Energy Economics 50 (2015) 207214

Contents lists available at ScienceDirect

Energy Economics

journal homepage: www.elsevier.com/locate/eneco

A multi-factor model with time-varying and seasonal risk premiums for

the natural gas market
Chengwu Shao a, Ramaprasad Bhar b,, David B. Colwell a
a
School of Banking and Finance, Business School, The University of New South Wales, Sydney, NSW 2052, Australia
b
School of Risk and Actuarial, Business School, The University of New South Wales, Sydney, NSW 2052, Australia

a r t i c l e i n f o a b s t r a c t

Article history: In this paper, we develop a quantitative model of the US natural gas market that explores its multi-factor structure
Received 16 June 2014 and its time-varying and seasonal risk premiums. With weekly spot and futures prices we show that three factors
Received in revised form 26 February 2015 are preferred to describe the futures term structure, and the time-varying risk premiums are also signicant.
Accepted 24 April 2015
Moreover, we found that the market implies a seasonal risk premium with two peaks and troughs in one year,
Available online 2 May 2015
which is important to correctly price the futures by maturity month. Finally, we link this seasonal risk premium
JEL classication:
to the uncertainty of the US natural gas demand and nd a positive relationship between them. These results
Q40 reveal the complex aspect of the market, and may have useful applications for other commodity sectors.
Q43 2015 Published by Elsevier B.V.
G13
C32

Keywords:
Natural gas
Short-term and long-term factors
Risk premium
Seasonality

1. Introduction
Global natural gas markets have undergone signicant changes in
regulation in recent decades. The US natural gas market is now the largest
one in the world and its deregulation started with The Natural Gas Policy
Act in 1978. Since then, the market has become more competitive and
relevant trading activities have greatly increased. Therefore it is impor-
tant for both academia and industry to obtain a deep understanding of
price movements and risk factors in the market.
The spot-based commodity models trace back to Brennan and
Schwartz (1985). This early model does not consider the mean-
reversion phenomenon and the Samuelson effect (i.e., the observation
that the volatility of a commodity futures contract tends to increase
when approaching its maturity) which are two properties usually seen
in commodity markets, including natural gas. To incorporate these
properties, Schwartz (1997) proposed a one-factor model where the
spot log-price follows an OrnsteinUhlenbeck process. However, the
futures term structures derived under this model are too strict and
inconsistent with empirical observations. In fact, the futures prices
over the whole term curve are perfectly correlated. Moreover, the
Corresponding author. Tel.: +61 2 9385 4930.
E-mail addresses: c.shao@student.unsw.edu.au (C. Shao), r.bhar@unsw.edu.au
(R. Bhar), d.colwell@unsw.edu.au (D.B. Colwell).
futures price volatility shrinks to zero in the long term, which is not
the case in reality.
To solve these drawbacks, several two-factor models have been con-
structed in the literature. Broadly speaking these models can be divided
into two types. The rst type is based on Gibson and Schwartz (1990)
where one factor is the spot price and the other is the stochastic conve-
nience yield. The second type was introduced in Schwartz and Smith
(2000) where one factor models short-term price deviations and the
other models the long-term equilibrium price evolution (hereafter
ST/LT). Schwartz and Smith also showed that both types are equivalent
in nature. Under these models the futures prices are no longer perfectly
correlated and so richer term structures can be produced. In addition,
because of the long-term factor, the futures price volatility converges
to a non-zero value in the long term, which is closer to actual situations.
Between the two types, the Schwartz and Smith (2000) structure has
advantages of being easy to interpret and having usually weakly corre-
lated factors. Since then, a number of papers have extended the original
ST/LT model to three factors by adding another short-term factor
(hereafter 2ST/LT) and found evidence in favour of such an extension.
Cortazar and Naranjo (2006) developed a general N-factor framework
and found that the 2ST/LT model ts the crude oil futures term struc-
ture much better than the ST/LT model. Bhar and Lee (2011) also
conducted a model comparison using crude oil futures. They found
that the 2ST/LT model shows signicant improvement and the ST/LT

http://dx.doi.org/10.1016/j.eneco.2015.04.013
0140-9883/ 2015 Published by Elsevier B.V.
208 C. Shao et al. / Energy Economics 50 (2015) 207214
model is misspecied. Therefore, in this paper we take the Schwartz and
Smith (2000) structure as the starting point for further development,
and make a comparison between the ST/LT version and the 2ST/LT
version.
As a special kind of commodity, natural gas exhibits signicant
seasonal behaviours which are well documented in much literature
such as Manoliu and Tompaidis (2002) and Mirantes et al. (2012).
Take crude oil as a counter example. One reason for the difference in
their seasonality is their different demand patterns. In the US, natural
gas is continuingly used as a fuel for heating purposes, and recently
there has been a growing trend to use natural gas for marginal power
generation in order to satisfy cooling needs. Therefore, the demand for
natural gas has a seasonal pattern with two peaks in winter and summer
respectively, and the weather conditions, especially the temperature,
play key roles. Mu (2007) analysed such seasonal demand patterns
and found signicant weather impacts. On the contrary, crude oil is
mainly used in the transportation and production industries which are
not strongly affected by seasons. Another reason is due to their different
physical structures. Compared with crude oil, natural gas is more costly
to transport and store, which makes its market more difcult to smooth
out its seasonal demand. The most common way to incorporate season-
ality into a commodity model is to add a deterministic seasonal compo-
nent in the spot price process. In this paper we use a sum of continuous
trigonometric functions similar to Srensen (2002).
For spot-based models, the spot price dynamics are directly specied
under the real-world measure (hereafter ). The martingale approach is
then employed to price futures as expectations of the future spot price
under an equivalent martingale measure (hereafter ) where the spot
price process follows another dynamics. For diffusion cases, such changes
in dynamics are captured by adjusting the drift terms, and risk premiums
are then the differences in the drift terms. In this paper we dene them as
the drift terms under minus the drift terms under . The intuition be-
hind the risk premiums becomes clear within the discussion of contango
and normal backwardation, in which one considers the difference be-
tween E ST  and the futures price, F 0; T E ST . If E ST NF 0; T
(normal backwardation), then the short futures hedger has weak
bargaining power, which pushes the futures price down relative to the
expected spot price. On the other hand, if E ST b F 0; T (contango),
then the long futures hedger has weak bargaining power, which pushes
the futures price up. Clearly, in order to compare the two expectations,
we only need to compare the drift terms under each measure. The classi-
cal setup is to assume constant risk premiums. However, there are also
some papers supporting time-varying risk premiums in commodity mar-
kets. For example, Considine and Larson (2001) and Chiou Wei and Zhu
(2006) found time-varying risk premiums in natural gas markets. To
account for time-varying risk premiums in our models, we apply the
framework in Duffee (2002) and Cheridito et al. (2007) and set the risk
premiums as afne functions of state factors. Casassus and Collin-
Dufresne (2005) adopted this approach under a stochastic convenience
yield model and found time-varying risk premiums in silver, gold and
copper markets. Cartea and Williams (2008) investigated the UK natural
gas market under the ST/LT model with a time-varying risk premium
whose estimation is not very sharp due to the limited amount of data
they used. In this paper, we use a dataset covering wider ranges in both
time-series and cross-section dimensions in order to obtain statistically
signicant results.
In terms of nancial interpretations, the risk premium can be treated
as the insurance paid by asset purchasers to hedge risk by entering into
long positions in futures contracts. On the other side, the negative risk
premium is the insurance paid by asset holders to hedge risk by entering
into short positions in futures contracts. If there is a higher requirement
for asset demand in some period, then the asset purchasers are more
willing to hedge risk than the assets holders. This means the aggregate
risk premium in this period is higher, and the prices of futures maturing
in this period are pushed up by the asset purchasers. As mentioned earlier,
the demand for natural gas in the US has a seasonal pattern with one peak
in winter and another in summer. Thus it is possible that the risk premi-
um in the US natural gas market also has some seasonality that is related
to such a demand pattern. In the literature, seasonality in the risk premi-
um has been considered by several authors. Villaplana (2003) modelled a
seasonal jump risk premium in electricity markets. Cartea and Villaplana
(2008) proposed a supplydemand model for electricity markets and
obtained a seasonal risk premium that is related to demand volatility.
Borovkova and Geman (2006) directly modelled natural gas forward
curves and introduced a seasonal risk premium as a function of the
forwards maturity month only. Jin et al. (2012) studied agricultural
commodities and set a seasonal risk premium as a sum of trigonometric
functions in the mean-reversion level of the convenience yield process
under . In this paper, we take a similar approach and specify two
peaks in one year in the mean-reversion level of a short-term factor pro-
cess under .
This paper has three contributions. First, we extend the model of
Cartea and Williams (2008) to three factors and justify such an extension.
Using a rich dataset we obtain two signicant time-varying risk
premiums. Second, we model a seasonal risk premium in the form of a
seasonal mean-reversion level under and nd statistical justications
for this component. This multi-factor spot-based model with such a sea-
sonal risk premium is new for energy markets. The two peaks in one year
in the seasonal risk premium are also newly detected for natural gas.
Third, we nd a positive relationship between the obtained seasonal
risk premium and the uncertainty of the US natural gas demand, which
is also new in the literature in terms of exploring quantitative relation-
ships between the risk premium and fundamental factors in natural gas
markets.
The remainder of the paper proceeds as follows. In Section 2 we pro-
pose our models and derive futures prices. In Section 3 we present the
data and estimation methodology. In Section 4 we provide and discuss
empirical results. Section 5 concludes the paper.
2. Model and futures price
The 2ST/LT model extends the classic Schwartz and Smith (2000)
ST/LT model by adding another short-term factor. The spot price is
decomposed as the sum of three factors: one long-term factor model-
ling the equilibrium price evolution related to technology advancement,
market structure change, resource discovery, etc; and two short-term
factors capturing price deviations caused by temporary supplydemand
imbalance or other short-term reasons.
The proposed Model 1 is a 2ST/LT version, and has a deterministic
seasonal component, time-varying risk premiums and a seasonal risk
premium. As usual, we dene a probability space (, F, ) where the l-
tration {F t}t 0 is generated by Brownian motions Wt = [W1,t W2,t W3,t].
The spot price process under is then modelled as
log St f t X 1;t X 2;t Y t 1a
dX 1;t 1 X 1;t dt 1 dW 1;t 1b
dX 2;t 2 X 2;t dt 2 dW 2;t 1c
dY t dt 3 dW 3;t : 1d
In the above equations, f(t) is the deterministic seasonal component,
X1,t and X2,t are the two short-term factors, and Yt is the long-term
factor. W1,t, W2,t and W3,t are correlated Brownian motions such that
dWi,tdWj,t = ijdt. 1 and 2 represent the mean-reversion speeds of
short-term price deviations in X1,t and X2,t respectively, while stands
for the average growth rate of the long-term equilibrium price evolu-
tion. 1, 2 and 3 are the volatility parameters.
There are several ways to specify the deterministic seasonal compo-
nent f(t) in the commodities literature. Some papers such as Lucia and
Schwartz (2002) use dummy variables. In this paper, we follow the
 C. Shao et al. / Energy Economics 50 (2015) 207214 209


method of Srensen (2002) and specify it as a sum of trigonometric For the factor X2,t, consider X 2;t e 2 t :
functions, which has advantages of producing a smooth component
    
and having fewer parameters to estimate: d X 2;t e 2 t 2 e 2 t X 2;t dt e 2 t dX 2;t
 
2;t e 2 t dt e 2 t 2 dW 2;t :
X
K
f t ssnconst k sin2kt k cos2kt 2
k1 Then
Z Z Z
tN   
tN

tN

where ssnconst is the mean value of f(t). d X 2;t e 2 t 2;t e 2 t dt e 2 t 2 dW 2;t
To price commodities futures by the martingale approach, we need to tn tn
Z
tn
Z
specify an equivalent martingale measure and then obtain futures X
N ti

tN

2;t e 2 t dt e 2 t 2 dW 2;t
prices as -expectations of future spot price. Assuming that the market in1 t i1 tn
is arbitrage-free, then according to Harrison and Pliska (1981), the exis- XN Z ti Z tN
 
tence of such a is guaranteed. However, its specic form depends on 2;t i e 2 t dt e 2 t 2 dW 2;t :
in1 t i1 tn
the market and also the way of changing the measure from . Here we fol-
low the specication in Cartea and Williams (2008), which is consistent
with Duffee (2002) and Cheridito et al. (2007), and set the spot price Thus
dynamics under as

XN
2;t i   t N t i  t N t i1
X 2;t N X 2;t n e 2 t N t n e 2 e 2
log St f t X 1;t X 2;t Y t 3a in1
2
Z tN

  e 2 t N t 2 dW 2;t :
dX 1;t 1 1 X 1;t dt 1 dW 1;t 3b tn

 
dX 2;t 2;t 2 X 2;t dt 2 dW 2;t 3c Now using standard formulae, the futures logprice is

log F t n ; t N log E St N j F tn 
dY t  dt 3 dW 3;t : 3d   1  
f t N E X 1;tN X 2;t N Y tN j F t n Var X 1;tN X 2;tN Y tN j F t n
2
1 t N t n 2 t N t n
At n ; t N X 1;t n e X 2;t n e Y tn
where 1 = 1 1, 2 = 2 2 and * = + . Now W1,t , W2,t
and
W3,t are correlated Brownian motions under such that dWi,t dWj,t =
6
ijdt. 1 and 1/1 are the mean-reversion speed and mean-reversion
level, respectively, of short-term price deviations in X1,t under . The where
same applies to 2 and 2,t/2 in X2,t. Comparing the two measures, 1
1  

and 2 are the differences in the mean-reversion speeds, while is the dif- At n ; t N f t N 1e 1 t N t n
1
ference in the average growth rate of the long-term equilibrium price XN
2;t i   t N ti  t N t i1
evolution.  e 2 e 2  t N t n

From Section 1, risk premiums are the drift terms under minus the in1 2

drift terms under . Therefore, for X1,t its risk premium is (1 1X1,t) 2  
2  

1 1e2 1 tN t n 2 1e2 2 tN t n
(1X1,t) = 1 + 1X1,t. Similarly, for X2,t its risk premium is 2,t + 2 X2,t, 4 1 4 2
and for Yt its risk premium is . They are all afne functions of the corre- 23 12 1 2 
1e 1 2 t N t n
 
t N t n 
sponding factors. The components 1X1,t and 2X2,t are the time-varying 2 1 2
risk premiums and become constant when 1 = 2 = 0. Note that the 1 3  
 

1e 1 t N t n 23  1e 2 t N t n
2 3
13 
mean-reversion level 2,t/2 is a function of calendar time instead of a 1 2
constant. In fact, the component 2,t is the seasonal risk premium in 7
Model 1. As mentioned in Section 1, we specify two peaks in one year.
Since we will use weekly data later, for simplicity we set it equal to Finally, we specify two sub models for comparison. Model 2 is a
a series of weekly constant values approximating a sum of continuous ST/LT version by dropping out the factor X1,t in Model 1, so logSt =
trigonometric functions as a whole. Specically, f(t) + X2,t + Yt. Model 3 differs from Model 1 by excluding the seasonal
risk premium, so 2,t 2 . The corresponding futures logprices are
X
2 straightforward to derive so we do not list them here.
2;t k sin2kt n k cos2kt n for t t n1 ; t n  4
k1
3. Data and estimation methodology
where t0 = 0, {tn}n 0 are equally spaced by t = 7/365.25 representing
3.1. Data description
a one-week interval.
Assuming constant interest rates, at date t the price of the futures
We use weekly observations of the Henry Hub natural gas spot price
contract maturing at T is
and the New York Mercantile Exchange (hereafter NYMEX) natural gas
F t; T E ST jF t : 5 futures prices obtained from DataStream. The Henry Hub is the largest
natural gas hub in the US with high trading activity and liquidity. It is
the standard delivery point of the NYMEX natural gas futures and also
In Model 1 under , conditional on F t, the vector [X1,T X2,T YT] follows
serves as the benchmark for the US natural gas markets. The NYMEX
a multivariate Gaussian distribution, thus X1,T + X2,T + YT is Gaussian and
natural gas futures market is the largest in the world and started trading
ST is then log-normally distributed. Therefore,
in April 1990. The underlying asset is the Henry Hub natural gas spot.
  Each futures contract is dened by its delivery month and its trading
1
E ST j F t  exp E logST j F t  Var logST jF t  terminates three business days before the rst day of its delivery
 2 
  1   month, see www.cmegroup.com for reference. The maturity time is
exp f T E X 1;T X 2;T Y T j F t Var X 1;T X 2;T Y T j F t :
2 set as the midweek of its delivery month. We choose 16 time series of
210 C. Shao et al. / Energy Economics 50 (2015) 207214

futures prices arranged by their time-to-maturity: F1, F2, F3, F4, F5, F6, Table 2
F9, F12, F15, F18, F21, F24, F27, F30, F33 and F36, where Fn is the PCA results for natural gas spot and futures.

contract that is the nth nearest to its maturity. The sampling period Factor Eigenvalue Cum. var.
covers ten years and is from January 2003 to December 2012. This 1 13.6791 0.8047
allows us to better describe the model's time-series properties which 2 2.0302 0.9241
is important for distinguishing its dynamics under from that under 3 0.6061 0.9597
and so detecting the time-varying risk premiums. 4 0.1828 0.9705
5 0.1562 0.9797
One reason to adopt such a wide range of times-to-maturity is Model
6 0.1118 0.9862
1's 2ST/LT structure. The long-term factor dominates the long-term 7 0.0686 0.9903
properties of futures curves, so it can be well represented by the futures 8 0.0386 0.9925
series with long time-to-maturity. The two short-term factors affect 9 0.0335 0.9945
short-end properties of futures curves, so the futures series with short 10 0.0232 0.9959
11 0.0206 0.9971
time-to-maturity (e.g. F1 to F6 successively) contain rich information 12 0.0156 0.9980
for accurate evaluations of the two factors. Another reason is that the 13 0.0112 0.9987
period of the seasonal risk premium is one year. Our 17 series cover 14 0.0074 0.9991
three whole years along the term structure dimension and many of 15 0.0069 0.9995
16 0.0063 0.9999
them are equally spaced by three months so that they provide enough
17 0.0020 1.0000
information for correct extraction of the seasonal risk premium.
The descriptive statistics are provided in Table 1. We observe that for This table reports eigenvalues and proportions of total variance cumulatively explained
(cum. var.) from PCA results. The data consists of standardized weekly returns of the
short-term contracts, the mean of the futures prices increases with Henry Hub natural gas spot and the NYMEX natural gas futures from January 2003 to
time-to-maturity, but for long-term contracts (starting from F18), the December 2012. Price changes at the time of contract switching are excluded.
mean decreases with time-to-maturity. The standard deviation of futures
returns shows a clear decreasing trend as time-to-maturity increases,
which is consistent with the Samuelson effect. factors are required. Thus the PCA results indicate that three factors
are appropriate for the data.
Finally, we group the futures contracts by their maturity months and
3.2. Preliminary analysis present the mean of the futures prices within each group in Table 3. We
see that the mean price is different across maturity months, suggesting
In general, more factors lead to richer term structures and superior the possible role played by the seasonal risk premium in futures pricing.
tting results. However, they come at the cost of increasing estimation
difculties and the concern of overtting. So there is a trade-off for the 3.3. Estimation methodology
number of factors included in the model. The Principal Component
Analysis (hereafter PCA) is a useful procedure to analyse the correlation We rst estimate the deterministic seasonal component f(t) in the
structure among a number of variables and suggest the appropriate spot price process. Cartea and Williams (2008) obtained this compo-
number of factors needed to explain the total variation as much as pos- nent using forward prices. In this paper, one of our purposes is to extract
sible. We conduct the PCA for the standardized returns of the spot and the risk premiums, especially the seasonal risk premium, from the
futures and display its results in Table 2. Hair et al. (1995) discussed futures prices. So to prevent possible inuences of the risk premiums
three criteria to determine the appropriate number of factors. In the implicit in the futures prices when estimating this component, we
nance literature, their Criterion 3: add more factors until the cumula- choose to estimate it using the spot price, which is the approach
tive percentage of the variance explained reaches a prespecied level is adopted in many other papers such as Cartea and Figueroa (2005).
usually adopted and the typical prespecied level is 95%. From Table 2, Fig. 1 shows the average spot logprice of each month. There are two
two factors are not enough to explain 95% of total variance and three peaks in one year: one is around January; and the other is around
June. So we set K = 2 in f(t) and estimate it by minimizing squared
errors. The estimated f(t) is expressed below and its curve is also plotted
Table 1
Descriptive statistics for natural gas spot and futures. in Fig. 1.

Price Return f t 1:6679 0:0258sin2t 0:0165cos2t

8
Series Mean S.D. Mean S.D. 0:0012sin4t 0:0470cos4t :
Spot 5.7441 2.3666 0.0008 0.0963
F1 5.8997 2.3615 0.0073 0.0709 To estimate the deseasonalized part, one approach in the literature is
F2 6.0718 2.3977 0.0078 0.0641 to estimate parameters relevant to the dynamics under using the
F3 6.2143 2.4335 0.0064 0.0573
historical spot price, and then estimate remaining parameters relevant
F4 6.3160 2.4084 0.0049 0.0521
F5 6.3939 2.3892 0.0041 0.0481 to the dynamics under using futures prices. However, this approach
F6 6.4582 2.3609 0.0040 0.0455 does not work well here. The reason is that there are three series of un-
F9 6.5704 2.2743 0.0022 0.0380 observable factors X1,t X2,t Yt to evaluate but only one series of observed
F12 6.5992 2.1270 0.0018 0.0330
F15 6.6396 2.1062 0.0022 0.0330
F18 6.6604 2.0887 0.0017 0.0288 Table 3
F21 6.6213 1.9804 0.0007 0.0266 Means of futures prices by maturity month.
F24 6.5643 1.8527 0.0010 0.0250
F27 6.5543 1.8394 0.0012 0.0258 Maturity month Jan Feb Mar Apr May Jun
F30 6.5466 1.8579 0.0010 0.0249
Mean 7.0918 6.9871 6.7260 6.1503 6.0627 6.0715
F33 6.4975 1.7735 0.0005 0.0236
F36 6.4552 1.6699 0.0007 0.0235
Maturity month Jul Aug Sep Oct Nov Dec
This table reports means and standard deviations (S.D.) of prices and returns of the Henry
Mean 6.1841 6.1949 6.1677 6.2512 6.5480 6.8607
Hub natural gas spot and the NYMEX natural gas futures. The data covers January 2003
December 2012 with 522 weekly observations for each series. For futures returns, changes This table reports means of the NYMEX natural gas futures prices grouped by maturity
due to contract switching are excluded. month. The data covers January 2003December 2012 with weekly observations.
 C. Shao et al. / Energy Economics 50 (2015) 207214 211

2
average
At each time point there are 17 cross-section observations: spot, F1,
1.9 f(t) F2, F3, F4, F5, F6, F9, F12, F15, F18, F21, F24, F27, F30, F33 and F36. So the
corresponding yt = [log St log F(t, t + 1) log F(t, t + 16)], and the
1.8
measurement equation is
1.7 2 3 2 3
1 1 1 f t
1.6 6 e 1 1 e 2 1 1 7 6 At; t 1 7
Ht 6
4
7 dt 6 7:
5 4 5
Logprice

1.5 1 16 2 16
e e 1 At; t 16
1.4
Given the considerable number of parameters we already have, to re-
1.3
duce further computational burden we follow Babbs and Ben Nowman
1.2 (1999), Cortazar and Naranjo (2006) and Mirantes et al. (2012) and
assume independent measurement errors with the same standard devia-
1.1
tion . So Rt = diag( 2, 2, , 2).
1 All programs are coded in MATLAB and the parameters are estimated
1 2 3 4 5 6 7 8 9 10 11 12
by minimizing the negative log-likelihood function using the fminunc
Month
medium scale algorithm. Because of the high-dimensional parameter
Fig. 1. Average spot logprice of each month and the estimated f(t).
space, we implement the fminunc with various initial values to achieve
the minimum point.

log St f(t) available. As pointed out in Schwartz and Smith (2000), it is
difcult to tell which factor or combination of factors causes a given
change in log St f(t) at some time point, and the evaluated factors
are perfectly negatively correlated. In other words, the information pro-
vided by one series of observations is not rich enough to support the in-
formation space spanned by three series of factors. Therefore, to
achieve clearer evaluations of these unobservable factors and conse-
quently better estimations of model parameters, we take another ap-
proach and estimate parameters relevant to and all together using
series of both spot and futures with various time-to-maturity.
The Kalman lter is the optimal procedure to evaluate unobservable
factors under the linear Gaussian state space framework. The likelihood
function can be obtained during its implementation and then model pa-
rameters can be estimated via maximum likelihood estimation (hereafter
MLE). The details are in Harvey (1991) for reference. The Kalman
lter has been widely used in numerous commodities papers such as
Schwartz and Smith (2000), Cortazar and Naranjo (2006), Cartea and
Williams (2008) and Mirantes et al. (2012), among others. Here the
discretized versions of our models t the linear Gaussian state space
framework, so we also use the Kalman lter and outline its setup for
Model 1 below for convenience.
The Kalman lter works on the state space form consisting of two es-
sential components: the transition equation and the measurement
equation. The general form is
4. Empirical results
The parameter estimates and standard errors are listed in Table 4.
Looking at Model 1, the two short-term volatilities 1 and 2 are
much higher than the long-term volatility 3, indicating that the two
short-term factors explain major variations in the spot price. The long-
term drift (under ) is positive but not signicant, suggesting no
clear linear trend in the spot price during the whole sampling period.
The long-term risk premium is negative and signicant at the
10% level, consistent with the results in Mirantes et al. (2012). This
nding can be explained by the hedging pressure in the US natural
gas market. A natural gas production project usually requires a heavy
pre-investment in exploration and facilities. Such pre-investment is
mainly a sunk cost and its pay back relies on a long series of future
cash ows from selling produced natural gas. Therefore, production
companies are sensitive to adverse (downward) movements in natural
gas prices in future, and so have strong motivations to hedge risk by en-
tering into short positions in the long-term futures contracts. Moreover,
from Hobk Haff et al. (2008) natural gas in the US is produced by
Table 4
Model estimation results.
Parameter Model 1 Model 2 Model 3
Estimate S.E. Estimate S.E. Estimate S.E.

1 1.2301 (0.0210) 1.2537 (0.0878)

x t ct At xtt ut ut N 0; Q t 9a
1 0.0218 (0.1145) 0.0848 (0.1069)
1 0.3365 (0.0174) 0.3390 (0.1086)
yt dt Ht xt vt vt N 0; Rt : 9b 1 0.4913 (0.0206) 0.4579 (0.0226)
2 2.6922 (0.1660) 1.4669 (0.0713) 1.9594 (0.1670)
0.0280 (0.1963) 0.0906 (0.1338) 0.1014 (0.2065)
The rst equation is the transition equation governing the
1 0.3182 (0.0041) 0.2902 (0.0042)
dynamics of state factors x; and the second equation is the measure- 1 0.1076 (0.0083) 0.0016 (0.0043)
ment equation relating the state factors to observations y. For Model 1, 2 0.3827 (0.0075) 0.3158 (0.0080)
the xt = [X1,t X2,t Yt], and the transition equation is determined 2 0.2175 (0.0073) 0.2726 (0.0077)
through the solutions to the three stochastic differential equations 2 0.5390 (0.0509) 0.6448 (0.0080) 0.8984 (0.1725)
2 0.5087 (0.0236) 0.7463 (0.0332) 0.4375 (0.0287)
under in Section 2. Specically, 0.0362 (0.0425) 0.0067 (0.0431) 0.0285 (0.0365)
2 3 2 3 0.0716 (0.0425) 0.0571 (0.0434) 0.0584 (0.0367)
e 1 t 0 0 0 3 0.1333 (0.0065) 0.1360 (0.0072) 0.1149 (0.0071)
At 4 0 e 2 t
0 5 ct 4 05 12 0.4487 (0.0119) 0.4435 (0.0291)
0 0 1 t
2 3 13 0.1265 (0.0255) 0.2099 (0.0630)
  21    13 1 3
6 1e2 1 t 1e 1 2 t 12
1 2
1e 1 t 7 23 0.1127 (0.0225) 0.3326 (0.0284) 0.1591 (0.0229)
6 2 1 1 2 1 7 0.0318 (0.0003) 0.0423 (0.0003) 0.0512 (0.0004)
6   22   23 2 3 7
Qt 6
6 1e 1 2 t
12 1 2
1e2 2 t 1e 2 t 7:
7 Log-likelihood 24,704.29 22,608.75 20,994.83
6 1 2 2 2 2 7
4   1 3   2 3 5
1e 1 t 13 1e 2 t 23 2
3 t This table reports parameter estimates and standard errors (S.E.) for Model 1, Model 2 and
1 2 Model 3. The procedure is implemented by the Kalman lter and MLE. The log-likelihood
values do not include the constant parts of log-densities of Gaussian distributions.
212 C. Shao et al. / Energy Economics 50 (2015) 207214

many small companies. This suggests a weak bargaining power on the
supply side and low risk tolerance due to small company sizes, which
further strengthens their motivations to hedge risk. On the other side,
the US natural gas buyers do not need such heavy pre-investments.
They are mainly concerned about short-term needs and can usually
pass high natural gas prices to their customers. Therefore, they do not
have strong motivations to hedge risk by entering into long positions in
the long-term futures contracts. In sum, the net hedging pressure on
the long-term futures contracts is from the shorting side. Consequently,
to balance this pressure the long-term futures prices are pushed down,
which is captured by the negative long-term risk premium that can
be treated as the insurance paid by the natural gas producers.
Most previous gas literature only assumed constant (or piecewise
constant) risk premiums in their stochastic factor models, such as
Kolos and Ronn (2008), Poblacin (2011) and Mirantes et al. (2012).
Here, we incorporate time-varying risk premiums into Model 1, and ob-
tain strongly signicant estimates: both 1 and 2 are signicant at the
1% level. It is often seen in the literature that estimates of time-varying
risk premium have low signicance levels. The main reason is that their
parameters do not explicitly enter the futures pricing formulae. Hence,
their estimates cannot be directly extracted from cross-section futures
prices, and have to be inferred from time-series dynamics. Compared
with Cartea and Williams (2008), the data here covers ten years, and
thus the signicance levels are greatly improved. The strongly signi-
cant time-varying risk premiums indicate that the mean-reversion
speeds under are different from those under , and thus provide a
more accurate description of the price dynamics under . This is impor-
tant when computing the holding period return distribution and value-
at-risk, as demonstrated by Casassus and Collin-Dufresne (2005).
Comparing Model 2 to Model 1, we notice that there are major dif-
ferences in parameter estimates for the short-term factor(s) but minor
differences for the long-term factor. This is understandable because
the long-term information is still captured by the same long-term factor,
while the short-term information is now merged into one short-term
factor in Model 2. To justify the improvement of Model 1 over Model
2, we rst observe that the estimates of mean-reversion speeds and
volatilities of both short-term factors in Model 1 are all signicant. The
standard deviation of Model 1's measurement errors is also smaller
than that of Model 2. Conducting the likelihood ratio (hereafter LR)
test, the degree of freedom is 6 and the LR statistic is 4191.08 that is
much higher than the critical value 20.01,6 = 16.81. Thus we reject the
null hypothesis of Model 2 in favour of Model 1. Finally we compare
both models' tting quality in Table 5. In general Model 1 has smaller
Table 5
Model errors by time-to-maturity.
Model 1 Model 2
mean errors and RMSE than Model 2, and the reduction of RMSE
becomes noticeable for short-term series. This means Model 1 yields
better tting quality with one more short-term factor, and the improve-
ment is more effective for short-term contracts. In sum, we conclude
that the three-factor Model 1 is preferred to the two-factor Model 2,
consistent with the PCA results in Section 3.2.
Although the seasonality in natural gas is well-known, most previous
gas literature only considered it in the spot price; see for example Manoliu
and Tompaidis (2002), Cartea and Williams (2008) and Mirantes et al.
(2012). Here, besides in the spot price, we introduce seasonality into
the risk premium of a short-term factor. In addition, the estimated
seasonal risk premium 2,t has two peaks and troughs in one year: four
relevant parameters 1, 1, 2 and 2 are all strongly signicant. Its curve
plotted in Fig. 2 shows that 2,t attains its local minima in late April and
early November, and attains its local maxima in January and early August.
To investigate the importance of this seasonal risk premium we compare
Model 3 to Model 1. We rst observe that the standard deviation of
Model 1's measurement errors is smaller than that of Model 3. Applying
the LR test, the degrees of freedom is 4 and the LR statistic is 7418.92
that is much higher than the critical value 20.01,4 = 13.28. Thus we reject
the null hypothesis of Model 3 in favour of Model 1. For further analysis
we compare both models' tting quality by maturity month in Table 6.
We see that Model 1 has smaller mean errors and RMSE than Model 3
for all maturity months. In particular, the pattern of the mispricing in
Model 3, represented by its mean errors across maturity months, can be
corrected by the seasonal risk premium 2,t introduced in Model 1. To
demonstrate this we calculate the conditional expectation of X2,T under
in Model 1 and plot its curve in Fig. 3. The maturity time T covers
36 months and X2,0 is assumed to be 0. From the deduction in Section 2,
this expectation expresses the effect of 2,t on futures pricing. Fig. 3
shows that the pattern of this expectation well matches the pattern of
the mispricing in Model 3. In sum, we conclude that the seasonal risk
premium 2,t plays a signicant role when pricing natural gas futures.
Furthermore, it renders a better picture of the risk premium structure in
the US natural gas market, and thus strengthens our understanding of
the relationship between futures prices and the spot price in the market.
This advancement is helpful for optimal hedging decisions and correct
valuations of real options, as pointed out by Kolos and Ronn (2008).
Finally, we link the seasonal risk premium 2,t to some fundamental
factor in the US natural gas market. Bessembinder and Lemmon (2002)
and Kolos and Ronn (2008) suggested a positive relationship between
the risk premium and the demand variance. Cartea and Villaplana
(2008) proposed a supplydemand model under which the electricity
spot price is determined by a demand factor and a capacity factor. The
demand factor follows two diffusion processes with different drift
0.6

Series Mean error RMSE Mean error RMSE

0.4
Spot 0.0038 0.0388 0.0108 0.0810
F1 0.0059 0.0284 0.0085 0.0434
F2 0.0024 0.0310 0.0006 0.0314 0.2
F3 0.0003 0.0281 0.0057 0.0331
F4 0.0007 0.0255 0.0077 0.0381
F5 0.0006 0.0267 0.0075 0.0421 0
F6 0.0010 0.0289 0.0071 0.0451
F9 0.0020 0.0369 0.0042 0.0428
0.2
F12 0.0003 0.0328 0.0017 0.0371
F15 0.0008 0.0280 0.0030 0.0333
F18 0.0017 0.0279 0.0032 0.0321 0.4
F21 0.0004 0.0302 0.0045 0.0312
F24 0.0030 0.0255 0.0070 0.0279
F27 0.0017 0.0269 0.0040 0.0288 0.6
F30 0.0001 0.0292 0.0002 0.0292
F33 0.0004 0.0259 0.0028 0.0361
0.8
F36 0.0018 0.0310 0.0069 0.0439 1 2 3 4 5 6 7 8 9 10 11 12
This table reports mean errors and root-mean-square-error (RMSE) of Model 1 and Model Month
2 for each series. The error is the difference between the tted logprice and the observed
logprice. Fig. 2. Estimated seasonal risk premium 2,t in Model 1.
 C. Shao et al. / Energy Economics 50 (2015) 207214 213

5
Table 6 x 10
2.5
Model errors by maturity month.

Model 1 Model 3

Maturity month Mean error RMSE Mean error RMSE 2

1 0.0013 0.0345 0.0233 0.0443
2 0.0022 0.0360 0.0427 0.0579
3 0.0113 0.0370 0.0492 0.0624 1.5
4 0.0092 0.0296 0.0128 0.0335
5 0.0026 0.0332 0.0491 0.0610
6 0.0013 0.0286 0.0736 0.0810
7 0.0001 0.0189 0.0500 0.0554 1
8 0.0017 0.0211 0.0035 0.0230
9 0.0026 0.0236 0.0253 0.0357
10 0.0010 0.0297 0.0251 0.0419
11 0.0047 0.0211 0.0159 0.0291 0.5
12 0.0023 0.0282 0.0047 0.0290

This table reports mean errors and root-mean-square-error (RMSE) of Model 1 and Model
3 grouped by maturity month. The error is the difference between the tted futures 0
logprice and the observed futures logprice. 1 2 3 4 5 6 7 8 9 10 11 12
Month

terms under and respectively. The difference in the drift terms be-
tween and and the seasonality in it affect the forward price and for-
ward premium. Therefore, this difference can be treated as the seasonal
risk premium for the demand factor. Moreover, this difference is speci-
ed as the product of the demand volatility and the market price of de-
mand risk, and a high demand volatility is found to be usually
accompanied with a large positive market price of demand risk.
Hence, the model implies a positive relationship between this difference
and the demand volatility, and thus a positive relationship between the
seasonal risk premium for the demand factor and the demand variance.
Although a positive relationship between the risk premium and the
demand variance has been well motivated by Bessembinder and
Lemmon (2002), Kolos and Ronn (2008) and Cartea and Villaplana
(2008), no previous gas literature studied it. Here, we explore the rela-
tionship between the seasonal risk premium 2,t and the uncertainty of
the US natural gas demand, which is an innovative step in the literature.
For each month, we measure the uncertainty by calculating the stan-
dard deviation of the US natural gas consumptions in that month. The
data is from the US Energy Information Administration website and
covers January 2003December 2012. The results are presented in
Fig. 4. We nd that there are two peaks and troughs across months,
and the pattern is similar to the pattern of the seasonal risk premium
2,t plotted in Fig. 2. The correlation between the value of 2,t at
mid-month and the uncertainty is 0.6545. Implementing a simple
linear regression of the value of 2,t at mid-month on the uncertainty,
Fig. 4. Uncertainty of natural gas demand (standard deviation of natural gas consumptions).
the coefcient is 7.0764 106 and its 95% condence interval is
[1.3170 106 12.8358 106]. In sum, these ndings indicate a sig-
nicant positive relationship between the seasonal risk premium 2,t
and the uncertainty of the US natural gas demand.
5. Conclusions
This paper proposes a three-factor model for the US natural gas
market. The model is equipped with two short-term factors, a long-
term factor, a deterministic seasonal component, two time-varying
risk premiums and a seasonal risk premium. The model is estimated
by the Kalman lter and maximum likelihood estimation using the
Henry Hub natural gas spot price and the NYMEX natural gas futures
prices. The long-term risk premium is signicantly negative, explained
by the net hedging pressure from the US natural gas supply side. The
time-varying risk premiums are also strongly signicant. The model is
statistically preferred to the two-factor version, justifying one more
short-term factor. The seasonal risk premium has two peaks and
troughs in one year, and is important for futures pricing. Finally, a pos-
itive relationship is found between the seasonal risk premium and the
uncertainty of the US natural gas demand.
Several kinds of extensions are possible for this model. One possibility
is to add a jump factor in the spot price and examine the effect of the jump

0.08

0.06

0.04

0.02

0.02

0.04

0.06

0.08

0.1
1 2 3 4 5 6 7 8 9 10 11 12 1 2 3 4 5 6 7 8 9 10 11 12 1 2 3 4 5 6 7 8 9 10 11 12
Month

Fig. 3. Expectation of X2,T given X2,0 = 0 under in Model 1.

214 C. Shao et al. / Energy Economics 50 (2015) 207214
risk premium on futures pricing. Another way is to specify 2 as a function
of relevant fundamental factors directly, such as some statistical measures
of temperature, heating degree days and cooling degree days, etc. We
leave these topics for future research.
References
Babbs, S., Ben Nowman, K., 1999. Kalman ltering of generalized vasicek term structure
models. J. Financ. Quant. Anal. 34 (1).
Bessembinder, H., Lemmon, M., 2002. Equilibrium pricing and optimal hedging in elec-
tricity forward markets. J. Financ. 57 (3), 13471382.
Bhar, R., Lee, D., 2011. Time-varying market price of risk in the crude oil futures market.
J. Futur. Mark. 31 (8), 779807.
Borovkova, S., Geman, H., 2006. Seasonal and stochastic effects in commodity forward
curves. Rev. Deriv. Res. 9 (2), 167186.
Brennan, M., Schwartz, E., 1985. Evaluating natural resource investments. J. Bus. 135157.
Cartea, A., Figueroa, M., 2005. Pricing in electricity markets: a mean reverting jump diffusion
model with seasonality. Appl. Math. Financ. 12 (4), 313335.
Cartea, A., Villaplana, P., 2008. Spot price modeling and the valuation of electricity for-
ward contracts: the role of demand and capacity. J. Bank. Financ. 32 (12), 25022519.
Cartea, A., Williams, T., 2008. UK gas markets: the market price of risk and applications to
multiple interruptible supply contracts. Energy Econ. 30 (3), 829846.
Casassus, J., Collin-Dufresne, P., 2005. Stochastic convenience yield implied from
commodity futures and interest rates. J. Financ. 60 (5), 22832331.
Cheridito, P., Filipovi, D., Kimmel, R., 2007. Market price of risk specications for afne
models: theory and evidence. J. Financ. Econ. 83 (1), 123170.
Chiou Wei, S., Zhu, Z., 2006. Commodity convenience yield and risk premium determination:
the case of the US natural gas market. Energy Econ. 28 (4), 523534.
Considine, T., Larson, D., 2001. Risk premiums on inventory assets: the case of crude oil
and natural gas. J. Futur. Mark. 21 (2), 109126.
Cortazar, G., Naranjo, L., 2006. An n-factor gaussian model of oil futures prices. J. Futur.
Mark. 26 (3), 243268.
Duffee, G., 2002. Term premia and interest rate forecasts in afne models. J. Financ. 57 (1),
405443.
Gibson, R., Schwartz, E., 1990. Stochastic convenience yield and the pricing of oil contingent
claims. J. Financ. 45 (3), 959976.
Hair Jr., J., Anderson, R., Black, W., Tatham, R., 1995. Multivariate Data Analysis: With
Readings. Prentice Hall, New Jersy.
Harrison, J., Pliska, S., 1981. Martingales and stochastic integrals in the theory of continuous
trading. Stoch. Process. Appl. 11 (3), 215260.
Harvey, A., 1991. Forecasting, Structural Time Series Models and the Kalman Filter.
Cambridge University Press.
Hobk Haff, I., Lindqvist, O., Lland, A., 2008. Risk premium in the UK natural gas forward
market. Energy Econ. 30 (5), 24202440.
Jin, N., Lence, S., Hart, C., Hayes, D., 2012. The long-term structure of commodity futures.
Am. J. Agric. Econ. 94 (3), 718735.
Kolos, S., Ronn, E., 2008. Estimating the commodity market price of risk for energy prices.
Energy Econ. 30 (2), 621641.
Lucia, J., Schwartz, E., 2002. Electricity prices and power derivatives: evidence from the
Nordic power exchange. Rev. Deriv. Res. 5 (1), 550.
Manoliu, M., Tompaidis, S., 2002. Energy futures prices: term structure models with
kalman lter estimation. Appl. Math. Financ. 9 (1), 2143.
Mirantes, A., Poblacin, J., Serna, G., 2012. The stochastic seasonal behaviour of natural gas
prices. Eur. Financ. Manag. 18 (3), 410443.
Mu, X., 2007. Weather, storage, and natural gas price dynamics: fundamentals and volatility.
Energy Econ. 29 (1), 4663.
Poblacin, J., 2011. Commodity models and investment under uncertainty: the optimal
contract determination. J. Deriv. Hedge Funds 16 (4), 240252.
Schwartz, E., 1997. The stochastic behavior of commodity prices: implications for valuation
and hedging. J. Financ. 52 (3), 923973.
Schwartz, E., Smith, J., 2000. Short-term variations and long-term dynamics in commodity
prices. Manag. Sci. 46 (7), 893911.
Srensen, C., 2002. Modeling seasonality in agricultural commodity futures. J. Futur. Mark.
22 (5), 393426.
Villaplana, P., 2003. Pricing power derivatives: a two-factor jump-diffusion approach.
Business Economics Working Paper.