tu)
VV

e
A

(tu)
du
Marcel Prokopczuk, ICMA Centre 22/28
1. Introduction 2. ABM-CARMA Model 3. Emprical Model Performance 4. Conclusion
Model Implementation: Kalman Filter Estimation
Add measurement error to the pricing formula:
Measurement equations
y
t
= d +Hx
t
+
t
with
d = [

(T
1
t) +
1
2

B(t, T
1
), ...,

(T
k
t) +
1
2

B(t, T
k
)]

H = [

A(t, T
1
), ...,

A(t, T
k
)]

E
t
[
t
] = 0
V
t
[
t
] =
Kalman lter maximum likelihood estimation of
parameters
Marcel Prokopczuk, ICMA Centre 23/28
1. Introduction 2. ABM-CARMA Model 3. Emprical Model Performance 4. Conclusion
In-Sample Pricing Errors
Benchmark: Schwartz/Smith (2000)
Root Mean Squared Error
Absolute %-Decrease Relative
F01 0.0409 0.0486 15.8% 1.26% 1.49%
F02 0.0283 0.0330 14.2% 0.87% 1.02%
F03 0.0207 0.0230 10.0% 0.64% 0.71%
All 0.0122 0.0141 13.5% 0.38% 0.43%
AIC
ABMCARMA
= 157, 285, AIC
SS2000
= 152, 935,
SIC
ABMCARMA
= 157, 131, SIC
SS2000
= 152, 795.
Marcel Prokopczuk, ICMA Centre 24/28
1. Introduction 2. ABM-CARMA Model 3. Emprical Model Performance 4. Conclusion
Out-of-Sample Pricing Errors: Time-Series
Split Data Sample into two periods
Estimation: First half
Prediction: Second half
Root Mean Squared Error
Absolute %-Decrease Relative
F01 0.0564 0.0627 10.1% 1.48% 1.59%
F02 0.0510 0.0543 5.9% 1.33% 1.38%
F03 0.0472 0.0488 3.2% 1.23% 1.24%
All 0.0375 0.0381 1.6% 0.94% 0.95%
Marcel Prokopczuk, ICMA Centre 25/28
1. Introduction 2. ABM-CARMA Model 3. Emprical Model Performance 4. Conclusion
Out-of-Sample Pricing Errors: Cross-Section
Split Data Sample into two parts
Estimation: F01 - F12
Prediction: F13 - F24
Root Mean Squared Error
Absolute %-Decrease Relative
F15 0.0068 0.0090 24.4% 0.21% 0.29%
F18 0.0115 0.0149 22.8% 0.36% 0.48%
F21 0.0173 0.0216 19.9% 0.55% 0.71%
F24 0.0237 0.0284 16.5% 0.75% 0.93%
All 0.0144 0.0179 19.6% 0.46% 0.59%
Marcel Prokopczuk, ICMA Centre 26/28
1. Introduction 2. ABM-CARMA Model 3. Emprical Model Performance 4. Conclusion
Conclusion
AR(1) poor description of the convenience yield
Extension of Schwartz/Smith model using continuous time
limit of ARMA processes to describe the convenience yield
Results in:
More exible futures curves
Without the use of additional risk factors
Applied to crude oil futures:
Better t/prediction at the short end
Better prediction of long maturity contracts from short
maturity contracts
Marcel Prokopczuk, ICMA Centre 27/28
1. Introduction 2. ABM-CARMA Model 3. Emprical Model Performance 4. Conclusion
More
Complete paper:
Commodity derivatives valuation with autoregressive and moving
average components in the price dynamics
(with R. Paschke)
forthcoming in Journal of Banking and Finance 2010
More work on commodity modelling:
Marcel Prokopczuk
ICMA Centre - Henley Business School
University of Reading
m.prokopczuk@icmacentre.ac.uk
www.icmacentre.ac.uk/about us/academic sta/dr marcel prokopczuk
Marcel Prokopczuk, ICMA Centre 28/28