Emmanuel Gincberg

MINI CASE STUDY SERIES:
TWO FACTORS MODEL CALIBRATION

Emmanuel Gincberg

CREDIT AGRICOLE CIB. Head of Commodity Quantitative Analytics.

The views expressed in this document are those of the author and do not necessarily
represent those of his employer.

Marcus Evans. Practical Quantitative Analysis in Com-
modities. 17th June 2010

MODELLING THE FUTURE CURVE

2


3 MODELLING THE FUTURE CURVE

The most activated traded instruments in commodity financial markets
are future and forward contracts.

NYMEX and ICE exchange markets enable to trade future con-
tracts up to 72 consecutive months for many energy assets.

Agricultural futures can be traded on the CBOT exchange.

Forward agreements are OTC contracts which generally settle on
monthly average of future closing prices.

Commodity Swaps are transactions where the floating price is
indexed on future prices.


Many commodity derivatives depend on different future contracts
(Swaptions, time spreads,).

It is therefore important to not only correctly model each future
contract but also to correctly represent the joint distribution of these
assets.

To do so it is convenient to consider the future curve as a whole
rather than a collection of individual points

BACKWARDED FUTURE CURVE

CONTANGO FUTURE CURVE

BACKWARDATION TO CONTANGO

FUTURE CURVE VARIATIONS

This particular observation illustrates general behaviours of commodity
future curves:

Commodity prices are governed by supply and demand.

Future curves short ends are generally much more volatile than long
ends (mean reversion).

Short dated future are used to cover unanticipated demand, their
prices are impacted by:
Weather conditions
Pipeline failures
Political events
Etc...

SHORT TERM / LONG TERM MODEL

Schwartz-Smith proposed a model to describe the variations of the fu-
ture curve where the spot price logarithm is driven by two factors:

ln(
t
) = _
t
+
t

_
t
is the short term process:
_
t
= _
t
+o

t
is the long term equilibrium process:
t
= p
{
+o
{

t
{

[
t
t
{
= p


Under these assumptions the future prices are lognormal variables:
ln
|
F(t, T)]
=
-k(1-t)
_
t
+
t
+ ( )

|ln
|
F(t, T)]
] =
-k1
_
0
+
0
+ p
{
+ ( )

vai|ln
|
F(t, T)]
] =
-2k1
(
2kt
1)
c
_
2
2k
+o
{
2
+2
-k1
(
kt
1)
c
_
c
(
k

() is a deterministic function:
() = p
{
+
1
2
(1
-2k1
)
o
2
2
+ o
{
2
+2(1
-k1
)
o
o
{


Although, as shown in the Schwartz-Smith paper, this approach is
equivalent to a stochastic convenience model, it is probably more intui-
tive due to the break down between short and long term drivers.
_
t
is a mean reverting process which has more impact on short
dated maturities.

t
creates parallel shifts of the curve

Short dated futures tend to revert to the long run equilibrium
regime which is driven by
t

These variables are however not observable

FUTURE CURVE MODEL

The previous approach consists in defining the spot price driving vari-
ables and to deduce the corresponding future prices dynamic. Alterna-
tively we can directly model the future price:
uF(t, T) = F(t, T)
I
(t, T) uW
t
I
I

The short term / long term model corresponds to a specific formula-
tion of the volatility functions
I
(t, T) where the number of factors is
set to 2.

In a more general setting the number of Brownian processes W
t
I
can
be reduced by PCA techniques.

MODELLING THE FUTURE CURVE 13
FUTURE CURVE MODEL

A more generic formulation of the short term long term future curve
model can therefore be expressed as follows:
F(t, T
I
) = F(u, T
I
) exp _
1
2
i
(t, T)
2
+
-k(T
-t)
_
t
+
t
_
Where
F(u, T
I
) is the current price of the future contract expiring at T
I
.
_
t
=
-kt
]
ku
o
()
u
t
0

t
= ] o
{
()
u
{
t
0

[
t
t
{
= p

I
(t, T) =
_
-2kT
]
2ku
o
()
2
+
t
0
] o
{
()
2
+
t
0

-kT
]
ku
o
()o
{
()
t
0

14 CALIBRATION

CALIBRATION

15 CALIBRATION
KALMAN FILTER

In their original papers Gibson-Schwartz and Schwartz-Smith propose to
estimate the model parameters with a Kalman filtering technique.

The kalman filter is a recursive method for estimating non observable
variables from observable variables.

Kalman results can be statically unstable.

There is no guarantee that this procedure would produce a set of pa-
rameters consistent with implied volatilities.

The model should be in line with the term structure of volatilities if we
want to use the corresponding options for hedging.

16 CALIBRATION
LISTED OPTIONS

Similarly to the future market there exists a listed commodity market
for options: For each future contract there generally exist Call and Put
options expiring shortly before the corresponding future maturity .
These options are the basic instruments of choice for the
hegdge of Vega risks linked to more complex derivatives

They imply a term structure of implied volatilities but contrary
to other markets the underlying is different for each maturity:

In Equity markets a term strucure of volatility defines a
strip of option prices on the same asset observed at
different times.

In most commodity markets each volatility point
corresponds to a unique future underlying expiry.

CALIBRATION 17
TERM STRUCTURE OF VOLATILITY

A term structure of volatility represents the variation of volatility with
maturity.

It is common to represent the at-the-money volatilities to illustrate
the term structure.

Commodity term structures tend to be downward sloping. It can be
explained by the mean reverting nature of these markets.

Each volatility point of the term structure corresponds to a unique
future contract point

18 CALIBRATION

19 CALIBRATION

Although volatility term structure varies with time, the downward slop-
ing profile is a persistent characteristic of volatility term structures.
By time homogeneity we can deduce that the volatility term struc-
ture of a fixed future contract should be upward sloping
(Samuelson effect).

This is consistent with the short term/long term model:
vai ln|F(t, T)] =
2
(
2
1)
o
_
2
2
+2
1)
o
_
o
+o
2

The first two terms are bounded as t & T tends to infinity
i.e. the corresponding volatility is decreasing towards 0.

When T is fixed the first two terms are increasing with t.

20 CALIBRATION
VOLATILITY TERM STRUCTURE CALIBRATION

But calibration to ATM volatilities will not provide a unique solution.

= 116%,
= 11u%,
= 2S%, = 88%

21 CALIBRATION

Another satisfactory calibration:

= 22S%,
= 9u%,
= 18%, = 2u%

CALIBRATION 22

The previous calibrations are obtained by global minimization of the
differences between listed options implied volatilities and the corre-
sponding volatilities computed with the two factors model parame-
ters (with constant instantaneous volatilities
and
{
).

Another approach consists in extrapolating
{
from the term struc-
ture of volatility as it corresponds to the implied volatility limit
(
_
Var|In|F(t,T)]]
t
) when (t, T) tend toward infinity:
Each point of the ATM volatility term structure can then be
exactly matched by introducing piecewise-constant instanta-
neous volatilities o
().
However there is no unique solution: each couple (, ) will
lead to a different o
() profile.

CALIBRATION 23

In order to illustrate the previous calibration method let us denote by
(t
I
, o
A1M
()) the market option maturities and corresponding ATM implied
volatilities (to simplify the calculations we assume p = u).

If we suppose that (, ) are known it remains to define o
() = o
() for
]
,
+1
] such as
-2kT
]
2ku
o
()
2
+
t
i
0
o
{
2

=
I
2

= 1

It implies the following result:

o
()
2
=
(o
A1M
()
2
o
{
2
)

2k(T
-T
-1
)
(o
A1M
( 1)
2
o
{
2
)
-1
-2kT
2
(
2kt

2kt
-1
)

It can only be solved if
(o
A1M
()
2
o
{
2
)

2k(T
-T
-1
)
(o
A1M
( 1)
2
o
{
2
)
-1
u

CALIBRATION 24

We have seen that calibrating (, ) to the term structure of volatility can
generate different results:

Two acceptable calibrations can create different correlation profiles
of the future curve.

The volatility term structure profileo(, ) of a given future contract
implied by the model is mainly driven by the mean reversion
level:
o(, ) = _
-2k1
(
2kt
1)
o
2
2
+2
-k1
(
kt
1)
o
o
{
+o
{
2
_
1
2
,

CALIBRATION 25

Volatility term structure of a fixed contract

CALIBRATION 26
HISTORICAL FUTURE PRICES

When no additional type of option quotes (swaptions, time spread
options) relevant for the calibration of the mean reversion/correlation
are available, an alternative solution is to use historical future prices.
Let us consider historical future prices
{
F(t, t +
I
)]
0<t<M
0<I<N
, where t
represent past observation dates, and
I
are constant expiries (they
are in fact approximated by the average time to maturity of the k-th
nearest future).

We focus on the stochastic term
i
(t) of the corresponding loga-
rithmic returns n _
F(t+t,t+t+
i
)
F(t,t+
i
)
_:
(t) = _
-k(t+t+
-u)
o
t+t
0
_
-k(t+
-u)
o
+
t
0

-kT
_ o
{

u
{
t+tt
t

CALIBRATION 27

With a local expansion around t~u, we can write
or|
(t)] = |
-k
]
2
+o
{
2
t +(t
2
)
Up to the first order of t, the stochastic terms of the logarithmic re-
turns are independent and equi-distributed:
(t)~u, |
-k
]
2
+o
{
2
t
Similarly we can compute the covariances c
]
(, o
, o
{
) corresponding
to two different contracts:
c
]
(, o
, o
{
) = _
-k(
+
]
)
o
2
+o
{
2
] t

The calibration method consists in minimizing the differences between
the model and the historical covariances.

CALIBRATION 28

= 84%,
= 17%,
{
= 2u%, = 19.S9%

29 CALIBRATION
VOLATILITY SEASONALITY

In some markets, for example Henry Hub natural Gas, the term struc-
ture of volatility is strongly seasonal.

It can be explained by higher demand, and therefore volatility, during
heating periods in winter.

The short term/long term breakdown cannot explain these term struc-
tures, alternative approaches have to be considered.

Each month of the year can be modelled by a distinct underlying creat-
ing twelve different term structure of implied volatilities.

30 CALIBRATION

A typical Natural Gas term structure of ATM volatilities

31 CONCLUDING REMARKS

Natural Gas ATM volatilities with the
August and November months highlighted.

CALIBRATION 32

Each season can be calibrated separately and it remains to define the
correlations between summer and winter:
ln
|
F
Summcr
(t, T)]
=
-k
S
(1-t)
_
t
S
+
t
S
+
S
( )
ln
|
F
WIntcr
(t, T)]
=
-k
W
(1-t)
_
t
w
+
t
w
+
w
( )
The summer-winter correlations can be estimated by historical
analysis or implied from Calendar spread options, the most liquid
pairs being March-April and October-January.
[
t
t
{
W
= p
{
Sw
, [
t
{
s
W
= p
{
Sw

[
t
W
= p
Sw
, [
t
{
s
t
{
W
= p
{
Sw

CALIBRATION 33
HEDGING CONSIDERATIONS

Different calibration methods will imply different hedging strategies.

Let us assume two different approaches where the mean reversion is
calibrated to historical covariances. We want to price a calendar spread
option:
|(
1
,
1
) (
1
,
2
)]
+

Forward model:
uF(t, T
I
) = F(t, T
I
)
I
(t, T
I
) uW
t
I

Each volatility function
I
(t, T
I
) is calibrated separately to the appropri-
ate ATM volatility point.

CALIBRATION 34

Short term/long term model ( = u):
ln
|
F(t, T
I
)]
= +
-k1
_
ku
t
0
o
(u)
t
+o
{

t
{

o
{
is calibrated to the implied volatility curve asymptote:
vai|ln|F(t, T)]]
t

t,T
---- o
{
2
= o
A1M
2
()
o
() is a piecewise constant function calibrated to the ATM volatility

curve.
-2k1
i
_
2ku
t
i
0
o
2
() +o
A1M
2
()
= o
A1M
2
(i)
2
() =

2k1
i

|o
A1M
2
() o
A1M
2
()]
2k1
i-1

-1
|o
A1M
2
( 1) o
A1M
2
()]
2kt
i

2kt
i-1
2

CALIBRATION 35


With the Forward model the value of the Calendar spread option de-
pends on both volatilities
ATM
(i
1
) and
ATM
(i
2
).

With the short term/long term model the value of the Calendar spread
option depends on both volatilities
ATM
(i
1
) and
ATM
().

Alternatively a short term/long term model can be calibrated to a few
fixed points of the ATM volatility term structure (the more liquid) and
the Vega profile of any options will be depend on the calibrated points
only.

CALIBRATION 36

CONCLUDING REMARKS

Multiple factors are necessary to correctly represent the variations
of the future curve.

To obtain satisfactory calibration - stable and in line with vanilla op-
tions - one has to combine different approaches (minimization, sta-
tistical estimation).

The calibration procedure impacts directly the risk profile and there-
fore should be decided accordingly to a hedging strategy.

37 REFERENCES
REFERENCES

Rajna Gibson; Eduardo S. Schwartz. Stochastic Convenience Yield and the Pricing of Oil
Contingent Claims. The Journal of Finance, Vol. 45, No. 3, Papers and Proceedings, Forty-
ninth Annual Meeting, American Finance Association, Atlanta, Georgia, December 28-30,
1989. (Jul., 1990), pp. 959-976.
Eduardo Schwartz; James E. Smith. Short-Term Variations and Long-Term Dynamics in
Commodity Prices. Management Science Vol. 46, No. 7, July 2000 pp. 893911
Helyette Geman. Commodities and Commodity Derivatives: Pricing and Modeling Agricul-
tural, Metals and Energy, January 2005, Wiley Finance

Emmanuel Gincberg

Uploaded by

Copyright:

Available Formats

Emmanuel Gincberg

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Emmanuel Gincberg

Uploaded by

Copyright:

Available Formats

MINI CASE STUDY SERIES:

TWO FACTORS MODEL CALIBRATION

CREDIT AGRICOLE CIB. Head of Commodity Quantitative Analytics.

() is a piecewise constant function calibrated to the ATM volatility

You might also like