Energy Economics 31 (2009) 736747

Contents lists available at ScienceDirect

Energy Economics
j o u r n a l h o m e p a g e : w w w. e l s ev i e r. c o m / l o c a t e / e n e c o

A supply and demand based volatility model for energy prices

Takashi Kanamura
J-POWER, 15-1, Ginza 6-Chome, Chuo-ku, Tokyo 104-8165, Japan

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

Article history: This paper proposes a new volatility model for energy prices using the supplydemand relationship, which
Received 5 December 2008 we call a supply and demand based volatility model. We show that the supply curve shape in the model
Received in revised form 29 March 2009 determines the characteristics of the volatility in energy prices. It is found that the inverse BoxCox
Accepted 1 April 2009
transformation supply curve reecting energy markets causes the inverse leverage effect, i.e., positive
Available online 7 April 2009
correlation between energy prices and volatility. The model is also used to show that an existing (G)ARCH-M
Keywords:
model has the foundations on the supplydemand relationship. Additionally, we conduct the empirical
Energy prices studies analyzing the volatility in the U.S. natural gas prices.
Volatility 2009 Elsevier B.V. All rights reserved.
Supply curve
Inverse leverage effect
Volatility-in-mean effect

1. Introduction
High volatility in price returns often appears in deregulated energy
markets. The market participants such as energy producers and
distributors always face such high volatility risks from energy markets.
As a simple example to illustrate the inuence of the volatility, let us
think of a thermal power plant procuring natural gas as the fuel from
the spot market. Since the natural gas prices are volatile due to the
supply and demand in the market, the risk manager in charge of the
procurement of natural gas needs to capture the volatility as accurately
as possible by using an energy volatility model. The volatility models
have been introduced into energy markets because of the needs from
energy market participants as described in the above example.
Although a lot of volatility models both in continuous and discrete
time were developed in nancial markets, they are directly applied to
the volatility models in energy markets without any adjustment for the
energy characteristics. The continuous time models in stock markets,
such as Heston model introduced by Heston (1993) and C.E.V.
(Constant Elasticity of Variance) models developed by Cox (1975)
and extended by Emanuel and MacBeth (1982), are directly used for
the models in energy markets (e.g., Eydeland and Wolyniec (2003)).
Views expressed in this paper are those of the author and do not necessarily reect
those of J-POWER. I am grateful to Toshiki Honda, Matteo Manera, Ryozo Miura, Izumi
Nagayama, Nobuhiro Nakamura, and especially Kazuhiko hashi for their useful
comments and suggestions. I also wish to thank Juri Hinz, Svetlozar Rachev, Christoph
Reisinger, Ryosuke Wada, and all seminar participants at the Bachelier Finance Society
2006 Fourth World Congress, the Commodity Finance Conference 2007, and University
of Karlsruhe for their helpful comments and suggestions. All remaining errors are mine.
Tel.: +81 3 3546 9375; fax: +81 3 3546 9531.
E-mail address: tkanamura@gmail.com.
Similarly, the discrete time models such as ARCH, GARCH, and ARCH-M
models in Engle (1982), Bollerslev (1986), and Engle et al. (1987) are
employed in energy market models in Dufe et al. (1999), Pindyck
(2004), and Deaves and Krinsky (1992), respectively. Energy markets
may utilize the same volatility models as nancial markets. However,
the volatility in energy markets is not necessarily the same as the
volatility in stock markets. For instance, the inverse leverage effect,
i.e., volatility increases in prices, often appears in energy markets,
while the analyses in stock markets illustrate the opposite relationship
between the volatility and the prices.
Energy price models have been developed as asset pricing models in
commodity markets. Gibson and Schwartz (1990) propose a two-factor
model for crude oil in which the log of spot price follows a normal
process and the convenience yield follows a mean reverting process.
Brennan (1991) also models commodity spot prices and convenience
yields as separate stochastic processes with a constant correlation.
Schwartz (1997) compares one-, two-, and three-factor models in which
the log of spot price follows a simple stochastic process to describe
commodity prices. One relevant paper in this class is Kolos and Ronn
(2008) where energy forward price returns precisely demonstrate
volatility-in-mean effect. While these models work well to describe
energy prices, supply and demand relationship of energy is not explicitly
structured in the models. This paper proposes the SDV model for energy
prices that can accurately demonstrate the characteristics of volatility in
energy prices using the supplydemand relationship.
We show time series of demand and price for natural gas in the U.S. in
Fig. 1. The gure seems to suggest that prices increase in demand,
especially in recent years because peaks of prices correspond to peaks of
demand. In order to express the curve using a simple model, we introduce
an equilibrium price model determined by the inelastic demand curve
uctuating stochastically and the upward-sloping supply curve xed in a

0140-9883/$ see front matter 2009 Elsevier B.V. All rights reserved.
doi:10.1016/j.eneco.2009.04.001
 T. Kanamura / Energy Economics 31 (2009) 736747
Fig. 1. The relationship between demand (= supply) and prices.
short period of time, what we call a supply and demand based volatility
(SDV) model. We show that the SDV model provides twofold character-
istics of the volatility in energy price returns. First, the volatility is time
varying owing to both the upward-sloping supply curve and demand
shocks. Second, the mean of price returns changes with the volatility,
which we call volatility-in-mean effect, due to the drift term of price
returns represented by the function of the volatility. Then, we specify the
model by employing the inverse BoxCox transformation supply curve
that can exhibit the drastic slope changes with an appropriate function
parameter. It is found that the model can produce both the inverse
leverage effect and volatility-in-mean effect. In addition, getting an idea
from the volatility-in-mean effect, we investigate the relationship
between the discrete time SDV model and (G)ARCH-M model. We
show that an existing (G)ARCH-M model has the foundations on the
supplydemand relationship.
We conduct the empirical analyses on the volatility in the U.S.
natural gas prices by using the SDV model and the implications from
the model. First, in an effort to examine the existence of the inverse
leverage effect in the U.S. natural gas market, we identify the model
parameters of the inverse BoxCox supply curve employing nonlinear
least squares. Both monthly equilibrium prices and demands are used
for the identication, supposing that the sum of consumption and
storage demands approximates to the equilibrium demand. The
results illustrate that the extremely large change of the gradient
causes the inverse leverage effect in the natural gas market. Second,
we examine the existence of the volatility-in-mean effect in the
natural gas market by using GARCH(1,1)-M model linked to the
discrete time SDV model. It is found that the estimation results
support the existence of volatility-in-mean effect.
Finally, we empirically investigate the validity of two assump-
tions in this paper. The rst assumption is that the sum of
consumption and storage demands approximates to the equilibrium
demand. The second assumption is that demand is inelastic to prices
in the SDV model. We show that both assumptions hold as the rst
order approximation.
737
The remainder of this paper is organized as follows. Section 2
proposes a new model for time-varying volatility in energy prices,
what we call SDV model and then investigates the relationship
between the SDV model and existing (G)ARCH-M models. Section 3
conducts empirical studies on the volatility in the U.S. natural gas
prices in attempts to examine the existence of the inverse leverage
and volatility-in-mean effects as the characteristics of the price
volatility. Section 4 addresses the validity of the assumption on the
equilibrium demand. Section 5 identies the simultaneous equa-
tions for the supply and demand curves of the natural gas market by
implementing the nonlinear two-stage least square estimation
in order to assess the validity of the demand inelasticity assump-
tion. Section 6 concludes and offers the directions for our future
research.
2. The model
2.1. A supply and demand based volatility (SDV) model for energy prices
Fig. 1 illustrates the relationship between monthly equilibrium
demands and prices for natural gas in the U.S.1 As we see, the prices
seem to increase in demand. The market observation has motivated us
to incorporate the supplydemand relationship directly into the
energy spot price model, i.e., the volatility model. In order to build the
model up, we employ the simplied relationship between the demand
and price for energy as follows.
As in Fig. 2, we assume that the equilibrium prices are determined
by the xed increasing supply curve and the stochastically uctuating
vertical demand curve. The assumption of the xed supply curve
comes from the observation on energy markets, i.e., both the number
and constitution of the production facilities are almost kept constant
in a short term. The assumption of the vertical demand curve arises
1
We use well head prices and total demand for natural gas.
738 T. Kanamura / Energy Economics 31 (2009) 736747
Fig. 2. Equilibrium prices.
from the demand inelasticity to prices in the short run, i.e., energy use
seems to be independent of the price change in that period if it is more
or less irrelevant to the seasonality of demand. In this sense, the model
is considered as the rst order approximation. Note that the validity of
the assumption is investigated in Section 5. The idea has been already
applied to the spot price model for electricity as in Barlow (2002) and
Kanamura and hashi (2007). It should be noted that as in Fig. 2 the
model is able to generate the large rate of price change with respect to
the demand as the demand (i.e., price) increases, supposing that the
supply curve is dramatically upward sloping.
Based on the above idea, we formulate the model as follows. We
denote the xed supply curve for energy by Pt = g(St) which is the
second order differentiable, monotone, and increasing convex func-
tion with respect to the supply St. As the supportive discussion,
Pirrong and Jermakyan (2008) offer a convex supply stack for the PJM
market. On the assumption of the demand inelasticity, the equilibrium
prices are determined by the supply curve function replacing the
supply St by the demand Dt. Here, we model the demand for energy,
assuming that the demand follows a stochastic process:
dDt = D dt + D dwt : 1 and whose solution exists by employing an explicit supply function
Proposition 1. Suppose that equilibrium prices for energy are given by
the function of the demand as Pt = g(Dt) where the demand process is
generated by Eq. (1). Then, a Supply and Demand based Volatility (SDV)
model for energy prices is expressed as follows:
dPt
= t dt + t dwt ; 2
Pt
g VDt
t = ; 3
g Dt D


D 1 gWDt
t = + D t: 4 mean reverting process: prices are obtained from the exponential
D 2 g VDt
Proof. We apply Ito's Lemma to Pt = g(Dt) by using Eq. (1).||
Eq. (3) shows that the volatility t is proportional to the ratio of the
rst derivative to the level of the supply curve function g, which
causes time-varying volatility by the demand shocks as long as ggVD t
Dt
does not become constant. Moreover, the drift term in Eq. (4) changes
with the volatility, which leads to the existence of the volatility-in-
mean effect. In addition, since the supply curve g is the second order
D
differentiable, monotonic, and increasing convex function, gW g D z 0
holds in Eq. (4). It implies that the volatility positively affects the drift
term in Eq. (4), when D is greater than or equal to 0. As we have
mentioned, the simplest model for energy prices based on the supply
demand relationship even can produce two signicant characteristics
on the volatility. First, the model exhibits time-varying property, since
the volatility of the price returns depends on the demand. Second,
there exists the volatility-in-mean effect since the drift term of price
returns is the function of the volatility. In an effort to deepen the
discussion on the volatility, the next section explicitly species the
supply curve function that reects energy markets more appropriately.
2.2. One-factor SDV model for energy prices
In order to examine the characteristics of the SDV model in detail,
we propose one-factor SDV model for energy prices that is tractable
that reects an energy market observation. In order to do so, we need
to estimate supply curve. While consumption demand for natural gas
may have market's seasonality, the storage of natural gas is used to
mitigate the seasonality of consumption demand. Hence, the total
demand expressed by the sum of consumption and storage demands
will be less seasonal than the consumption demand. As the rst order
approximation, we aggregate the data across the seasons and years in
Fig. 3.
The exponential function may be a reasonable choice to express
the second order differentiable, monotonic, and increasing convex
function assumed in Proposition 1. It is because as is seen in
commodity asset pricing models the log transformations of prices
follow a simple stochastic process such as Brownian motion and a
transformation of a simple stochastic process. However, Fig. 3
regressing the exponential curve on the data (also see Appendix E)
implies that the rate of change of price with respect to the demand for
the scatter plots seems to be much larger than that for the expo-
nential function. For this reason, we employ the inverse BoxCox
 T. Kanamura / Energy Economics 31 (2009) 736747 739

Fig. 3. Supply curves.

transformation as the supply curve that can not only have the existence of the solution, we have the one-factor SDV model for
exponential function built in but also exhibit larger gradients of the energy prices that is tractable as in Eqs. (7)(9):
supply curve than the exponential function, although there exists a
problem on the asymptote for the supply function. In order to avoid dPt
= t dt + t dwt ; 7
the problem, we propose the following supply curve supposing the Pt
8
existence of a constant c close to and less than the asymptote ac2,3 > 1
>
>
where we introduce the inverse BoxCox supply curve until the >
> Pta Pt V1 + aa
>
>
< 0
supply limit c in the natural gas market and more than c use the
t = 11 8
exponential supply curve whose rst and second derivatives are the > 1
> 1 a B
>
>
aC
B1 + a 1 + a C Pt z1 + a a ;
same as the rst and second derivatives of the inverse BoxCox supply >
> @ A
>
: 1 + a Pt 1 a
curve at c, respectively.
8
>
>
1 2
t = k1 t + k2 t ; 9
>
> St a
>
> 1 + a St V c
>
< c D D 1 a
where = ; k1 = D, and k2 = 2 .
Pt = 5 c
>
> 1 The rst intuition of the model lies in the volatility term in Eq. (8)
>
> 

 
>
> 1 + aa 1 a St where it represents inverse leverage effect because t increases in Pt
>
: exp a St z c;
1 a 1 + a c due to negative a, i.e., for a b 0, t is monotonically increasing with
respect to Pt and bounded.5,6 Even if we use this tractable model, it can
where a is assumed to be less than or equal to 0. The intuition of this still represent the inuence of the supply curve convexity to the
supply curve is that smaller is a more convex is the supply curve. If a volatility that we intend to investigate in this paper. In addition, the
in Eq. (5) is zero, it collapses to the exponential function which second intuition lies in the drift term where it represents volatility-in-
corresponds to exponential transformations of a simple stochastic mean effect because t is a function of t.
process as in commodity price models such as Schwartz (1997). Else if While we provide the stochastic process for energy price volatility
a is less than 0, it has larger slope changes than the exponential.4 based on the supply and demand relationship, one may doubt if the
On the other hand, we, for simplicity, model demand for energy as solution exists or not. In order to clarify this problem, we offer the
a one-factor process: following lemma.
dDt = D dt + D dwt : 6 Lemma 1. Suppose the one-factor SDV model as in Eqs. (7)(9). Then,
there exists the solution.
Supposing the inelasticity of demand to prices, the energy prices
are given as the inverse BoxCox transformation of the demand. Proof. We offer the proof in Appendix A.||
Applying Ito's Lemma to Eq. (5) with Eq. (6) and then replacing the Hence we could secure the existence of the solution regarding one-
drift term for St c by that for St c in order to guarantee the factor SDV model.
We investigate the characteristics of the volatility in the one-factor
SDV model. According to Eq. (8),7 the volatility is proportional to the

5
Recall the expression of t and 1 + a N 0, then t is monotonically increasing with
2 a
P(t,S) 2C2 ([0, ) R). respect to Pt and bounded such as t V limPt Y t = 11 + a.
3 6
We suppose that c is the extremely high supply level that does not reach in In addition, we present the two-factor SDV model in Appendix B.
1
7
reality. We examine t for Pt V1 + a a , because the price level in reality is almost
1
4
Otherwise, i.e., a is positive, there does not exist the problem of this asymptote. included in this region. However, t for Pt z1 + a a also represents the inverse
 1
Thus, in the whole supply region, Pt = 1 + a Sct a holds. leverage effect we will later explain.
740 T. Kanamura / Energy Economics 31 (2009) 736747

price to the power of the negative BoxCox transformation para- where

meter.8 Since it stands for the curvature of the supply curve, the
t D n t
o
i1
magnitude of the volatility changes by the curvature. If the curvature Mt = 1 D D0 + 1 1 D ; ki = D 1 D :
D
is larger than that of the exponential function, i.e., a is negative, the
price gives a positive impact to the volatility. Else if a is zero implying
Suppose, in addition, that the squares of the past demand shocks
the exponential function, the price does not affect the volatility. It
only dominate the volatility as in Eq. (16). Then, the SDV model
means that if the supply curve becomes steeper than the exponential
collapses to ARCH(t)-M model:
function as in Fig. 3, the volatility increases in the prices, which is
often observed in energy markets and called as the inverse leverage 
2 2D X
t
2
effect. t 2
Ai;0 + Ai;2 t i ; 17
As another point of the SDV model, the volatility governs the drift c i=1
term of the price returns, implying that the model may be related to where
the volatility-in-mean effect as in (G)ARCH-M models. The next
section sheds light on the relationship between both models. X
a l
l
Ai;k = 1 l + 1 a b ; 18
l=k
c l k;i;l i;k
2.3. Why are (G)ARCH-M models applied to energy prices?
j n j 1 n
Many previous works on energy markets introduce (G)ARCH aj;i;n = n Cj Mt ki j n; aj;i;n = M j = n;
t t
models to capture the heteroskedasticity of the volatility in the prices.
However, to our knowledge the goodness of t to market data seems and

k
the main reason for the model selection. Down the line this section c + aMt + 1i
explores an economic reason behind the selection of a (G)ARCH-M bi;k = :
D
model for energy prices employing the SDV model.
To investigate the relationship between the SDV model to (G)ARCH The proofs to obtain Eqs. from (14)(18) are offered in Appendix C.
models, we represent the SDV model in discrete time transformed We investigate the inuence of the supply curve shape, i.e., the
from that in continuous one as in Eqs. (7) and (9) as follows: inverse Box-Cox transformation parameter (a), to the volatility. If it
2
is zero implying that the supply curve is exponential, the volatility
rt = t + t t ; t = 1 t + 2 t ; 10 2
is 2D and constant, injecting a = 0 into Eq. (16). What is more
c

important for energy markets, we analyze the inuence of more
where r t = log(P t + 1 ) log(P t ), 1 = DD , and 2 = 2a. Using upward sloping supply curve than the exponential one to the
Eqs. (5) and (9),9 2t leads to the function of demand for energy: volatility, i.e., a b 0. We then obtain the property for Ai,k such that if
a is negative, Ai,k is positive for all i and k. (The proof is offered in
2 2D
t = : 11 Appendix D.) Hence, if a is negative, then the volatility in Eq. (17)
c + aDt 2 always increases in the past demand shocks, because it includes the
squares of the past demand shocks and the positive coefcients. In
We assume that the demand follows an AR(1) model:10 addition, the negative a guarantees nonnegativeness of 2t in
Eq. (17).
Dt = 1 D Dt 1 + D + D t : 13
It is well-known that ARCH() collapses to GARCH(1,1), taking
time t as innity. The volatility, then, is likely to be modeled by ARCH
In addition, suppose that energy price returns are governed by the
(t) or GARCH(1,1) models, if the squares of the past demand shocks
one-factor SDV model. Then, the discrete time SDV model is expressed
approximately dominate it. Recalling that the model accommodates
by followings:
the volatility-in-mean effect as in Eq. (14), an existing GARCH(1,1)-M
2 model for energy prices has the foundations on the supplydemand
rt = t + t t ; t = 1 t + 2 t ; 14
relationship.
t = t t ; 15 Using the simple supply curve that reects energy markets,
Section 2 obtained the SDV model that can produce both the inverse
!k ! leverage and volatility-in-mean effects. Justifying the model for
2D X
a k Xt
2 k
t = 1+ 1 k + 1 Mt + ki t + 1i ; energy prices, Section 3 is devoted to empirically analyze the volatility
c2 k=1
c i=1 in the U.S. natural gas prices.
16
3. Empirical studies on the U.S. natural gas price volatility

3.1. Data
8
The one-factor SDV model is partly categorized in the C.E.V. (Constant Elasticity of
Variance) models because the volatility term is expressed by t = Pt a. The We use historical prices, storage, and consumption for the U.S.
remarkable point is that the C.E.V model parameter a has an economic standpoint natural gas to analyze the characteristics of the price volatility. The
such that a represents the curvature of supply curve. wellhead prices are used as the proxy of the equilibrium prices owing
1
9
We focus on the price region Pt V1 + a a for the supply curve and t, because the
to availability of the data. The storage and consumption are employed
price in reality is almost included in this region.
10
Since energy is often used for heating or cooling, it is well known that demand for to calculate the total equilibrium demand.11 These data are originally
energy links to temperature. Recalling that temperature is often modeled using an provided by the Energy Information Administration (EIA),12 including
Ornstein-Uhlenbeck process, we employ the process for the demand:

dDt = D D Dt dt + D dwt :
We obtain the AR(1) model by transforming the continuous time model for the
demand as in Eq. (12) to the discrete one. Note that the mean reversion of the demand
restricts the coefcient to 0 1 D 1 in general.
11
The validity of the assumption that total demands come from the sum of
consumption and storage demands is investigated in Section 4.
12
As of the rst two data, we directly download them from the homepage of the EIA.
The data is obtained from www.eia.doe.gov. On the other hand, the website of
Economagic.com offers the consumption demands collected from the EIA.
 T. Kanamura / Energy Economics 31 (2009) 736747 741

Fig. 4. Consumption demands/prices.

339 monthly observations from January 1976 to March 2004,
respectively.
3.2. Inverse leverage effect in the U.S. natural gas market
As we modeled in Section 2, the volatility of the price returns in the
one-factor SDV model is the function of the price whose parameter
represents the curvature of the supply curve shape. Using the
volatility model whose parameter is estimated from the U.S. natural
gas market, we show that the supply curve shape causes the inverse
leverage effect: the increase of volatility in prices.
Supposing that demand for natural gas is approximately inelastic
to the prices, the demandprice relationship leads to the supply
curve. In order to secure the inelasticity of demand, we need to

Fig. 5. Storage demands/prices.

742 T. Kanamura / Energy Economics 31 (2009) 736747
Fig. 6. The relationship between demand (= supply) and prices.
examine the irrelevancy of the total demandprice relationship to
seasonality. Fig. 4 shows that time series of consumption demands
and prices. The gure suggests that consumption demands are
strongly cyclical due to seasonality while prices are less cyclical than
the consumption demands. On the other hand, Fig. 5 shows that time
series of storage demands and prices. The gure suggests that peaks
of storage demands seem to occur in lowest prices, i.e., seasonality
adjustment of consumption demands will be occurred using storage
demand.
If these two demands are summed up, the cyclical trend of
consumption demand will be alleviated by storage demand. We
calculate the total demand in Fig. 1. The total demand seasonality is
more reduced than consumption demand comparing Figs. 1 to 4.
Hence, the total demandprice relationship is assumed to be
irrelevant to seasonality as the rst order approximation. Based on
this assumption, we use scatter plots as in Fig. 6 between total
demands and prices to estimate the supply curve as the rst order
approximation. According to the gure the equilibrium prices are a
monotonically increasing function of the equilibrium demands.
Then we model the supply curve that represents the scatter plots
as in Fig. 6. Because they seem more upward-sloping curve than the
exponential one, we employ the inverse BoxCox transformation for
the relationship between the equilibrium demands and prices (the
supply curve) as follows:

1
D a
Pt = 1 + a t + t :
c
According to the parameter estimates by the Nonlinear Least
Squares (NLS) as in Table 1, both a and c are statistically signicant.
Especially, the negative a ( 1.816) implies that the estimated
supply curve possesses a larger curvature than the exponential. Fig. 7
Table 1
Parameter estimation of the inverse BoxCox supply curve.
Parameter a c
Estimate 1.816
t-statistic 23.18
Log likelihood 449.29
AIC 902.59
SIC 910.23
drawing both the estimated exponential and inverse BoxCox supply
curves on the scatter plots illustrates that the latter ts better than
the former.13,14 Thus judging from the historical data, it is highly
probable that the U.S. natural gas market has the inverse BoxCox
transformation supply curve with the negative parameter a. Recalling
that the volatility of the one-factor SDV model is t =Pt a as in Eq. (8)
in the normal price range, the negative a implies that volatility t rise in
prices Pts, i.e., the inverse leverage effect, in the market.
As an interesting discussion, the scatter plots for the natural gas
demands and prices are comparatively deviated from the estimated
convex supply stack. According to Pirrong and Jermakyan (2008), the
deviation stems from the existence of market power. Hence, the
scatter plots may suggest the existence of market power in the U.S.
natural gas market.
3.3. Volatility-in-mean effect in the U.S. natural gas market
The drift term of the SDV model and the supply curve shape of the U.S.
natural gas market as in Subsections 2.3 and 3.2, respectively, predict the
existence of the volatility-in-mean effect in the natural gas prices. In
addition, as an example of energy prices, Deaves and Krinsky (1992)
obtains the effect in nearby oil futures prices.15 We, then, estimate the
EWMA GARCH(1,1)-M model linked to the SDV model using the monthly
natural gas prices.16 For simplicity, we employed the drift term is linear to
the volatility based on Eq. (4). The model is given by
rt = t + t t ; t = k0 + k1 t ; 20
19
t = t t ; 21
2 2 2
t = + t 1 + 1 t 1; 22
where k1 expresses the volatility-in-mean effect and represents the
GARCH(1,1) effect.
13
The parameters estimation for an exponential supply curve is shown in Appendix E.
14
The exponential function can represent the monotonically increasing trend of the
scatter plots, but both in high and low demand regions it causes large tting errors as
in Fig. 7. The observation is consistent with the estimation results because both AIC and
4.136 106
SIC (902.59 and 910.23) in Table 1 for the inverse BoxCox one are less than those
27.69
(932.01 and 939.65) in Table 7 for the exponential.
15
It is considered as the proxy of the spot prices.
16
We use the EWMA GARCH(1,1) model with the volatility-in-mean effect because
the GARCH(1,1) model violated the stationary condition for 2t .
 T. Kanamura / Energy Economics 31 (2009) 736747 743

Fig. 7. Supply curves.

Using the Maximum Likelihood Estimation (MLE) with the initial Eqs. (23) and (24) are reduced to the discrete time linear models:
values from the ordinary least squares, we identify the parameters. 
1 1
The results illustrated in Table 2 suggest the existence of the Xt = 1 X t Xt 1 + X t + X t uf1 Xt 1 ; Yt 1 ; t ; 26
volatility-in-mean effect in the U.S. natural gas prices because k1 and

are statistically signicant, respectively. 2 2
Yt = 1 Y t Yt 1 + Y t + Y t uf2 Xt 1 ; Yt 1 ; t ; 27

4. Validity of the assumption of equilibrium demand

which are linear time update equations in the EKF system. Eq. (19) is
the nonlinear model17:
The empirical studies in Subsection 3.2 substituted the sum of
consumption and storage demands for the equilibrium one. However,
1
it is not easy to guarantee the validity of the approximation, since the Xt + Yt a
Pt = 1+a + t uhXt ; Yt ; t ; 28
equilibrium one is not observed in the natural gas market. In an effort c
to do this, we offer a simple idea by using the supply curve shape. In
which constitutes the nonlinear measurement equation. Following
the beginning, we estimate the inverse BoxCox transformation
Welch and Bishop (2004),
parameter using an Extended Kalman Filter (EKF) where equilibrium
demand is unobservable and then compare it with the parameter from
 ! ! 1
!
Xt X t Xt 1 X t 1 t
the NLS as in Table 1. If both estimates are close enough, we consider = + At + Wt ; 29
Yt Y t Yt 1 Y t 1 2t
that the sum of consumption and storage demands may represent the
equilibrium demand as the rst order approximation under the !

assumption of inelastic demand to prices. Xt X t
Pt h X t ; Y t ; 0 + Ht + Vt t ; 30
The EKF is used for the parameter estimation of the supply curve Yt Y t
function and demand processes. Note that the improvement of the
estimation accuracy introduces two factor model for demands (Xt and where
Yt) as in Appendix B: 
X t = f1 X t 1 ;Y t 1 ; 0 ;
dXt = X X Xt dt + X dw1t ; 23


1 X t 0
dYt = Y Y Yt dt + Y dw2t ; 24 Y t = f2 X t 1 ;Y t 1 ; 0 , At = ,
0 1 Y t

 !
Et dw1t dw2t  = w dt: 25  1 1  1 1
X 0 1
Wt = , Ht = 1 + aXt +c Yt a 1c 1 + aXt +c Yt a ,
0 Y c

h i

t w t
Vt = 1, V 1t ; 2t = uQt , V t  = muRt .
Table 2 w t t
Parameter estimation of GARCH-M model.
Tables 3 and 4 offer the complete set of the EKF equations which
Parameters k0 k1 include time and measurement update equations so as to calculate
Estimates 0.021 0.290 2.87 10 5 0.212 the a priori estimate error covariance matrix ( t ) and the a
Standard errors 0.000 0.008 3.71 10 10 0.002 posteriori estimate error covariance matrix (t), respectively. Note
Log likelihood 4.83 102
AIC 9.64 102
SIC 9.60 102 17
We assume that seasonality can be negligible i.e., DCt = Xt and DtS = Yt.
744 T. Kanamura / Energy Economics 31 (2009) 736747

Table 3 the assumption, we model the demand curve by the inclined linear
EKF time update equations. function, though the supply curve is the same as that in the previous
X
t = f1(Xt 1,Yt 1,0)
sections. The inelasticity is then examined by applying the nonlinear

Y
t = f2(Xt 1, Yt 1,0) two-stage least square estimation (N2SLS) in Amemiya (1974) to the
T T
t = Att 1At + WtQt Wt simultaneous equations of the supply and demand curves.
For the identication, it is important to determine the instruments.
Although the candidate that causes the equilibrium demand uctua-
Table 4 tion is temperature, it is known that the equilibrium demand for
EKF measurement update equations. energy linearly increases in the temperature. In addition, the variable
Kt = T T T 1
t Ht (Htt Ht + VtRtVt )
at the prior period t 1 is often employed as the instrument for time
X t = X


t + Kt(Pt h(Xt ,Yt ,0)) series data, since it must be orthogonal to model errors. Using the
Yt = Y


t + Kt(Pt h(Xt ,Yt ,0)) demand at time t 1 as the instrument, we, then, conduct the N2SLS
t = (I KtHt)
t
estimation of the supply and demand equations.19 The model is

1
D a
Table 5
g1 Pt ; Dt = Pt 1 + a t = t ; 34
c
Parameter estimation by using an extended Kalman lter.

Model parameter a c X X X g2 Pt ; Dt = Pt + Dt + 1 = t : 35
Estimate 0.58 4.11 106 0.84 1.25 106 1.09 105
Standard error 0.05 1.00 0.25 1.10 10 1 9.96 10 1
Eqs. (34) and (35) represent the supply and demand curves,
Model parameter Y Y Y w m respectively. If is not statistically signicant and is negative and
Estimate 0.14 1.48 106 869.19 1.00 0.01 statistically signicant, then the demand curve becomes vertical
Standard error 0.03 9.99 10 1 1.00 7.07 10 4 0.00 implying that the demand is inelastic to the prices.
Log likelihood 7.76
AIC 3.55 101
Based on these regression equations as in Eqs. (34) and (35), we
SIC 1.55 101 estimate the parameters by solving the following minimization
problem:
 1
that 0
we dene 1 the a priori estimate error and the covariance by T T
i = arg min gi Z Z Z
T
Z gi ; 36
i
X X
e
t u
@ t t A
and T
t E[et et ], respectively and also dene
!
Yt Y t where i = 1,2, 1 = (a,c), 2 = (, ), and Zt = [1Dt 1] is the matrix of
the a posteriori estimate error and the covariance by et u X t X t
the instruments.
Yt Y t
Judging from the standard errors in Table 6, the parameters except
and t E[et eTt ], respectively. are statistically signicant. In addition, is negative. These results
Using the recursive updates of time and measurement update guarantee the inelasticity of the demand to the prices. Furthermore,
equations as in Tables 3 and 4, measurement errors (ePt) and the since a is negative (1.666), it suggests that the slope changes of the
covariance matrices (t) are supply curve are more dramatic than the exponential. It is consistent
 with the result as in Subsection 3.2 where the supply curve is
ePt = Pt h X t ; Y t ; 0 ; 31 estimated subject to the demand inelesticity to the prices. Thus, even
the identication of the simultaneous equations of the supply and
t = Ht
T T
t Ht + Vt Rt Vt : 32 demand curves has validated the demand inelasticity assumption.

Using the measurement errors and the covariance matrices, the 6. Conclusions and directions for future research
parameters set () in Eqs. (19), (23), (24), and (25) are estimated by
the MLE: This paper has proposed a volatility model, which we call a supply
and demand based volatility (SDV) model for energy prices
X
N X
N
1 T
= arg min lnjt j + ePt t ePt ; 33 characterized by the supplydemand relationship. We have illustrated

t =1 t =1 that it can produce both the time-varying volatility and the
volatility-in-mean effect that characterize energy prices. Addition-
where = (a,c, X, X, X, Y, Y, Y, w, m).
ally, it has been shown that the SDV model can produce the inverse
According to the results in Table 5, the BoxCox transformation
leverage effect often seen in energy markets, supposing that the
parameter (a) is 0.58 (negative and statistically signicant),
inverse BoxCox transformation function represents the supply curve.
although it is a little larger than that by the demand-price based
Moreover, this paper has shown that the existing (G)ARCH-M model
estimation ( 1.816) as in Table 1.18 In the sense that both parameters
has the foundations on the supplydemand relationship, being
take negative values, we have obtained almost the same result as that
derived from the discrete time SDV model with some approximation.
in Table 1, which assumes that the sum of consumption and storage
The empirical studies have analyzed the volatility in the U.S.
demands represents the equilibrium demand. It indicates that the
natural gas prices using the SDV model. First, to examine the existence
assumption may be valid as the rst order approximation.

5. Simultaneous equations with elastic demand   

 

 ui;t E ui;t
19
Setting u1,t =t and u2,t =t, we have E Z Vt ui;t = E =
 =
 
Dt 1 ui;t E D t1 E ui;t
This section investigates the validity of the demand inelasticity 0
for i = 1,2. As is explained in e.g., (Wooldridge, 2002), these instruments
assumption in the SDV model by using the natural gas data. Relaxing 0
h i
satisfy the efciency condition of the N2SLS estimation for i=1,2: E u2i;t Z Vt Zt =
" !#
1 Dt 1 h i
18
The inverse leverage effect is still observed, even if the equilibrium demand is E u2i;t = E u2i;t EZ Vt Zt :
assumed to be unobservable. Dt 1 D2t 1
 T. Kanamura / Energy Economics 31 (2009) 736747 745

Table 6 Then for a 0, s(x) and m(x) are continuous and differentiable
Simultaneous estimation of supply and demand curve parameters by N2SLS. with respect to x. It is veried by the followings. Note that for x b 0 we
Parameter a c set s(x) = 0 and m(x) = 0.
Estimate 1.666 3,939,852 1.158 2.10 10 6
9.97 10 7 Asx a
Standard error 0.423 567,355 0.769 lim = lim a + 1x =0 A8
xY + 0 Ax xY + 0

Asx 1 a
lim 1 = A9
xY1 + aa Ax 1 + a
of the inverse leverage effect: volatility increases in prices, we
estimated the inverse BoxCox transformation parameter for the
supply curve using the historical prices and demands. The negative Asx 1 a
lim = A10
estimate implied that the U.S. natural gas possesses the effect. It is xY + Ax 1 + a
consistent with energy market observations in Eydeland and
Wolyniec (2003). Second, we showed that volatility-in-mean effect Amx a 1 a 2 2a
lim = lim a + 1 D x + 2a + 1 x =0
xY + 0 Ax xY + 0 D 2
is statistically signicant in GARCH(1,1)-M model. It supports the
A11
volatility-in-mean models for energy prices.
Finally in order to enhance the SDV model, we have empirically
examined the validity of two model assumptions using the natural gas Amx 1 a 1 a1 2a 2
lim 1 = D + A12
data. First, we have shown that as the rst order approximation the xY1 + aa Ax D 1 + a 2 1 + a2
sum of consumption and storage demands approximates to the
equilibrium one by comparing the inverse BoxCox transformation

Amx 1 a 1 a 2 1a 2
parameter estimated from the extended Kalman lter with that from lim = D + A13
xY + Ax D 1 + a 2 1 + a
the nonlinear least squares. Second, we have shown that the demand
is inelastic to the prices by conducting the identication of The results satisfy the Lipschitz condition to the coefcients s(x)
simultaneous equations for the supply and demand curves with the and m(x) as in SDE (A1). In conclusion, it suggests that the one-factor
nonlinear two-stage least square estimation. SDV model as in Eq. (A1) has a unique strong solution.
For future research, we will conduct the empirical studies using
more frequent natural gas prices like weekly and daily ones. The Appendix B. Two-factor SDV Model
applications to other energy prices like crude oil and heating oil are
possible extensions for our further studies. We formulate the two-factor SDV model that incorporates
consumption and storage demands. Denoting by DCt consumption
Appendix A. Existence of solution to one-factor SDV model demand for energy, the deviation (Xt) from the average demand (D Ct )
is assumed to follow a mean reverting process, as is often observed in
We show the existence of the solution to the one-factor volatility energy markets:
model for a 0 and constant D. Eqs. (7)(9) are given by the
following: C C
Dt = Xt + Dt ; B14

dPt = mPt dt + sPt dwt ; A1 dXt = X X Xt dt + X dw1t : B15

sx = xx; A2 Then, we introduce the storage demand (DS) which is also divided
into the mean reverting deviation process (Yt) and the average (D tS).
mx = xx; A3 The decomposition expresses the seasonal pattern and the deviation
of storage demand. The storage demand model is given by
8
>
> a
1 S
Dt = Yt + Dt ;
S
B16
>
> x x V1 + a a
< 0 1
x = 1 a @
1
a1 + aa A A4
>
> x z1 + aa ;
1
dYt = Y Y Yt dt + Y dw2t : B17
>
> 1+
: 1 + a 1 ax
The mean reversion reects the characteristic of storage such that
1 a 2
the deviation is expected to reach a certain level in a long run if it is far
x = D x + x : A5 from the level.
D 2
In order to obtain the prices, we use a supplydemand relationship
Since (x) is bounded where the upper bound of (x)is set as k, where we assume that the demand curve is vertical and the supply
we have curve is an inverse BoxCox transformation function. The equilibrium
prices are recovered from the supply curve, substituting the demand

2 2 2 2 2
OsxO = O xO OxO V k 1 + OxO : A6 for the supply:
8
>
1
>
> Xt + Yt + f t b a
In addition, (x) is also bounded (k) since it is expressed by the >
> Xt + Yt + f t b V c
< 1+a c
quadratic function of (x): Pt =
>
> 1

 
>
> 1 + a exp 1 a Xt + Yt + f t b
> a
 : 1 a a Xt + Yt + f t b z c;
2 2 2 2 2 1 + a c
OmxO = O xO OxO V k 1 + OxO : A7
B18

The results satisfy the growth condition to the coefcients (s(x) where f(t) denotes the sum of D  Ct and D tS. We dene by Z and qt as
and m(x)) of the SDE (A1) (e.g., (Dufe, 2001)). follows: Zt = Xt + Yt +c f t b , qt = 1c 1 Y Yt . Applying Ito's Lemma to
X
746 T. Kanamura / Energy Economics 31 (2009) 736747
n o
Table 7 We set Mt = 1 D t D0 + D
D
1 1 D t and ki =D(1D)i 1.
Parameter estimation of an exponential supply curve.
Substituting Eq. (C29) into Eq. (C27), we obtain Eq. (16). We, then, have
Parameter b c Eqs. from (14)(16) as the discrete time SDV model.
Estimate 1,035,777 853,644 If the past demand shocks and the squares dominate the volatility,
t-statistic 12.07 8.38 k j = 0 holds for k j. We, then, write Dt to the power of n as:
Log likelihood 464.00
AIC 932.01 X
n X
t
n n j
SIC 939.65 Dt aj;i;n t + 1 i; C30
j=0 i=1

Eq. (B18) by using Eqs. (B15) and (B17), the two-factor SDV model for where aj,i,n = nCj M jt kni j (j n) and aj;i;n = 1t M tn j = n. Injecting
energy prices is expressed by followings20: Eq. (C30) into Eq. (C27), we have

dPt 2 2D X
X
t
k
= t dt + t dwt ; B19 t i;k t + 1 i ; C31
Pt 2
c k=0 i=1
 P  l
1 2 where i;k = 1l l + 1 ac al k;i;l . Expressing t by Dt
dZt = Z + X qt Zt Z dt + Z dwt ; B20 l = k
2 from Eq. (11) and employing the denition of t, we have


dqt = q Y qt dt + q dw2t ; B21 c + aDt
t = t : C32
D
8
>
> P a 0
1
Notice that
>
>
< Z t 11 Pt V1 + a a
t = 1 a B aC
B1 + a1 + a C
1
B22 X
t
>
> Pt z1 + a a ;
> Z 1 + a @
> 1 aPt A Dt = Mt + kj t + 1 j: C33
: j=1

 
1 2 1 1 2 Substituting Eqs. (C32) and (C33) into Eq. (C31), we have
t = 1 a t + Z Z + X qt Zt t ; B23
2 Z 2 Z Pt !k
+ 1i
D X X
2
2
t c + aMt + 1i +a j=1 kj t + 1ij k
t i;k t + 1 i:
c2 k = 0 i = 1 D
Et dwt dw2t  = dt; B24
C34
1 1 2
Z = f Vt + X + Y + X f t b + Z ; B25 Remind that by denition t + 1 i is expressed by t + 1 i and that
c 2
kj = 0 for k j when the past demand shocks and the squares
q dominate the volatility, we have
1
Z = 2X + 2Y + 2w X Y ; B26
c
2 2D X
X
t
k
1 X Y X Y t 2
Ai;k t + 1 i; C35
where dwt = X dw1t + Y dw2t , q = Y , q = Y, c k=0 i=1
c Z cX cX
X w + Y
and = .
Z
 k
We nd that, like the one-factor SDV model, the two-factor SDV where Ai,k = i,kbi,k and bi;k = c + aMt + 1 i
D
. If the squares of the past
model is also classied in the 1C.E.V. models, judging from the volatility demand shocks only dominate the volatility, we have Eq. (17).||
in Eq. (B22) for Pt V1 + a a. The drift term in Eq. (B23) stems from
the volatility. It leads to the volatility-in-mean model. Appendix D. Proof of positive Ai,k

Appendix C. Proof of discrete-time SDV model The mean reversion of the demand process whose implication is
0 1 D 1 causes positive ki. It also causes the positive and constant
Employing the Taylor expansion with respect to Dt around 0 for long term average of Mt represented by D. These positives of ki and Mt
D
Eq. (11), we have lead to positive al k,i,l as in Eq. (18). Additionally bi,k, c, and a are all
! positive by denition. These results lead to positive Ai,k.||
2 X
a k
2 k k
t = 2D 1+ 1 k + 1 Dt : C27
c k=1
c Appendix E. Exponential supply curve parameters

We assume that the demand process follows AR(1) model: We model the supply curve in the U.S. natural gas market using
St b

Dt = 1 D Dt 1 + D + D t : C28 Pt = e c . Since the supply is equivalent to the sum of consumption

It is reduced to

D n o Table 8
t t
Xt
i1
Dt = 1D D0 + 1 1D + D 1 D t + 1 i:
Parameter estimation of consumption and storage demand processes using MLE.
D i=1
Model parameter cX X 2X cY Y 2Y XY
C29 Estimate 0.00 0.89 7.34 10 9
0.00 0.39 7.10 109 0.59
Standard error 1.00 0.02 1.00 1.00 0.04 1.00 0.03
20 Log likelihood 8.53 103
We employ Eq. (B23) in order that the two-factor SDV model is consistent with the
AIC 1.71 104
one-factor SDV model although Eq. (B23) is different from that obtained by the supply
SIC 1.71 104
curve in the supply region more than c.
 T. Kanamura / Energy Economics 31 (2009) 736747 747

and storage demands in equilibrium, we have St = DCt + DtS. Substitut- consumption and storage demands may be justied in employing the
ing it into the exponential function, we have mean reverting processes.

DCt + DSt b
Pt = e c : E36 References

Amemiya, T., 1974. The nonlinear two-stage least-squares estimator. Journal of

We estimate both b and c in Eq. (E36) by using the NLS. The results Econometrics 2, 105110.
are illustrated in Table 7. Judging from the t-statistics as in Table 7, Barlow, M.T., 2002. A diffusion model for electricity prices. Mathematical Finance 12,
287298.
both parameters are statistically signicant. Bollerslev, T., 1986. Generalized autoregressive conditional heteroskedasticity. Journal
of Econometrics 31, 307327.
Appendix F. Parameter estimation of demand processes Brennan, M.J., 1991. The price of convenience and the valuation of commodity
contingent claims. In: Lund, D., Oksendal, B. (Eds.), Stochastic Models and Option
Models. North Holland, Amsterdam.
We calibrate the consumption and storage demand models in Clewlow, L., Strickland, C., 2000. Energy Derivatives: Pricing and Risk Management.
Appendix B to the historical data. Since the corresponding deviations Lacima Publications, London.
Cox, J., 1975, Notes on Option Pricing I: Constant Elasticity of Variance Diffusions,
are Eqs. (B15) and (B17), respectively, the discrete time models are Working paper, Stanford University (reprinted in Journal of Portfolio Management,
1996, 22 1517).
Xt + 1 = cX + X Xt + X ; F37 Deaves, R., Krinsky, I., 1992. Risk premiums and efciency in the market for crude oil
futures. The Energy Journal 13, 93117.
Dufe, D., 2001. Dynamic Asset Pricing Theory, Third Edition. Princeton University
Press, New Jersey.
Yt + 1 = cY + Y Yt + Y ; F38 Dufe, D., Gray, S., Hoang, P., 1999. Volatility in energy prices. In: Kaminski, Vincent
(Ed.), MANAGING ENERGY PRICE RISK. RISK PUBLICATIONS, London.
! Emanuel, D., MacBeth, J., 1982. Further results on the constant elasticity of variance call
2X XY X Y option pricing model. Journal of Financial and Quantitative Analysis 17, 533554.
where =(X,Y)~N(,), =(0,0), and = . Engle, R., 1982. Autoregressive conditional heteroscedasticity with estimates of the
XY X Y 2Y variance of united kingdom ination. Econometrica 50, 9871007.
The MLE is applied to the models using the initial values from the least Engle, R., Lilien, D., Robins, R., 1987. Estimating time varying risk premia in the term
structure: the Arch-M model. Econometrica 55, 391407.
squares for each process. Eydeland, A., Wolyniec, K., 2003. Energy and Power Risk Management: New
According to the results shown in Table 8, c X is not statistically Developments in Modeling, Pricing, and Hedging. John Wiley & Sons, Inc., Hoboken.
signicant, while X and 2X are statistically signicant judging from Gibson, R., Schwartz, E., 1990. Stochastic convenience yield and the pricing of oil
contingent claims. Journal of Finance 45, 959976.
the standard errors. Similarly, c Y is not statistically signicant, Heston, S., 1993. A closed-form solution for options with stochastic volatility with
while Y and 2Y are statistically signicant. Using the conversion applications to bond and currency options. The Review of Financial Studies 6,
formula as in e.g., Clewlow and Strickland (2000), we have the 327343.
Kanamura, T., hashi, K., 2007. A structural model for electricity prices with spikes:
parameters of i, i, and i for i = X and Y in Eqs. (B15) and (B17):
r measurement of spike risk and optimal policies for hydropower plant operation.
2 ln Energy Economics 29, 10101032.
i = lni ; i = ci 1 lni ; i = i 2 i1 a n d w : w = Kolos, S.P., Ronn, E.I., 2008. Estimating the commodity market price of risk for energy
i i
X Y X + Y prices. Energy Economics 30, 621641.
XY : We, then, have
X Y 1 e X + Y Pindyck, R.S., 2004. Volatility in Natural Gas and Oil Markets, Working paper.
Massachusetts Institute of Technology.
4 Pirrong, C., Jermakyan, M., 2008. The price of power: the valuation of power and
X = 0:11; X = 0; X = 9:05 10 ; Y = 0:94; weather derivatives. Journal of Banking and Finance 32, 25202529.
5 Schwartz, E.S., 1997. The stochastic behaviour of commodity prices: implication for
Y = 0; Y = 1:26 10 ; and w = 0:61: valuation and hedging. Journal of Finance 52, 923973.
Welch, G., Bishop, G., 2004. An Introduction to the Kalman Filter, Working paper.
University of North Carolina at Chapel Hill.
The estimates indicate that the long term averages of both Xt and Wooldridge, J.M., 2002. Econometric Analysis of Cross Section and Panel Data. The MIT
Yt are zero due to insignicance of cX and cY. In addition, since the Press Cambridge.

parameters for the two-factor model are signicant besides X and Y,