You are on page 1of 43

MODELING SEASONALITY IN COMMODITY PRICE DYNAMICS

YANPING CHONG, WEIXIAN KONG, MATTHEW TOCHTERMAN, JUNYUE XU, AND HANGYING YU
A BSTRACT. The industry standard model for commodity price dynamics implies constant correlation between returns of futures with different tenors. We extend the model by allowing its
parameters to vary over time. This practice enables us to capture the seasonality effect embedded in the evolution of a commodity futures curve in addition to the seasonal patterns observed
in the futures curve term structure. We use a storage valuation problem as an example to
illustrate how to apply our model in practice.
K EYWORDS: Commodity, natural gas, forward curve, HJM, Clewlow-Strickland, seasonality,
storage valuation.
ACKNOWLEDGMENTS : The authors would like to thank Nancy Wallace for discussions and
comments as well as for providing some part of the data. The authors would also like to thank
Matthew Ohara and Eric Reiner for comments.

1. I NTRODUCTION
The Clewlow and Strickland (1999a) model is the industry standard model for commodity
price dynamics. We extend it by allowing its parameters to vary across different months.
With such a modification, the model better captures the seasonality effect embedded in the
evolution of the commodity futures curve. We use the natural gas storage valuation problem
as an example to show how to use our model in practice. Specifically, we investigate three
different methods for valuing a storage facility and discuss and interpret the differences in
these valuation methods.
Many derivatives and real options are valued based on the spot/future price of a commodity.
Common examples include spark spread options, tolling agreements, crack spread options,
Date: March 20, 2013.
1

Electronic copy available at: http://ssrn.com/abstract=2503252

Y. CHONG, W. KONG, M. TOCHTERMAN, J. XU, AND H. YU

and commodity storage valuation. In order to price these derivatives, we need a model that
can accurately characterize the evolution of the commodity price. There are two popular
approaches in the literature. The first approach starts by modeling the spot price evolution.
For example, Schwartz (1982), Gibson and Schwartz (1990), and Schwartz (1997) all fall into
this category. Under the risk neutral measure, models under this approach resemble the Hull
and White (1990) interest rate models and have the general form:
dSt
= (t t ) dt + t dWt .
St
Most of the related work has been focused on finding/calibrating a proper process for the
convenience yield t and the instantaneous volatility t . There are two main goals. One is to
achieve empirically observed mean-reversion in the spot price (Schwartz (1997)). The second
is to fit a futures price volatility term structure in which the short-dated futures usually have
higher volatility than long-dated futures (Bessembinder, Coughenour, Seguin, and Smoller
(1996)).
As pointed out by Clewlow and Strickland (1999b), in order to be calibrated to the term
structure of the futures curve, spot price models usually involve state variables that are unobservable, so the estimation and implementation of them can be very complicated. Perhaps
more importantly, these models cannot be easily calibrated to the term structure of futures
prices, which is especially important in valuing commodity contingent claims as explained
below.
Recall that the forward price of a conventional tradable financial asset has a strict relationship with the spot price in the absence of arbitrage opportunities. Assuming a deterministic
interest rate path r(t) and no dividend yield to the asset, the futures price is equal to forward
price and we have
Z

F (t, T ) = S(t) exp


ru du ,

u=t

Electronic copy available at: http://ssrn.com/abstract=2503252

(1)

MODELING SEASONALITY IN COMMODITY PRICE DYNAMICS

where F (t, T ) is the futures price at time t with maturity date T and S(t) is the spot price.
The futures and spot prices of a commodity, however, do not need to have such a relationship
due to the inability of market participants to carry out arbitrage strategies. On the one hand,
if the futures price of a commodity is lower than implied by equation (1), traders cannot
short natural gas in the spot market and long the future to make an arbitrage profit as is the
case in the stock market. On the other hand, if the commodity futures price is higher than the
normal level, traders cannot simply buy in the spot market and hold the natural gas and short
futures unless they have a physical place to store the purchased commodity. Empirically, we
observe futures curves for many commodities in contango or backwardation for long periods,
implying that those arbitrage trading opportunities are either inexecutable or too costly to
market participants. Furthermore, some commodity prices exhibit seasonal patterns, such as a
peak during the winters and a trough during the summers.
The HJM framework introduced by Heath, Jarrow, and Morton (1992) allows us to easily
capture the observed shape of the commodity futures curve. Under this framework, the dynamics of the futures price are modeled directly, with the current futures curve as the initial
condition. Therefore, the term structure of the futures curve is calibrated automatically. Cortazar and Schwartz (1994) are the first to use this method. Noticing that the instantaneous
return on all futures contracts is equal to zero under the risk neutral measure, they assume the
futures price follows a multi-factor stochastic process:
K

dF (t, T ) X
=
bk (t, T ) dWk .
F (t, T )
k=1
where Wk is a Brownian motion. Factor loadings bk (t, T ) are estimated using PCA for observable futures contracts and then scaled for longer maturity futures contracts to match the
historical volatility.
Clewlow and Strickland (1999a) extend this model and develop a practical and efficient
framework for estimating the multi-factor HJM model and the pricing and risk management

Y. CHONG, W. KONG, M. TOCHTERMAN, J. XU, AND H. YU

of energy derivatives. It has become the industry standard and is widely used in practice.
Clewlow and Strickland (1999b) connect the futures price dynamics to the spot price dynamics
and show that the single factor HJM model is a generalized version of the Schwartz (1997)
model.
Since the Clewlow-Strickland model is perfectly calibrated to the current futures curve,
it accurately characterizes the seasonal pattern observed in the futures curve term structure.
However, it does not model the seasonality embedded in the evolution of the entire futures
curve. For example, as we shall see in Section 3, the correlation between futures contracts
with different maturities changes through time and also exhibits seasonality. However, the
Clewlow-Strickland model implicitly assumes a constant correlation structure between different futures contracts. In light of this, we extend the Clewlow-Strickland framework such that
it takes into account the seasonality embedded in the evolution of the futures curve. Based on
this extension, we then apply three different methods to value a storage facility using market
data. Each valuation method has different underlying assumptions about the trading strategies
used to exploit the seasonality in natural gas (futures/spot) prices.
The rest of the paper is organized as follows. In Section 2, we describe the data used in
this paper. We discuss in detail the seasonality effect in the natural gas market in Section 3.
Section 4 presents our model and explains how to calibrate it to market data. In Section 5, we
describe the storage valuation problem and introduce three different ways to solve it using our
model. Section 6 concludes the paper.

2. DATA
We have daily data of Henry Hub natural gas futures prices from Bloomberg. Our sample
spans from January 1, 2000 to February 12, 2013. On each day, we have prices of 12 shortestdated futures contacts, with monthly maturity dates from January to December. Named after

MODELING SEASONALITY IN COMMODITY PRICE DYNAMICS

their maturity months, these contracts are called January, February, . . ., December contracts.

Twelve daily log-return series are computed based on the futures prices, one for each group of
contracts with same contract name (i.e. contracts that expire in the same month but possibly in
different years). Data on weekends and national holidays are excluded. To reduce data noise,
returns of contracts on the day right before and after expiry dates are removed.
Note that the original futures prices on each day are not sorted according to contract tenors.
To facilitate our analysis, we sort the prices and returns on each day according to contract
tenors in ascending order. In this case, we still have 12 daily log-return series. Now each of
them represents the return of futures with a fixed ranged of time to maturity. For example, the
first series consists of prompt month contracts (0-1 month maturity), while the second series
is made of the 2nd nearest to expiry month contracts (1-2 months maturity).

3. S EASONALITY
Like other commodity futures curves, the futures curve for natural gas has a pronounced
shape that incorporates seasonality. The market for natural gas is traded at various locations
throughout the United States. The most liquid of these locations is the Henry Hub in southern
Louisiana, this location acts as the major trading hub for the eastern United States while the
Waha location in Texas is the major trading hub for the western United States (Eydeland
and Wolyniec (2003)). The Henry Hub futures contract is traded on the NYMEX and is the
second-highest volume futures contract in the world based on a physical commodity.

Figure 1 plots the futures prices of January-December contracts from January 1, 2000 to
February 12, 2013. We can see a strong seasonal pattern. On any given day, the winter month
contracts usually have the highest prices among all outstanding contracts, regardless of their
1

The actual expiry dates are usually several days before the beginning of the month. For example, the February
contract usually has a expiry date around January 27.
2
http://www.cmegroup.com/trading/energy/natural-gas/natural-gas learn more.html

Y. CHONG, W. KONG, M. TOCHTERMAN, J. XU, AND H. YU

F IGURE 1. Futures Curve of Nature Gas

16
14

Price ($/mmBtu)

12
10
8
6
4
2
0
2014
2012
2010
2008
2006
2004
2002
2000

Jan

Feb

Mar

Apr

Date

May

Jun

Jul

Aug

Sep

Oct

Nov

Dec

Expiry Month

This figure plots the Henry Hub natural gas futures curve from January 2000 to February 2013.

tenors. According to Eydeland and Wolyniec (2003), this is due to the increased residential
consumption of natural gas during winter for home heating.
The relationship between weather and the seasonal pattern of the natural gas futures curve
is well documented by Eydeland and Wolyniec (2003). While Henry Hub natural gas prices
are highest during winter, seasonal natural gas prices at other points vary across the United
States. Prices for natural gas in colder climates (north-east, north-west, mid-west, etc) display
relatively higher natural gas prices in the winter months, however in the warmer climates the
price also rises in the summer months. In the states with a warmer climate the demand for
natural gas is also driven by electricity consumption for air-conditioning which is greater in
the summer months. In summary, this seasonality is driven by demand dynamics specific to
the climate in which the natural gas is delivered. Figure 2 shows the basis swap price for two
hubs: Algonquin City Gates and Southern California. In the natural gas market, basis is the
difference between the price of gas at one delivery point and gas at another. The standard
reference is the differential for gas at a specific hub to the price of gas quoted at Henry Hub

MODELING SEASONALITY IN COMMODITY PRICE DYNAMICS

F IGURE 2. Basis Swap Price for SoCal and Algonquin City Gates
Basis swap price for Socal and Algonquin City Gates pricing points as at 4Feb2013
0.3

6
SoCal (LHS)
Algonquin City Gates (RHS)

0.2

0.1

0
Feb13

Apr13

Jun13

Aug13

Oct13

Dec13

Feb14

0
Apr14

This figure plots the Basis swap price for SoCal and Algonquin City Gates of different maturities. The current
date is February 4, 2013. The unit is $/mmBtu. The prompt month contract prices appear to be driven by
trading particular to the month ahead contract.

via the NYMEX futures price. Algonquin City Gates is located in a colder climate where the
natural gas price is higher during winter. On the contrary, Southern California has a warmer
climate. Its natural gas price relative to Henry Hub peaks twice due to higher demand for
natural gas both in summer (as there is increased demand for natural gas due to an increase
in demand for electricity for air-conditioning) and again during winter when gas is used for
heating.
Aside from the seasonality directly observed in the term structure in the futures curve, the
evolution of the futures curve also displays seasonality. For example, returns of all outstanding
contracts may have a different correlation matrix for different months. Figure 3 shows the

Y. CHONG, W. KONG, M. TOCHTERMAN, J. XU, AND H. YU

F IGURE 3. Return Correlation During Different Months


Prompt Month Contract
1
2nd Nearest
3rd Nearest
4th Nearest
5th Nearest
6th Nearest
7th Nearest
8th Nearest
9th Nearest
10th Nearest
11th Nearest
12th Nearest

0.95

Correlation

0.9
0.85
0.8
0.75
0.7
0.65
0.6

Jan

Feb

Mar

Apr

May

Jun

Jul

Aug

Sep

Oct

Nov

Dec

2nd Nearest Month Contract


1
3rd Nearest
4th Nearest
5th Nearest
6th Nearest
7th Nearest
8th Nearest
9th Nearest
10th Nearest
11th Nearest
12th Nearest

0.95

Correlation

0.9
0.85
0.8
0.75
0.7
0.65
0.6

Jan

Feb

Mar

Apr

May

Jun

Jul

Aug

Sep

Oct

Nov

Dec

In the upper part of the figure, each line plots the correlation between prompt month futures contracts log return
and nth nearest month contracts log return across different months. The lower part plots the correlation between
2nd nearest month futures contract and longer dated contracts returns in each month of the year.

correlation between the futures return of short-dated and long-dated contracts across months.
We can clearly see that the correlation tends to be lower during the winter (around January)
and late summer (around September). A plausible explanation for the winter trough is that the
short-dated contract prices track the spot price closely. During winter months, the demand for
natural gas is largely driven by potentially extreme weather conditions. Therefore, changes in
current weather conditions stir the demand, which then causes the spot price and thus shortdated futures price to fluctuate. But the coldness of winter clearly has little impact on longerdated contracts, and serves only as an idiosyncratic factor to short-dated contracts rather than
a common factor to all contracts, lowering the correlation between short-dated and long-dated
contracts.
A similar explanation can be given for the September dip, albeit from a supply rather than
demand side. While the Atlantic hurricane season is officially from June 1 to November 30,
95% of all intense hurricane activity occurs during August to October (Landsea (1993)). Since

MODELING SEASONALITY IN COMMODITY PRICE DYNAMICS

F IGURE 4. Principal Component Loading During Different Months

PC Loading

Prompt Month Contract PC

2nd Nearest Month Contract PC

0.5

0.5

0.5

0.5
2

10

12

PC Loading

3rd Nearest Month Contract PC


1

0.5

0.5

0.5

0.5
4

6
8
Month

10

12

10

12

4th Nearest Month Contract PC

6
8
Month

10

12

This figure plots the factor loadings of the four shortest dated contracts in each month of the year. The blue line
denotes the coefficients of the first factor. The red line denotes the coefficients of the second factor. The black
line denotes the coefficients of the third factor.

severe hurricanes can disrupt the supply of natural gas, which in turn affects spot and shortdated futures prices but not long-dated futures prices, the correlation between short-dated and
long-dated contracts is expected to be low during hurricane seasons.
Another way to look at this is through the so called seasonal principal component analysis.
This method is proposed by Blanco and Stefiszyn (2002). In the seasonal PCA, the return
of futures contracts with different maturities are separated into 12 groups according to the
observation month. Normal PCA is then conducted within each group, where each individual
contracts return is treated as a variable.

10

Y. CHONG, W. KONG, M. TOCHTERMAN, J. XU, AND H. YU

Figure 4 shows the principal component loadings of the four shortest dated contracts in each
group. We can observe strong seasonal fluctuation patterns in the second (slope) and third (tilt)
components, and that the prompt month contract has an opposite pattern compared to longerdated contracts. For example, the loading on the second factor peaks around February and
September for the prompt month contract while it bottoms during the same periods for longerdated contracts. This implies that the differential (in return) between the these contracts varies
more than usual during winter and late summer, which would translate to a lower correlation
between these contracts. Thus, the observed pattern from seasonal PCA is consistent with the
time varying correlation recorded in Figure 3.
The different dimensions of seasonality embedded in the commodity futures curve have a
huge impact on our choice of model as illustrated in the next section.

4. M ODEL AND C ALIBRATION


Similar to modeling interest rate dynamics, there are two general approaches to modeling
commodity prices. The first approach models spot price dynamics directly. A good example
can be seen in Schwartz (1997). This is akin to a Hull-White type of interest rate model.
The second approach starts from the forward price dynamics and backs out the spot price
process. This second approach adopts the HJM framework. The multi-factor model proposed
by Clewlow and Strickland (1999a) falls into this category.
Models in both categories can be calibrated to current market observed forward curves
and option implied volatilities, although it is easier to do this with an HJM type of model.
Prices of many derivatives such as caps and American options can be obtained by Monte
Carlo simulations of the spot price process. We show in Section 5 that the valuation of certain
products, such as a gas storage facility is more complex since the generated cash flows depend
not only on the path of spot prices but also on the evolution of the entire forward curve. This
means a good model should accurately depict the correlation of prices of forward contracts

MODELING SEASONALITY IN COMMODITY PRICE DYNAMICS

11

F IGURE 5. Loading of PCA (2000-2013)


PCA by Time to Maturity
0.8
1st PC
2nd PC
3rd PC

0.6

PCA Loading

0.4

0.2

0.2

0.4

0.6

4
5
6
7
8
9
10
11
12
Time to Maturity (Months)
In this figure, we plot the factor loadings of the first three principal components of natural gas future returns.
PCA is performed over the entire sample space.

with different maturities. For this reason, the HJM type of model is more suitable since it is
relatively easier to model and estimate the covariance matrix of different forward contracts
under this framework. The industry standard multi-factor HJM model proposed by Clewlow
and Strickland (1999a) assumes that the price of a futures contract with maturity T has the
following dynamics:
n

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

(2)

where zi (t) are Brownian motions. Clewlow and Strickland (2000) suggest to set
i (t, T ) = i exp [i (T t)] .

(3)

12

Y. CHONG, W. KONG, M. TOCHTERMAN, J. XU, AND H. YU

The parameters i and i can be calibrated via principal component analysis (PCA) to best
match the historical covariance matrix of returns of futures contracts with different maturities.
Specifically, we first sort those contracts according to their time to maturity on each day. We
then group daily linear returns into 12 time-series according to their contract tenors. In other
words, futures contracts with tenors 0-1 month (or 1-2 months, 2-3 months, etc) are spliced
into one time series so that it represents daily return of futures with a fixed range (one-month)
of maturities. For example, from January 2, 2013 to January 31, 2013, the first time-series
consists of daily returns of the February 2013 contract, while the second time-series consists
of daily returns of the March 2013 contract, etc. On February 1, 2013, the first time-series
consists of daily returns of the March 2013 contract while the second time-series consists of
daily returns of the April 2013 contract. We then perform principal component analysis on
these 12 time-series. The first three loadings are plotted in Figure 5. The first two principal
components explain about 98% of the total variations of those 12 series. This suggests a twofactor HJM model is appropriate to model the correlation between different forward contracts.
We now estimate the parameters in the Clewlow-Strickland model using Nonlinear Least
Squares (NLS). Assuming the PCA loading of contracts with maturity w.r.t factor i is ,i ,
the variance of factor is score (i.e. the ith eigenvalue of the covariance matrix) is i , we have
the relationship
,i

i = i exp [i ( )] + ,

(4)

where the value of i and i can be estimated using NLS. Note that in each daily return series
the contract maturity actually varies within a one-month range. We set to be the mean of
these maturities for each series so that is equal to 0.5 month, 1.5 months, . . ., and 11.5
months respectively. Figure 6 compares the estimated HJM factor coefficients (i.e. the right
hand side of equation (4)) and PCA loading times (i.e. the left-hand side of equation (4)).
We can see that the first principal loading is fitted by an exponential form very well. However,
the exponential specification forces the coefficients of the second factor to have the same

MODELING SEASONALITY IN COMMODITY PRICE DYNAMICS

13

F IGURE 6. Clewlow-Strickland Coefficients VS PCA Factor Coefficients

0.04
PCA First Coeff
Fitted Value
0.03

0.02

0.01

0.2

0.4

0.6

0.8

15

x 10

PCA Second Coeff


Fitted Value

10
5
0
5

0.2

0.4
0.6
0.8
1
Year to Maturity
In this figure, we plot the coefficients of PCA factors and the estimated values of HJM coefficients.

sign, while the actual second PCA coefficients have positive sign for short-dated contracts but
negative for long-dated contracts. An easy way to remedy this is to add a constant term in
equation (3) for the second factor:
2 (t, T ) = c2 + 2 exp [2 (T t)] .

(5)

Equation (5) allows the factor coefficient to have different signs for contracts with different
tenors, which would allow a better fit to principal component loadings. Figure 7 shows the
modified Clewlow-Strickland factor exposures. We can see that the fit to the second factor
is significantly better than the original model. This is important as the second factor is the
main driver of the seasonality embedded in the time-varying correlation matrix, as shown

14

Y. CHONG, W. KONG, M. TOCHTERMAN, J. XU, AND H. YU

F IGURE 7. Modified Clewlow-Strickland Coefficients VS PCA Factor Coefficients

0.04
PCA First Coeff
Fitted Value
0.03

0.02

0.01

0.2

0.4

0.6

0.8

15

x 10

PCA Second Coeff


Fitted Value

10
5
0
5

0.2

0.4
0.6
0.8
1
Year to Maturity
In this figure, we plot the coefficients of PCA factors and the estimated values of modified Clewlow-Strickland
coefficients.

in Section 3. One of the main differences between the modified and original model lies at
the model derived forward volatility term structure. The long-end volatility implied by the
modified model does not converge to zero as maturity goes to infinity. Also, if the value of
c2 is large, the volatility may not decrease monotonically w.r.t to maturity. However, this
phenomenon is almost unobservable if the parameters are calibrated to real data. If the time
horizon for the product to be priced is not very long, the modified model tends to have a very
similar volatility term structure as the original version. Figure 8 compares the volatility term
structure implied by the modified Clewlew-Strickland model with that implied by the original
model(non-time varying). We can see that as long as the maturity is less than two years, the

MODELING SEASONALITY IN COMMODITY PRICE DYNAMICS

15

F IGURE 8. Modified VS Original Clewlow-Strickland Implied Forward


Volatility

0.5
Orig. CS Model
Mod. CS Model

0.45
0.4

Annual Volatility

0.35
0.3
0.25
0.2
0.15
0.1
0.05
0

20

40

60
80
100
120
Maturity (month)
In this figure, we compare the forward volatility implied by the original and modified version of two-factor
Clewlow-Strickland model. Model parameters are estimated by PCA during the entire sample period from January 2000 to February 2013.

two models have an almost identical volatility term structure. While the general framework of
Clewlow-Strickland seems reasonable, setting i and i as constants implicitly assumes that
the covariance matrix does not vary over time. Equation (6) shows the covariance between
returns of futures contracts with tenors and 0 :
Cov (r , r 0 ) =

2
X
i=1

i2 exp (i ( + 0 )) .

(6)

16

Y. CHONG, W. KONG, M. TOCHTERMAN, J. XU, AND H. YU

We can see that the covariance is independent of time. However, as we have shown in Section
3, the covariance matrix should vary across different seasons, with lower correlation between
short-dated and long-dated contracts during the winter and late summer.
In light of this, we modify equation (3) and make i and i be time variants:
i (t, T ) = ci,t + i,t exp [i,t (T t)] , i = 1, 2, c1,t = 0

(7)

where ci,t , i,t , and i,t are step functions w.r.t t and are constant within the same month of
different years. This means that for a given i, we have 12 different values of i,t and i,t . To
calibrate these time varying parameters, we resort to seasonal PCA, which was discussed in
Section 3.
Following the guide-line of seasonal PCA introduced by Blanco and Stefiszyn (2002), we
separate our whole sample into 12 groups by the observation month. In other words, futures
curves observed in January (or February, March, etc) are put into the same group. We can then
calibrate ci,t , i,t , and i,t to market historical prices by running PCA within each subgroup and
selecting the parameter values that best mimic estimated exposures to the first two principal
components.3 Figure 9 shows the first three factor loadings for each group.
Note that once we calibrate the model to market historical prices and return dynamics,
we have implicitly achieved a volatility term structure that is equal to the average historical
volatility term structure. However, this feature may not be desirable as the current volatility
level may be very different from the historical average. Following the practice of Clewlow
and Strickland (2000), we add a scaling factor to equation (7)
i (t, T ) = t {ci,t + i,t exp [i,t (T t)]} , i = 1, 2.

c1,t is always fixed to be 0

(8)

MODELING SEASONALITY IN COMMODITY PRICE DYNAMICS

17

F IGURE 9. Loading of Seasonal PCA (2000-2013)


Jan

Feb

Mar

6 8 10
Apr

6 8 10
May

6 8 10
Jul

6 8 10
Aug

6 8 10
Oct

6 8 10
Nov

8 10

8 10

6 8 10
Jun

6 8 10
Sep

6 8 10
Dec

8 10

In this figure, we plot the factor loadings of the first three seasonal principal components of natural gas future
returns. PCA is performed over the entire sample space.

Doing this allows us to change the level of model derived volatility while keeping the correlation matrix intact. We can use the least squares method to calibrate t such that the model derived volatility term structure best match the term structure of either recent historical volatility
or implied volatility. In this paper, we choose to calibrate t with recent historical volatilities.
Specifically, we estimate 12 different recent historical volatility term structures using futures
returns from each of the most recent 12 months. For example, if the current date is January
15, 2013, to calibrate t for February, we estimate the historical volatility term structure using
return data only in February, 2012. Figure 10 compares model derived and recent historical
volatility term structure for months January to April on January 2, 2013. Even though we can

18

Y. CHONG, W. KONG, M. TOCHTERMAN, J. XU, AND H. YU

F IGURE 10. Calibrate Volatility Term Structure


January

Feburary

Annual Vol

0.8

0.8
Model Vol
Historical Vol

0.6

0.6

0.4

0.4

0.2

0.2
2

10

12

March
0.8
Annual Vol

Model Vol
Historical Vol

10

12

April
0.8

Model Vol
Historical Vol

0.6

0.6

0.4

0.4

0.2

0.2

Model Vol
Historical Vol

4
6
8
10 12
2
4
6
8
10 12
Maturity (month)
Maturity (month)
In this figure, we compare the model calculated volatility term structure and actual historical volatility term
structure in the past year, for month January to April.

only adjust one parameter, the variance term structure is well matched. This happens because
the shape of the volatility term structure remains relatively unchanged across our sample. Table 1 records all the estimated coefficients.
One remaining question is whether our modified model can reasonably capture the seasonality embedded in the evolution of the futures curve; that the correlation between short-dated
and long-dated futures returns are relatively low in the winter and late summer. The standard
Clewlow-Strickland model implies a constant correlation across different months. By allowing the model parameters to change over time, the model implied correlations now fluctuate
throughout the year. Figures 11 and 12 compare the historical correlation with our modified

MODELING SEASONALITY IN COMMODITY PRICE DYNAMICS

19

F IGURE 11. Return Correlation During Different Months


Empirical Correlation
1
2nd Nearest
3rd Nearest
4th Nearest
5th Nearest
6th Nearest
7th Nearest
8th Nearest
9th Nearest
10th Nearest
11th Nearest
12th Nearest

0.95

Correlation

0.9
0.85
0.8
0.75
0.7
0.65
0.6
Jan

Feb

Mar

Apr

May

Jun

Jul

Aug

Sep

Oct

Nov

Dec

Model Implied Correlation


1
2nd Nearest
3rd Nearest
4th Nearest
5th Nearest
6th Nearest
7th Nearest
8th Nearest
9th Nearest
10th Nearest
11th Nearest
12th Nearest

0.95

Correlation

0.9
0.85
0.8
0.75
0.7
0.65
0.6
Jan

Feb

Mar

Apr

May

Jun

Jul

Aug

Sep

Oct

Nov

Dec

In the upper part of the figure, each line plots the historical correlation between prompt month futures contracts log return and nth nearest month contracts log return across different months. The lower part plots the
corresponding model implied correlation.

TABLE 1. Calibrated Coefficients for Modified Clewlow-Strickland Model


Month
Jan
Feb
Mar
Apr
May
Jun
Jul
Aug
Sep
Oct
Nov
Dec

1
0.0361
0.0302
0.0280
0.0269
0.0289
0.0321
0.0310
0.0316
0.0308
0.0307
0.0318
0.0368

1
0.8094
0.7126
0.5996
0.6800
0.7245
0.7560
0.8912
0.9065
0.8282
0.7650
0.6900
0.9749

c2
0.0039
0.0018
0.0017
0.0048
0.0059
0.0073
0.0068
0.0036
0.0030
0.0029
0.5481
0.0057

2
0.0256
0.1168
0.0330
0.0115
0.0147
0.0168
0.0160
0.0228
0.0384
0.0296
0.5423
0.0262

2
8.6614
44.0933
21.1076
2.4091
2.6210
2.3695
2.5419
8.4814
17.1680
13.4304
0.0241
6.2380

1.2758
1.1563
0.8330
0.9173
1.1321
1.1365
1.0079
0.9239
0.8387
0.7560
0.8597
0.7221

In this table, we report the estimated coefficients in a modified Clewlow-Strickland model. The current observation date is January 2, 2013. The value of c1 is fixed at 0 in all specifications.

20

Y. CHONG, W. KONG, M. TOCHTERMAN, J. XU, AND H. YU

F IGURE 12. Return Correlation During Different Months


Empirical Correlation
1
3rd Nearest
4th Nearest
5th Nearest
6th Nearest
7th Nearest
8th Nearest
9th Nearest
10th Nearest
11th Nearest
12th Nearest

0.95

Correlation

0.9
0.85
0.8
0.75
0.7
0.65
0.6
Jan

Feb

Mar

Apr

May

Jun

Jul

Aug

Sep

Oct

Nov

Dec

Model Implied Correlation


1
3rd Nearest
4th Nearest
5th Nearest
6th Nearest
7th Nearest
8th Nearest
9th Nearest
10th Nearest
11th Nearest
12th Nearest

0.95

Correlation

0.9
0.85
0.8
0.75
0.7
0.65
0.6
Jan

Feb

Mar

Apr

May

Jun

Jul

Aug

Sep

Oct

Nov

Dec

In the upper part of the figure, each line plots the historical correlation between 2nd nearest month futures
contracts log return and nth nearest month contracts log return across different months. The lower part plots
the corresponding model implied correlation.

model implied correlation. As can be seen, the model implied correlation mimics the historical
correlation, with lower value during the winter and late summer. This is not surprising, as the
modified model is estimated using seasonal PCA. As we have shown in Section 3, the loading of the 2nd principal component displays seasonality in a way consistent with the seasonal
pattern of correlation fluctuations.

5. S TORAGE VALUATION P ROBLEM


5.1. Problem Description. Thus far, we have proposed and estimated models that captures
the seasonality of natural gas. In this section, we apply our model to the storage valuation
problem.
The existence of seasonality in the natural gas futures curve term structure implies arbitrage
opportunities for gas storage operators. To construct an arbitrage, we want to find the contango
section of the natural gas forward curve, then long the shorter-term forward/futures contract,

MODELING SEASONALITY IN COMMODITY PRICE DYNAMICS

21

short the longer-term forward/futures contract and store the natural gas for the relevant period.
For example, on October 23, 2009, the Henry Hub natural gas futures price for the Nov-09
contract was $4.787. On the same day, the futures price for the Feb-10 contract was $5.869.
We could long the Nov-09 contract and short same amount of Feb-10 contract, receive delivery
of natural gas in November, store the gas, and finally deliver in February at the agreed price.
Assuming that the storage cost is zero and the borrowing cost is $0.05, the dollar profit would
be $1.032 with an annualized log return of ln((5.869 0.05)/4.787) 4 = 78% during the
three months when we hold the gas.
In the example above, we assume that we always hold forward contracts to their maturities.
And we can call the $1.032 Intrinsic Value. In practice, however, this buy-and-hold strategy
may not be optimal. If the spread becomes smaller in the next period, it may be optimal to
unwind the initial positions and realize the gain immediately without further incurring the gas
injection/extraction cost and borrowing cost. For instance, if on October 26, the future price
for both the Nov-09 and Feb-10 contracts become $5.2, i.e., the calendar spread drops to 0, it
is better off for the arbitragers to unwind the initial positions by shorting the Nov-09 contract
and longing the Feb-10 contract. While the total dollar profit from trading forward contracts
would still be $1.082, the storage and borrowing costs are reduced. If the spread widens
instead in the next period, we can always do nothing but stick with the original buy-and-hold
strategy.
Thus, it is evident that the value of the gas storage depends not only on the current future
curve, but also on the entire evolution of the futures curve. Therefore, it is essential to have
a model that accounts for the seasonality of the futures curve dynamics when valuing a gas
storage facility. It is also clear that the valuation depends on the exact trading strategy used by
the storage operator. We provide three examples below showing alternative methods to value
the gas storage facility.

22

Y. CHONG, W. KONG, M. TOCHTERMAN, J. XU, AND H. YU

TABLE 2. Natural Gas Contract


Term
Maximum Working Gas Capacity
Initial Working Gas Capacity
Ending Working Gas Capacity
Maximum Injection Rate
Maximum Withdrawal Rate
Injection Cost
Withdrawal Cost
Fuel Injection Loss

1 year
1 Bcf
0 Bcf
0 Bcf
1/30 Bcf/day
1/30 Bcf/day
0.01 $/Bcf
0.01 $/Bcf
0%

In this table, we report the basic assumptions used in all three natural gas trading strategies.

5.2. Rolling Intrinsic Method. The dynamic trading strategy described in Section 5.1 is
called a rolling intrinsic strategy. In this case, the storage operator rebalances a portfolio
consisting of futures contracts and the physical gas storage facility daily. On each day, the
trader chooses to hold a portfolio that yields the highest profit if held to maturity.
The value of a gas storage facility for a trader with a rolling intrinsic strategy can be estimated using Monte Carlo simulation. First, we calibrate the parameters in equations (2) and
(7) using the method shown in Section 4. We then simulate many paths of futures curves
using the discretized version of equation (2). Figure 13 shows 20 cross-sectional samples of
simulated futures curves, with January 2, 2013 to be the initial date.
Once we have the simulated paths of futures curves, we can simulate the trading profit on
each path and take the mean of the discounted value of profit as the storage value. Some basic
assumptions about our trading strategy are listed in Table 2. Figure 14 reports the histogram
of profits from 10, 000 paths of simulated futures curves starting from January 2, 2013. The
corresponding value of the gas storage during year 2013 is $1.09 per mmBtu capacity. Notice
that there is a clear lower boundary for simulated profits. This boundary represents the intrinsic
value of the storage at the starting date, which is determined by the initial calendar spreads on
the futures curve.

MODELING SEASONALITY IN COMMODITY PRICE DYNAMICS

23

F IGURE 13. Cross-Sectional Samples of Simulated Forward Curves with


Seasonal PCA
20 Samples of Simulated Forward Curves after 10 Days
4.4
Actual Forward Curve in 10 Days
Current Forward Curve
Simulated Forward Curve in 10 Days

4.2
4
3.8

Price

3.6
3.4
3.2
3
2.8
2.6
Jan

Feb Mar

Apr May Jun Jul Aug Sep Oct Nov Dec


Contract Expiry Month
In this figure, we plot 20 cross-sectional samples of simulated forward curves after 10 business days. The HJM
model is fitted by seasonal PCA. The current observation date is 01/02/2013. The simulated forward curves are
supposed to mimic the forward curve to be observed on 01/17/2013.

5.3. Trinomial Tree Method. The rolling intrinsic method above is based on a trading strategy in the futures market. Alternatively, we can build a similar trading strategy in the spot
market and evaluate the storage using a trinomial tree method. At each time interval, we make
one of the three operational decisions - to inject, withdraw or retain the inventory level. If it
is optimal to inject at one point, we buy the natural gas at the spot market; if it is optional to
withdraw at one point, we sell the natural gas at the spot market.
Existing literature Mihaela (2004) discusses the tree/discretization models to price natural
gas storage using the spot price dynamics, which we briefly review below.

24

Y. CHONG, W. KONG, M. TOCHTERMAN, J. XU, AND H. YU

F IGURE 14. Profits From Rolling Intrinsic Trading Strategy


3500

Number of Occurances

3000

2500

2000

1500

1000

500

0.5

1.5
2
2.5
3
3.5
Actual Profits
In this figure, we plot the histogram of profits using rolling intrinsic trading strategy from 10000 paths of simulated futures curves with the strategy starting from January 2, 2013.

Equation (2) details the evolution of the futures curve. We note that any explicit form of
the futures price dynamics corresponds to a spot price process.4 For a one-factor Clewlow and
Strickland (1999a) model the log spot price is described as follows:

 1 2
ln F (0, t)
2
2t
dx(t) =
+ (ln F (0, t) x(t)) +
1e
dt + dz(t), (9)
t
4
2


where x(t) is the log of the spot price x(t) = ln(S(t)). We can simplify the formula as
dx(t) = [(t) x(t)] dt + dz(t),
4

Recall we assume a deterministic interest rate process. Under this condition, the futures price is equal to the
forward price.

MODELING SEASONALITY IN COMMODITY PRICE DYNAMICS

25

where

(t) =


 1 2
ln F (0, t)
2
2t
.
+ ln F (0, t) +
1e
t
4
2

Similar to the Hull-White short rate model, discretization is adopted according to the dynamics of the log of the spot price. The tree discretization involves two steps.
In the first step the (t) parameter is assumed to be zero, leaving the process as:
d
x(t) =
x(t) dt + dz(t).
Forward induction is used to obtain the tree of state prices and the closed-form result for
up, flat and down probabilities as derived by Clewlow and Strickland (1999b):
pu,i,j



1 2 t + 2 x2i,j t2
xi,j t
2
=
+ (k j)
(1 2(k j)) (k j) ,
2
x2
x

pd,i,j



1 2 t + 2 x2i,j t2
xi,j t
2
=
+ (k j) +
(1 + 2(k j)) + (k j)
2
x2
x
pm,i,j = 1 pu,i,j pd,i,j

Secondly, (t) is added back to the process as a calibration term, making the tree fit the
observed forward prices. The probability remains unchanged in this step.
In our study, we allow i and i to be time varying over the 12 different months. This
is theoretically valid as the log spot price process in equation (9) still holds piecewise (12
individual values of i and i for each month) and the time step of the discretization remains
equally spaced.

For the space step, Hull and white (1993) recommend a discretization x = 3t. Now
with the time-varying , x is also time-varying, meaning that with the time-varying , the
tree may not recombine. Thus, we make the following adjustment:

x = max(t ) 3t.

26

Y. CHONG, W. KONG, M. TOCHTERMAN, J. XU, AND H. YU

This allows the space step to remain constant over time. With a constant time step and a
constant space step, the tree will recombine. The probabilities of up, down and flat moves will
differ from the original model with constant i and i . The differing probabilities will better
model the seasonality of the evolution of natural gas forward curve, i.e. the covariance matrix
should vary across different seasons.
After constructing the discretization framework, we utilize the same methodology as Clewlow
and Strickland (1999b) to price the storage: the backward induction technique appears in three
dimensions: (1) time; (2) space; and (3) volume. For each loop, we compare values under each
of the three possible actions respectively - to hold the same inventory, to inject to the maximum quantity and to withdraw to the maximum quantity. In this way, we can finally derive
the storage value under the optimized set of actions at each node on the lattice.
Specifically we construct the tree under the assumption that injection is done to full capacity,
therefore we have two final sets of volume nodes. Full capacity to be withdrawn at the relevant
spot prices on the tree; and (2) empty capacity.
We then step back through the tree determining if the maximum value for the storage facility is obtained by either: (1)holding the current level of inventory; (2)injecting gas; or (3)
withdrawing gas. Given this optimal decision, we can then calculate the value of storage
facility.
Intuitively, if at time t = 0 we are in a winter month the decision will be to hold zero
inventory until the gas price is lower at which point the algorithm should buy and then sell
when the gas price is higher. We observe that the value of the storage facility using this method
in our example over 12 months does this and yields a value very similar to the intrinsic value.
These values are listed in Table 3. In Appendix A we provide an explanation of the mechanics
of the trinomial tree method by using a simple three-period example.
5.4. Least-Square Monte Carlo Method. As previously discussed the rolling intrinsic method
is based on trading futures contracts while the trinomial tree method is based on trading in the

MODELING SEASONALITY IN COMMODITY PRICE DYNAMICS

27

spot market. Each of these two methods has its own limitations. The rolling intrinsic method
selects the optimal futures spreads conditional on the futures dynamics. However, it may
not be able to fully account for the extrinsic value of the storage. The trinomial tree method
assumes no trading in the futures market, and therefore may not fully capitalize on the available arbitrage opportunities. To compensate for these limitations, we adopt a third method,
the Least Squares Monte Carlo simulation (LSMC) method, to account for both multi-factor
futures dynamics and optionality. We do not claim that this third method is superior to the
former methods as we are fully aware of the convergence issues in implementing LSMC.
The LSMC method was first introduced by Longstaff and Schwartz (2001) to price American options on stocks. Boogert and Jong (2008) further modified the LSMC method to value a
natural gas storage facility by simulating the dynamics of the spot price as a one-factor meanreverting process. Below we briefly explain the rationale of using the LSMC method to value
a storage facility, and then describe how it is implemented. The optimization problem is set
up so that at each point in time, the storage holder makes a decision from several operational
choices (for example, inject v1 , hold, withdraw v2 , etc.), with the objective of maximizing
the payoff. The payoff is then calculated as the sum of the discounted expected future values
(the continuation values):
P V (t) =

max

CF (t, F (t), v) +

vD(t,v(t))

DF (t, t + 1)E [P V (t + 1, F (t + 1), v(t + 1)|F (t), v(t))] .


The difficulty lies in computing the expectation of the future present value. The LSMC method
is a regression based simulation method which approximates the continuation value by a linear
combination of the basis functions of current state variables. In our paper, a state can be
uniquely defined by the time t, the inventory level v(t) and the price of a set of futures contract
F (t). We define continuation value C(t, v(t + 1), X(t)) as the value attached to the inventory

28

Y. CHONG, W. KONG, M. TOCHTERMAN, J. XU, AND H. YU

level v(t + 1) at time t + 1 evaluated at time t. This value is a linear combination of the basis
function j of futures prices X(t) at time t. Therefore, continuation value can be expressed
in the following form:
E [P V (t + 1, F (t + 1), v(t + 1)|F (t) = X(t))] = C (t, v(t + 1), X(t))
=

nBasis
X

aj (t, v(t + 1)) j (X)

At time t, for a certain level of inventory, the investor evaluates all the feasible actions D(t, v(t))
and chooses a strategy (t, v(t), X(t)) which delivers the highest payoff:
(t, v(t), X(t)) = arg

max

CF (t, F (t), v) + C (t, v(t) + v, X(t))

vD(t,v(t))

where CF (t, F (t), v) is the net cash inflow/outflow due to change in the inventory level v.
In our paper, the cash flow is calculated as
CF (t, F (t), v) = F 0 (t) v
where F 0 (t) is the prompt month futures price.
For instance, in the case of three feasible actions (inject at v1 , hold, withdraw at v2 ), if
the top path gives the largest payoff given X(t), then the optimal strategy is (t, v(t), X(t)) =
v1 .

CF (
v(t)

)
, v 1
)
t
(
F
t,
0

v(t + 1) = v(t) + v1

v(t + 1) = v(t)

C F (t

, F (t)
, v

2)

v(t + 1) = v(t) + v2
Time = t

Time = t+1

MODELING SEASONALITY IN COMMODITY PRICE DYNAMICS

29

In this paper, we apply this method to value the storage by incorporating the underlying
futures price, which is different from the traditional LSMC method which uses backward
induction only. We follow a two-phase valuation algorithm similar to Mantel, Peterson, and
Turner (2007) which consists of a backward induction phase followed by a forward induction
phase. We provide further details on this method with the forward and backward induction
phases described below.
Phase 1 - Backward Induction

(1) Time/Volume Discretization: Both volume and time are discretized into a fine grid
(nStepsnV olumes). The number of time steps is denoted as nSteps and the number
of volume steps is denoted as nV olumes. By discretizing the volume, we introduce
additional choices for injection and withdrawal decisions.
(2) As performed for the rolling intrinsic method, we simulate a number of independent
n
). X is a vector containing the futures prices of
sample futures paths (X0n , ..., XnSteps

the remaining time steps observed at the current time step. For instance, at tm , Xm =
n
] is a vector of active forward price whose maturity coincides with the
[Fmn , ..., FnSteps

remaining time steps.


(3) NPV Initialization: Assign continuation value P VnSteps+1 (F, v) = 0 for all simulations and volume levels at tnSteps .
(4) Regression: A regression is then done between the continuation value C(tm , v(tm+1 ), Xm )
and Xm .
(5) Update NPV: The strategy with largest payoff is chosen and the NPV at tm , with volume level v(tm ) and current futures price X(tm ) can be calculated using the following
equation:
P V (tm , X(tm ), v(tm )) =

max

CF (tm , F (tm ), v) + C(tm , v(tm ) + v, X(tm ))

vD(t,v(t))

30

Y. CHONG, W. KONG, M. TOCHTERMAN, J. XU, AND H. YU

In this way, at t = tm , for each volume level v(tm ), the NPV is updated and a decision
rule (tm , v(tm ), X(tm )) is determined.
(6) Step (4) and Step (5) are repeated recursively backward through the grid until we reach
t = 1.
After steps (1) through (6) have been completed we obtain the coefficient aj (t, v(t))
for each time and volume step. These coefficients are then used to determine the
optimal operation policy in the forward induction phrase.

Phase 2 - Forward Induction


In this phase, we step through the time-volume space to maximize profit by choosing the
optimal operational strategy at each time step.

(1) Futures Dynamic Simulation: A different set of futures curves is simulated to reduce
the valuation bias (Tavella (2002)).
(2) Injection/Withdrawal Decision: For each simulation path, given the current volume
level v(tm ), an optimal strategy is chosen among the feasible choices to maximize the
NPV given current information X(tm ). This step utilizes the same method as the one
in Backward Induction Step (5).
(3) Once we finalize the best decision, we move to the node in the next period with
v(tm+1 ) = v(tm ) + (tm , v(tm ), X(tm )), where v(tm+1 ) is the inventory in the next
period after optimal operational policy is adopted. The discounted cash flow is updated
to incorporate the volume change from tm to tm+1 .
(4) Repeat Step (2) and Step (3) until the last period is reached.
(5) The NPV of the storage is calculated as the average of all the simulated paths.
The forward induction effectively applies the optimal operation policy (based on the
rules determined in the backward phrase) and determines the corresponding inventory

MODELING SEASONALITY IN COMMODITY PRICE DYNAMICS

31

F IGURE 15. Profits from LSMC Strategy


Actual Profit (LSMC)

Number of Occurances

1500

1000

500

0
0.8

0.9

1.1

1.2

1.3
1.4
Actual Profits

1.5

1.6

1.7

1.8

Notes: This histogram shows LSMC strategy profit distribution on 10, 000 paths of simulated futures curves with
the strategy starting from January 2, 2013.

level throughout the time period for all the simulation paths. All the cash flows incurred during the operation are discounted back to determine the NPV of the storage
facility.
Figure 15 plots the histogram of profits from 10, 000 paths of simulated futures
curves with the strategy starting from January 2, 2013. The corresponding value of the
gas storage during year 2013 is $1.21 per mmBtu capacity.
5.5. Comparison Among Three Valuation Methods. We calibrate the two-factor futures
price dynamics specified in equations (2) and (7) to Henry Hub natural gas futures curve on
12 different dates. These dates are set to be the first business day of each month from February

32

Y. CHONG, W. KONG, M. TOCHTERMAN, J. XU, AND H. YU

2012 to January 2013. Based on the estimated model, we evaluate an one-year ownership of
a gas storage with capacity 1 mmBtu, using three valuation methods introduced above.5 Table
3 summarizes the values of storage with different starting dates.
TABLE 3. Storage Value (1 Year Tenor)
Starting Date Rolling Intrinsic Rolling Intrinsic Trinomial Tree
LSMC
(Daily Rbl.)
(Monthly Rbl.) (Monthly Rbl.) (Monthly Rbl.)
Feb-12
1.0515
0.9793
0.8699
1.1188
Mar-12
1.0326
0.9485
0.8027
1.0750
Apr-12
1.5151
1.2349
1.0211
1.5327
May-12
1.3845
1.0539
0.8241
1.3974
Jun-12
1.3373
1.0211
0.7910
1.3605
Jul-12
1.1682
0.8023
0.6234
1.1993
Aug-12
1.0737
0.6568
0.5006
1.1026
Sep-12
1.1124
0.7447
0.5256
1.1468
Oct-12
1.0931
0.6613
0.4498
1.1214
Nov-12
0.9498
0.4874
0.3318
1.0169
Dec-12
1.0052
0.5630
0.5133
1.1094
Jan-13
1.0928
0.7115
0.6129
1.2056
In this table, we report the estimated value of a natural gas storage per year per mmBtu capacity based on different
trading strategies and valuation methods. Each row of the table assumes a distinct initial trading date. These dates
are set to be the first business day of each month.

We observe that the value of the storage facility calculated using the rolling intrinsic method
is materially greater for daily rebalancing when compared to monthly rebalancing. This is expected and akin to the difference between an American option and a Bermudan option with a
significantly limited exercise window. As we increase the rebalancing frequency, more optionality value is captured by the trading strategy. Similarly, the monthly rolling intrinsic strategy
always yields a higher value than storage intrinsic value, which is equal to the buy-and-hold
value of the optimal futures contracts with no rebalance.
The trinomial tree method produces the lowest value for the storage facility. With this
method we use the dynamics at a particular point in time and construct the tree based on one
5

For the trinomial tree method, we base our evaluation on an one-factor model due to the relative convenience to
link futures dynamics to spot price dynamics under such a setting.

MODELING SEASONALITY IN COMMODITY PRICE DYNAMICS

33

forward curve. The trading strategy for this method is more constrained with trading only in
the spot market with a simple monthly decision to inject or withdraw gas. These limitations
yield a value that incorporates less extrinsic value and appears to be very similar to the intrinsic
value of the storage facility.
LSMC produces the highest value for the storage facility as the optionality is more fully considered. In the backward induction phase, the global optimal decision policy is constructed by
considering the evolution of the forward curve. In the forward induction phase, given the current forward curve, the optimal operational decision is made to maximize the overall profit.
Given these factors we expect the value of storage facility using LSMC to be greater than
that of the Rolling Intrinsic method with the same rebalancing frequency. As the rebalancing
frequency of the Rolling Intrinsic method increases, we expect this value to more closely resemble that of the LSMC method. We observe this in the results of the rolling intrinsic method;
when we change the rebalancing from monthly to daily, the result more closely matches that
of the LSMC method.
As detailed in Table 3, we observe that the value of the storage facility displays seasonality
in that for all three methods the value of the facility is highest in April and lowest in November. This reflects the shape of the forward curve at these times. We capture the additional
optionality in April as there is the maximum number of summer months to inject when the
price of gas is lowest and then we have the maximum number of winter months in which the
price of gas is highest to withdraw. In November, the forward curve will be reaching its peak,
meaning that if a market participant entered into the contract in November they would have
to buy at a higher price or buy in the summer months and then have fewer opportunities after
the summer to withdraw and achieve a maximum profit. For clarity, we plot the value of the
storage facility by month for each method in Figure 16.
5.6. Market Implication. While the results in Section 5.5 suggest the existence of arbitrage
opportunities in both the future and spot market of natural gas, we have not considered the

34

Y. CHONG, W. KONG, M. TOCHTERMAN, J. XU, AND H. YU

F IGURE 16. Storage Value Simulation Result


Simulation Result
1.6
Rolling Intrinsic Daily Reblancing
Rolling Intrinsic Month Reblancing
Trinomial Tree Monthly Reblancing
LSMC Monthly Reblancing
Intrinsic Value

1.4

Storage Value ($/MMBtu)

1.2

0.8

0.6

0.4

0.2

Feb12

Mar12

Apr12

May12

Jun12

Jul12

Aug12
Starting Date

Sep12

Oct12

Nov12

Dec12

Jan13

In this figure, we plot the estimated value of a natural gas storage per year per mmBtu capacity based on different
trading strategies and valuation methods. Each set of bars corresponds to a distinct initial trading date. These
dates are set to be the first business day of each month.

TABLE 4. Gas Storage Facility Operations


Type

Cushion to Working Injection Period Withdrawal Period


Gas Ratio
(Days)
(Days)
Aquifer
Cushion 50% to 80%
200 to 250
100 to 150
Depleted Oil/Gas Reservoirs
Cushion 50%
200 to 250
100 to 150
Salt Cavern
Cushion 20% to 30%
20 to 40
10 to 20
Source: Federal Energy Regulatory Commission

fixed cost of owning a natural gas storage. If we view such a cost as an investment, the
arbitrage profits generated by trading natural gas futures/spot is the payoff of the investment.
It is interesting to see if a natural gas storage is a fair investment, in the sense that whether the
investment return exceeds risk free rate under the risk neutral measure. In order to answer this
question, we need to evaluate the cost of developing a gas storage first. There are three main
types of gas storages available. Table 4 compares their main features.

MODELING SEASONALITY IN COMMODITY PRICE DYNAMICS

35

TABLE 5. Development Cost of Working Gas Storage Facility


Type

Development Costs
Per Bcf of Working
Gas Capacity
$5 - $6 million

2-Cycle Reservoir
6-to-12 Cycle Salt Cavern
Gulf Coast
$10 - $12 million
Northeast and West
As much as $25 million
Source: Federal Energy Regulatory Commission

The salt cavern is the choice for most natural gas seasonality arbitragers. This is mainly
because it has the highest gas injection and extraction speed, ensuring storage operators the
ability to timely cycle the natural gas according to the optimal holding determined by their
trading strategies. Therefore, the cost of developing/owning a salt cavern gas storage is the
best representation of the fixed cost of arbitrage opportunities in the natural gas market.
Among all gas storage types, the salt cavern appears to be the most costly to build. Table 5
summarizes the total development cost of different gas storage facilities. The depleted reservoirs usually cost 5-6 million $/Bcf of working gas capacity to develop. In the Gulf Coast,
a 6-12 cycle salt cavern normally costs 10-12 million$/Bcf. In other regions, the cost ranges
from 10 million $/Bcf to 25 million $/Bcf.
While the storage facility owner expects a risk adjusted return in the real world, under our
risk-neutral pricing framework, the storage facility owner would expect to earn a risk free rate
return. We use the 30-year US treasury rate to approx the risk-free rate, which is 3.22% as
of February 12, 2013. Hence, the annual expected cost of the storage facility under the risk

36

Y. CHONG, W. KONG, M. TOCHTERMAN, J. XU, AND H. YU

neutral measure is expected to be:


Annual expected cost = total storage cost risk free rate
' [10, 25]million$ per Bcf 3.22%
' [9.74, 24.34]$ per mmBtu 3.22%(Note: 1 Bcf = 1.027 million mmBtu)
' [0.3, 0.73]$ per mmBtu.
Although the costs of different storage facilities vary based on different specifications, the
values above provide us with a good practical benchmark. We note that our values of the gas
storage facility (per year per mmBtu) are in general higher than risk free rate of return (on the
upper bound of required initial investment). However, these profits can only be guaranteed
if the storage operator can fully hedge his trading position. Also, the high fixed costs and
initial infrastructure outlays would also provide a significant barrier to entry. Nevertheless,
our analysis indicates that owning or developing a natural gas storage facility appears to be a
good investment.

6. C ONCLUSION
The Clewlow-Strickland model of commodity futures price does not allow its parameters
to vary over time. As a result, the model implied correlation matrix between contracts of
different maturities remains constant throughout the year. However, we show in this paper that
the historical correlation between short and long dated natural gas contracts is lower during
the winter and late summer. This is likely caused by spot market supply and demand shocks,
which are induced by idiosyncratic fluctuations in extreme weather conditions. In light of
this, we extend the Clewlow-Strickland model and let its parameters to take different values
during different months. We explain how to estimate and calibrate the extended model to

MODELING SEASONALITY IN COMMODITY PRICE DYNAMICS

37

market data. We find that our model manages to capture the seasonal fluctuation in correlation
between different futures contract reasonably well.
Based on our modified model, we proceed to evaluate a natural gas storage facility assuming
the investor using one of the three different trading strategies. We find that the value of storage
per year per mmBtu capacity exceeds the risk-free return on the initial investment required to
own a salt cavern gas storage facility.

38

Y. CHONG, W. KONG, M. TOCHTERMAN, J. XU, AND H. YU

R EFERENCES
B ESSEMBINDER , H., J. C OUGHENOUR , P. S EGUIN , AND M. S MOLLER (1996): Is There a Term Structure of
Futures Volatilities? Reevaluating the Samuelson Hypothesis, The Journal of Derivatives, 4(2), 4558.
B LANCO , C., AND P. S TEFISZYN (2002): Valuing Natural Gas Storage Using Seasonal Principal Component
Analysis, The Risk Desk, 2(3).
B OOGERT, A., AND C. D. J ONG (2008): Gas Storage Valuation Using a Monte Carlo Method, The Journal
of Derivatives, 15(3), 8198.
C LEWLOW, L., AND C. S TRICKLAND (1999a): A Multi-Factor Model for Energy Derivatives Risk Management, Working Paper.
(1999b): Valuing Energy Options in a One Factor Model Fitted to Energy Prices, Working Paper.
(2000): Energy Derivatives: Pricing and Risk Management. London: Lacima Publications.
C ORTAZAR , G., AND E. S CHWARTZ (1994): The Valuation of Commodity Contingent Claims, The Journal
of Derivatives, 1(4), 2739.
E YDELAND , A., AND K. W OLYNIEC (2003): Energy and Power Risk Management: New Developments in
Modeling, Pricing, and Hedging. Hoboken: John Wiley & Sons.
G IBSON , R., AND E. S CHWARTZ (1990): Stochastic Convenience Yield and the Pricing of Oil Contingent
Claims, Journal of Finance, 45(3), 959976.
H EATH , D., R. JARROW, AND A. M ORTON (1992): Bond Pricing and the Term Structure of Interest Rates: A
New Methodology for Contingent Claims Valuations, Econometrica, 60(1), 77105.
H ULL , J., AND A. W HITE (1990): Pricing Interest-Rate Derivative Securities, The Review of Financial Studies, 3(4), 573592.
L ANDSEA , C. (1993): A Climatology of Intense (or Major) Atlantic Hurricanes, Monthly Weather Review,
121, 17031713.
L ONGSTAFF , F., AND E. S CHWARTZ (2001): Valuing American Options by Simulation: A Simple LeastSquares Approach. The Review of Financial Studies, The Review of Financial Studies, 14(1), 113147.
M ANTEL , J., D. P ETERSON , AND M. T URNER (2007): Gas Storage Valuation Models, Working Paper.
M IHAELA , M. (2004): Storage Options Valuation Using Multilevel Trees and Calendar Spreads, International
Journal of Theoretical and Applied Finance, 07(04).
S CHWARTZ , E. (1982): The Pricing of Commodity-Linked Bonds, Journal of Finance, 37(2), 289300.

MODELING SEASONALITY IN COMMODITY PRICE DYNAMICS

39

(1997): The Stochastic Behavior of Commodity Prices: Implications for Valuation and Hedging, The
Journal of Finance, 52(3), 923973.
TAVELLA , D. (2002): Quantitative Methods in Derivative Pricing: An Introduction to Computational Finance.
Wiley.

40

Y. CHONG, W. KONG, M. TOCHTERMAN, J. XU, AND H. YU

A PPENDIX A. A N E XAMPLE OF T RINOMIAL T REE


We provide an example of the mechanics of the trinomial tree method for valuing the storage facility. This example is based on the trinomial tree method provided by Clewlow and
Strickland (1999b), extended by Mihaela (2004) for the storage valuation problem.
In this example we have four time periods where the natural gas price is used for decision
making in terms of injection or withdrawals from the storage facility6.
We build the trinomial tree and have the spot price of natural gas, state prices and probabilities each move. From here we work back recursively through the tree and explain the
methodology for this below for a four time period tree.
TABLE 6. Stock price tree for four period trinomial model
T=0

T=1

T=2

1
2
4.9134
3
3.6339 3.6479
4 2.737 2.6979 2.7083
5
2.003 2.0107
6
1.4928
7

T=3
7.1526
5.3103
3.9426
2.9271
2.1732
1.6134
1.1979

Using the methodology of Mihaela (2004) we have three decisions to make:


(1) Hold decision value
(2) Inject decision value
(3) Withdrawal decision value
Looking at t = 3, there is no value to holding the same inventory or injecting and therefore
withdraw the inventory at this level. We have a forest with different trees for different quantities for a specified discretization. The final sets of nodes will have a number of different
trees representing different levels of volume. At the terminal node (t = 3) we will have five
6

We gratefully acknowledge the assistance provided by Nancy Wallance and Paulo Issler for providing valuable
Matlab codes.

MODELING SEASONALITY IN COMMODITY PRICE DYNAMICS

41

sets of trees in the forest for volumes of 0, 0.25, 0.5, 0.75 and 1.0 and calculate the value of
withdrawing this inventory.
We then compare the value of injecting, withdrawing and holding the same level of inventory for each node recursively through the tree and at each of these nodes we take the
maximum value.
For example, given the stock price tree above, at the final node the maximum values are all
for withdrawal of gas. It leads to the storage valuations in Table 7.
TABLE 7. Storage Value
T = 3 Z = 0 Z = 0.25 Z = 0.50 Z = 0.75 Z = 1.00
1
0
1.7631
3.5263
5.2894
7.0526
2
0
1.3026
2.6052
3.9077
5.2103
3
0
0.9606
1.9213
2.8819
3.8426
4
0
0.7068
1.4135
2.1203
2.8271
5
0
0.5183
1.0366
1.5549
2.0732
6
0
0.3784
0.7567
1.1351
1.5134
7
0
0.2745
0.5489
0.8234
1.0979

We then work backwards through the tree. Keeping with the same example above we then
evaluate the decision at t = 2. For Z = 0 for example, we evaluate the tree at the point where
the spot price is 4.9134. Here we need to look at the tree and determine how we can derive the
most value from the storage facility. At this point Z = 0, we have no inventory in the storage
and will either inject or keep holding no inventory. If we inject to the maximum capacity of the
facility, the value of the storage in the next period is calculated to be: Pu Vu +Pm Vm +Pd
Vd , In this example, the value would be: 0.05757.053+0.75405.2103+0.18853.8426 =
5.0583. If we were to inject at this level, the cost would be: 4.9134 + 0.1000 = 5.0134, giving
a value of the storage facility as 0.0421. We do this for each value of the quantity of storage
in the facility. This is detailed in Table 8.
The same iteration is then done for time T = 1 which yields Table 9.

42

Y. CHONG, W. KONG, M. TOCHTERMAN, J. XU, AND H. YU

TABLE 8. Storage Value


T=2
1
2
3
4
5
6
7

Z=0
0.0421
0.0555
0.0471
0.0269
0.0014

Z = 0.25 Z = 0.50 Z = 0.75 Z = 1.00


1.2954
0.9925
0.7492
0.5546
0.3996

2.5488
1.9295
1.4513
1.0823
0.7978

3.8021
2.8664
2.1533
1.61
1.196

5.0555
3.8034
2.8554
2.1377
1.5942

TABLE 9. Storage Value


T = 1 Z = 0 Z = 0.25 Z = 0.50 Z = 0.75 Z = 1.00
1
2
3 0.0529
0.9789
1.9048
2.8308
3.7567
4 0.0845
0.7839
1.4834
2.1829
2.8824
5 0.1032
0.629
1.1547
1.6805
2.2062
6
7
At time T = 0, we then look at the options for the position where the level of gas in the
storage facility is equal to 0 (Z = 0). We have the following options:
(1) Hold the same level of inventory Z = 0. At this level the value of the storage is the
continuation value of the level of storage: Pu Vu + Pm Vm + Pd Vd . Which is:
0.1248 0.0529 + 0.7504 0.0845 + 0.1248 0.1032 = 0.0828.

(2) We can inject gas into the facility. If we do this, we calculate the value as the probabilityweighted value of a full facility in the next time period: Pu Vu + Pm Vm + Pd Vd .
This is: 0.1248 3.7567 + 0.7504 2.8824 + 0.1248 2.2062 = 2.9071. We then
calculate the cost of injecting gas at T = 0: 2.7370 + 0.1000 = 2.8370. The value if
we inject gas at this node is then: 2.9071 2.8370 = 0.0693.

MODELING SEASONALITY IN COMMODITY PRICE DYNAMICS

43

(3) Another option would be to withdraw if the amount in storage was greater than zero.
The maximum value is 0.0828 which becomes the value of the storage facility at time T = 0.
(Y. Chong) H AAS S CHOOL OF B USINESS , U NIVERSITY OF C ALIFORNIA , B ERKELEY, USA.
E-mail address: yanping chong@mfe.berkeley.edu

(W. Kong) H AAS S CHOOL OF B USINESS , U NIVERSITY OF C ALIFORNIA , B ERKELEY, USA.


E-mail address: weixian kong@mfe.berkeley.edu

(M. Tochterman) H AAS S CHOOL OF B USINESS , U NIVERSITY OF C ALIFORNIA , B ERKELEY, USA.


E-mail address: Matthew Tochterman@mfe.berkeley.edu

(J. Xu) H AAS S CHOOL OF B USINESS , U NIVERSITY OF C ALIFORNIA , B ERKELEY, USA.


E-mail address: junyue xu@mfe.berkeley.edu

(H. Yu) H AAS S CHOOL OF B USINESS , U NIVERSITY OF C ALIFORNIA , B ERKELEY, USA.


E-mail address: hangying yu@mfe.berkeley.edu

You might also like