The Journal of Finance - 2002 - Bessembinder - Equilibrium Pricing and Optimal Hedging in Electricity Forward Markets

15406261, 2002, 3, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/1540-6261.00463 by Indian Institute Of Technology, Kharagpur Central Library, Wiley Online Library on [23/12/2022].
See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
THE JOURNAL OF FINANCE * VOL. LVII, NO. 3 * JUNE 2002
Equilibrium Pricing and Optimal Hedging

in Electricity Forward Markets
HENDRIK BESSEMBINDER and MICHAEL L. LEMMON*
ABSTRACT
Spot power prices are volatile and since electricity cannot be economically stored,
familiar arbitrage-based methods are not applicable for pricing power derivative
contracts. This paper presents an equilibrium model implying that the forward
power price is a downward biased predictor of the future spot price if expected
power demand is low and demand risk is moderate. However, the equilibrium for-
ward premium increases when either expected demand or demand variance is high,
because of positive skewness in the spot power price distribution. Preliminary em-
pirical evidence indicates that the premium in forward power prices is greatest
during the summer months.
WHOLESALEPOWERMARKETS,where producers trade electricity among them-

selves and with power-marketing and power-distribution companies, have
grown rapidly in recent years.1 The U.S. Department of Energy (2000) re-
ports that U.S. wholesale power transactions during 1999 amounted to ap-
proximately 2.6 billion megawatt hours (MWh), or about $85 billion. The
U.S. wholesale power market is soon likely to comprise the world's largest
commodity market.
Electricity as a commodity has many interesting characteristics, most of
which stem from the fact that it cannot be economically stored.2 Market-
* Bessembinder is from the David Eccles School of Business, University of Utah and the
Goizueta Business School, Emory University. Lemmon is from the David Eccles School of Busi-
ness, University of Utah. This paper is derived from an earlier manuscript titled "Pricing, Risk
Sharing, and Profitability in the Deregulated Electric Power Industry" that was developed
while both authors were on the faculty of the Arizona State University College of Business. The
authors thank Douglas Cochran, Alexander Eydeland, Jean Gray, Paul Joskow, Vince Kamin-
ski, John Dalle Molle, Steve Norris, Duane Seppi, Ren6 Stulz, Catherine Wolfram, and an
anonymous referee for many helpful comments. The authors also benefited from the partici-
pation of seminar participants at the University of Arizona, Arizona State University, the Uni-
versity of Colorado,Emory University, Rice University, the University of Texas at Austin, the
University of Utah, Washington University, Instituto Technol6gicoAut6nomo De Mexico, the
1999 University of California Energy Institute POWER conference, and the 2000 American
Finance Association Meetings.
1 The growth in power trading is partially attributable to ongoing deregulation of the indus-
try. Most deregulation scenarios call for the separation of power production from transmission
and retailing, with production and retailing opened to competition.
2 Though power cannot be stored, potential energy can be stored in the form of fuel stock-
piles or water behind dams. The capacity to quickly convert potential energy to power remains
limited, however, as evidenced by the differential between daytime and nighttime power prices.
1347
15406261, 2002, 3, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/1540-6261.00463 by Indian Institute Of Technology, Kharagpur Central Library, Wiley Online Library on [23/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
1348 The Journal of Finance
1000
100
10
04/01/1997 10/18/1997 05/06/1998 11/22/1998 06/10/1999 12/27/1999 07/14/2000
Date
Figure 1. Average daily on-peak power prices in dollars per megawatt hour, for de-
livery at PJM from April 1997 through July 2000.
clearing prices are volatile because inventories cannot be used to smooth

supply or demand shocks. The absence of storage also allows for predict-
able intertemporal variation in equilibrium prices. Power prices for day-
time delivery are typically more than twice as high as for nighttime delivery,
and power prices during the summer are predictably higher than during
the temperate months of spring and fall. In addition, power prices are
subject to sudden, but generally temporary, upward spikes. Some of these
features of wholesale power prices are evident in Figure 1, which displays
average daily on-peak spot power prices in the Pennsylvania, New Jersey,
Maryland (PJM) market over the period April 1997 through July 2000. The
standard deviation of percentage changes in the average daily power prices
displayed on the figure is 34 percent. By comparison, the standard devia-
tion of daily returns on the S&P 500 Index during the most volatile single
month in recent decades, October 1987, was 5.7 percent. Large, temporary,
upward spikes in PJM prices are evident in Figure 1 during each summer
season. In another dramatic and well-publicized example, power prices in
Equilibrium Electricity Forward Pricing 1349
the U.S. Midwest briefly rose from near $30/MWh to over $7,000/MWh
during June 1998.3
Finally, and most important for purposes of this paper, the inability to
store power means that the no-arbitrage approach to pricing derivative se-
curities cannot be applied in the usual manner. The well-known cost-of-carry
relationship links spot and forward prices as a no-arbitrage condition. How-
ever, the arbitrage strategies required to enforce the cost-of-carry relation-
ship include purchasing the asset at the spot price and storing it for subsequent
sale at the forward price.4 Since this strategy cannot be executed in power
markets, forward prices for electricity need not conform to the cost-of-carry
relationship.
Pirrong and Jermakyan (1999) and Eydeland and Geman (1999) also ob-
serve that the no-arbitrage approach does not apply to power derivatives.
Pirrong and Jermakyan note that electricity forward prices will differ from
expected delivery date spot prices due to an endogenous market price of
power demand risk. However, they do not attempt to model the determi-
nants of the market price of power risk, while we do. Eydeland and Geman
focus on electricity options, noting (p. 72) that:
there is another important consequence of non-storability: using the spot

price evolution models for pricing power options is not helpful, since
hedges involving the underlying asset, i.e., the famous delta hedging,
cannot be implemented, as they require buying and holding power for a
period of time.
Eydeland and Geman present a pricing model for power options that relies
on assumptions regarding the evolution of forward power prices. We adopt
an equilibrium approach and explicitly model the economic determinants of
market clearing forward power prices.
The equilibrium model presented here relies on the assumptions that prices
are determined by industry participants rather than outside speculators,
and that power companies are concerned with both the mean and the vari-
ance of their profits. We evaluate power producers' and retailers' net de-
mands for forward contracts, and obtain closed form solutions for the
equilibrium forward power price and for optimal forward positions. The im-
plications of the model are illustrated with a set of simulations.
The model generates the testable implication that the forward risk pre-
mium is a function of the difference between two covariance terms that can
be related to the variance and skewness of spot power prices. Positive skew-
ness in wholesale power prices is attributable to the nonstorability of power
3 For example, see "Staff Report to the Federal Energy Regulatory Commission on the Causes
of Wholesale Electric Pricing in the Midwest During June 1998," available at http://www.
ferc.fed.us.
4 For a description of the standard cost-of-carry forward pricing model see, for example,
MacKinlay and Ramaswamy (1988) or Kawaller, Koch, and Koch (1987).
combined with convexity in the industry supply curve. When expected power
demand is low and demand variability is modest (as might be expected dur-
ing the temperate months of spring and fall), there is little skewness in spot
prices, and power retailers' desire to hedge their revenues leads to a down-
ward bias in equilibrium forward prices. In contrast, when expected demand
is high relative to capacity or when demand is more variable, the distribu-
tion of spot power prices becomes positively skewed. Short forward positions
incur large losses if upward spikes in spot prices occur, and the equilibrium
forward price is bid up to compensate for skewness in the spot price distri-
bution. The U.S. demand for power is largest and most variable in the sum-
mer.5 The model, therefore, predicts an upward bias in U.S. forward power
prices for summer delivery.
We also evaluate optimal forward positions. Power-producing firms' opti-
mal forward market positions depend on forecast output and on the skew-
ness of power demand. Power retailing firms' optimal forward positions depend
on forecast usage, and on the interaction between local and system demand
as measured by power demand betas.
Finally, we conduct some preliminary empirical analysis using the avail-
able short time series of electricity forward and spot delivery prices. The
evidence is generally supportive of the model's predictions. Forward prices
for summer delivery exceed forecast spot prices, while forward prices for
spring and fall delivery are similar to, or slightly less than, forecast spot
prices. Regression analysis confirms that the premium in forward power
prices as compared to expected spot prices is positively related to the level of
power demand.
Routledge, Seppi, and Spatt (1999) also consider the equilibrium pricing of
electricity contracts.6 They focus on linkages between the natural gas and
electricity markets (i.e., the "spark spread") that arise from the fact that
natural gas can be stored or converted into electricity. They obtain several
implications, including predictions of mean reverting spot prices and unsta-
ble electricity/fuel price correlations. Their model also implies that electric-
ity prices will be positively skewed, while we obtain the implication that
skewness will affect the equilibrium forward premium and optimal forward
positions.
Despite the unique features of power markets that arise from nonstorabil-
ity, studying power markets is likely to lead to broader insights regarding
market learning and financial innovation. As noted above, the model pre-
sented here is one of limited participation, with prices determined by the
trades of those who produce and deliver power rather than by speculators
from outside the power industry. The limited participation assumption ap-
pears reasonably accurate at present, as discussed in Section I.C below. How-
5 See Wall Street Journal (2000).

6 There are also several papers that consider pricing of derivatives on storable energy com-
modities such as oil and natural gas. These include Amin, Ng, and Pirrong (1995), Gabillion
(1995), and Routledge, Seppi, and Spatt (2000).
ever, the existence of a nonzero premium in forward power prices as implied

by the model and the preliminary data provides incentives for financial in-
termediaries to create instruments, for example, power-indexed bonds, to
allow outside speculators to include power positions in their portfolios. The
presence of outside speculators would then be expected to decrease the mag-
nitude of the forward premium. A similar evolution has been described by
Froot (1998, 2001) in the case of catastrophe insurance, an industry which
evolved during the 1990s from a limited participation market with premi-
ums well above actuarial value, to a broader market with more modest pre-
miums, after insurers issued catastrophe-linked bonds. It will be of interest
to observe whether financial contracts that allow outside speculators to take
positions in power markets are introduced, and if power forward prices then
converge toward average spot prices.
Finally, although we focus explicitly on hedging decisions and equilibrium
in wholesale electricity markets, we believe that some insights obtained in
our analysis will prove useful to a broader set of hedging problems. Previous
studies concerning forward hedging strategies, such as Anderson and Dan-
thine (1980) and Hirshleifer and Subramanyam (1993), among others, de-
velop expressions showing that optimal forward positions depend on the
covariance of revenues (price times quantity) with prices. We extend the
analysis by making endogenous both market prices and output decisions,
obtaining the implication that price skewness is relevant to hedging deci-
sions. Our conjecture is that equilibrium approaches like ours will also prove
relevant in a broader array of commodity hedging applications.
This paper is organized as follows. Section I outlines the general setup of
the model and discusses the demand for risk reduction on the part of power
industry participants. Section II derives the equilibrium forward price and
the optimal forward positions of power producers and retailers. Section III
presents some preliminary empirical evidence, and Section IV concludes.
I. Power Production, Wholesale Markets,

and the Demand for Risk Reduction
A. The General Setup
The inability to store power implies that traditional cost-of-carry models
for pricing forward contracts do not readily apply to power markets. We
therefore use an equilibrium approach. Our primary goal is to assess equi-
librium forward power prices and optimal hedge positions for power firms.
To do so requires a model of the underlying spot market and of the trans-
actions that participants will optimally make there. We analyze power pro-
duction during a single future time period, and assume that there is no
uncertainty in spot markets, that is, that power companies are able to fore-
cast demand in the immediate future with precision, and are able to enter
contracts in the wholesale market at known prices. Actual spot power mar-
kets operate with short delivery horizons: The most active markets are for
next-hour and next-day delivery. In practice, power demand over a short

horizon such as the next hour can be forecast with substantial, though not
complete, precision.
The model includes NP power producers, each producing power using iden-
tical technology and selling it to power retailers in a competitive wholesale
market.7 We assume that power can be transmitted without cost.8 There are
NR power retailing (or power marketing) firms that purchase power in the
wholesale market and sell it to final consumers at a fixed unit price, de-
noted PR. Retail customers consume as much power as they desire at that
price.9 The realized demand for retailer i is denoted by the exogenous ran-
dom variable QRi-
Each producer i is assumed to have a power production cost function of the
form
a
TCi = F +-(Qpj)c, (1)
where F represents fixed costs, Qpi is the output of producer i, and c is a

constant greater than or equal to two. This simple cost function captures
many features relevant to the production of electricity. It implies that mar-
ginal production costs increase with output, which reflects the fact that the
industry employs an array of production technologies and fuel sources with
differing out-of-pocket costs, including hydro, nuclear, coal, oil, and natural
gas. Increasing marginal costs are also consistent with empirical regulari-
ties such as higher daytime than nighttime power prices.
7 Whether wholesale power markets are in fact competitive has been the subject of consid-
erable analysis. Green and Newbery (1992), Newbery (1995) and Wolfram (1999) conclude that
there are an insufficient number of suppliers in the British power markets. Borenstein and
Bushnell (1999) conclude that the California electricity markets have some potential for market
power, but view the resulting costs to be small relative to the potential gains from deregulation.
Joskow and Kahn (2001, p. 30) conclude that recent prices in the California market "far ex-
ceeded competitive levels." See also the survey by Joskow (1997). In relying on the assumption
that wholesale power markets are competitive, our analysis can be viewed as the limiting case
obtained when market mechanisms are developed to mitigate possible producer market power.
8 We abstract from analyzing transmission costs in order to focus on an alternate set of
issues. The pricing of transmission rights is itself an interesting and challenging issue, made
complicated by the presence of externalities within the distribution system. For a discussion of
possible market-oriented solutions to the allocation of transmission capacity, see Joskow and
Tirole (1999) and Harvey, Hogan, and Pope (1996).
9 Note that this approach does not rule out predetermined seasonal variation in retail prices-
the assumption is that the retail price does not respond to real-time shocks. Despite the famil-
iarity of the fixed-price retail contract, it is actually a complex derivative: a variable quantity
forward contract. The main argument in favor of the fixed price contract is that retail custom-
ers do not have a comparative advantage in bearing price risk. However, a fixed retail price
fails to provide incentives to curtail demand at times when wholesale prices are high. Hybrid
contracts that allow for some variation in retail prices in response to wholesale prices have
begun to evolve, being offered mainly to larger and more sophisticated final customers.
If the cost parameter c is greater than two, marginal costs increase at an

increasing rate with (are convex in) output. This allows some flexibility to
account for complexities that are not formally modeled, including the use of
relatively inefficient peaking plants to meet high levels of demand, as well
as the possibility that constraints on production and transmission capacity
will become binding when demand is high. The rapid increases in marginal
costs implied by production levels that approach capacity can be approxi-
mated by considering the effect of increasing the cost convexity parameter, c.
Note also that if c is greater than two, the distribution of power prices will
be positively skewed even when the distribution of power demand is symmetric.
B. The Demand for Risk Reduction

The literature on corporate risk management argues that firms can ben-
efit from hedging market risks, because excessive volatility increases the
expected costs of financial distress and can lead to suboptimal investment.
For example, Smith and Stulz (1985) show that risk hedging can reduce
expected tax liabilities, expected bankruptcy costs, and equilibrium wages
paid to risk-averse managers. Stulz (1990), Bessembinder (1991), and Froot,
Scharfstein, and Stein (1993) present models incorporating contracting costs
and costly external financing, and show that a policy of hedging market
risks can lead to more efficient capital investment outcomes.
Relying on the corporate hedging literature, we argue that companies in
the power industry are likely to benefit from reducing the risk of their cash
flows. The extreme volatility of wholesale power prices implies that even
well-capitalized power firms may have power price exposures sufficiently
large that adverse price changes could lead to corporate default or bank-
ruptcy. Indeed, two major California power retailers, Southern California
Edison and Pacific Gas and Electric, defaulted on scheduled payments to
creditors and suppliers during January 2001, attributing the defaults to the
high costs of purchasing power on the wholesale markets.10 Further, the
large capital investments involved in power production and distribution in-
crease the relevance of risk-related changes in investment incentives. In
practice, power producers have used bilateral forward contracts for decades,
and there is an active brokered market in power forward and option con-
tracts. The New York Mercantile Exchange and the Chicago Board of Trade
have each introduced power futures contracts, and there is an emerging
market in weather-related derivatives targeted to power producers.
We assume that power firms consider both expected profits and the vola-
tility of profits. Higher expected profits from power transactions improve
firm value, while higher volatility imposes costs due to increases in the like-
lihood of financial distress and/or effects on future investment incentives.
We assume that power firms' objectives are linear in expected profit and the
10 See also Anderson (2000) for a description of how events in June 1998 caused power prices
in the U.S. Midwest to increase from $30/MWh to over $7,000, leading to contract defaults and
the near bankruptcy of some power firms.
variance of profit. Optimal hedging strategies when the objective is linear in

expected profit and variance of profit have been studied by Rolfo (1980) and
Hirshleifer and Subramanyam (1993), among others.
C. Forward Market Participation

We model the forward market as a closed system, where only producers
and retailers (power marketing firms) can take positions. While this over-
simplifies the actual situation, it seems a reasonable starting point. Since
power cannot be stored, each power marketer must arrange to deliver the
power it purchases to an ultimate retail customer who will consume it."
Outside speculators, that is, those who neither produce power nor have de-
livery contracts with final customers who will consume the power, cannot
take positions in contracts that require physical power delivery. Futures con-
tracts allow for participation by outside speculators, so long as they offset
their positions prior to the delivery date. Although power futures contracts
are traded, activity levels are extremely low.12 Outside speculators can take
positions in cash-settled contracts, but prices of cash-settled contracts re-
main linked to prices of physical-delivery contracts only by the trades of
those participants who can accomplish physical delivery.13 In practice, trad-
ing in cash-settled electricity contracts has not developed rapidly. Krapels
(2000, pp. 13 and 72) observes that:
regional over-the-counter markets are the centers of price dynamics, with

very limited potential for participation in trading by organized or indi-
vidual speculators . . . [and] . . . electricity markets have not yet devel-
oped effective mechanisms for speculator participation.
Ong (1996) reports the market for cash-settled trades in electricity (mainly
options and swaps) to be only about five percent as large as the physical
delivery market. He attributes the slow growth in cash-settled trading to the
lack of definitive spot power price indices.14
" A power-marketing firm that contracts to purchase power must also arrange a "sink" for
the power. Conversations with industry participants indicate that some power marketing firms
have defaulted on agreements to purchase power, not for financial reasons, but because they
were unable to arrange for a sink.
12 For example, on October 2, 2000, the New York Mercantile Exchange listed futures prices
for power delivery at the Pennsylvania, New Jersey, Maryland (PJM) market for each month
from October 2000 through November 2001. However, open interest in every contract except
November 2000 was zero, and only four November contracts were open.
13 Eydeland and Geman (1999) also consider the role of physical production capacity, noting
that the safest way to hedge the risks involved in trading power options is to operate a power
plant.
14 The best publicized spot power indices are those disseminated by Dow Jones, & Company,
Inc. However, the Dow Jones data reflect the average price paid for power delivered at a loca-
tion on a given day. Since many deliveries result from previously arranged forward or option
contracts, the data does not accurately represent actual spot prices.
Though we do not formally present an alternative model with costless

participation by unlimited numbers of outside speculators, it is easy to en-
vision the possible equilibria that would result. With risk-neutral outside
speculators, the forward price would converge to the expected delivery date
spot price. If the outside speculators are risk-averse but hold diversified
portfolios, then a CAPM-style result will be obtained, with the bias in the
forward price as a predictor of the spot price dependent on the covariance
between power prices and overall market returns. If (as might be expected)
power prices are not significantly correlated with aggregate market returns,
then frictionless models with unlimited numbers of either risk-averse or risk-
neutral speculators will imply a zero risk premium for power contracts.
An alternative to assuming either the complete absence of outside specu-
lators or costless participation by an unlimited number of outside specula-
tors is to consider the case where a finite number of outsiders endogenously
choose to speculate, after bearing an entry cost. Hirshleifer (1988) presents
a model where outside speculators bear a fixed setup cost to trade in the
forward market. In the resulting equilibrium, the sign of the forward pre-
mium is determined by systematic risk and by the net hedging pressure of
producers and consumers of the underlying good. This suggests that the sign
(though not the magnitude) of the forward risk premium in a market with-
out systematic risk is not altered by the introduction of limited outside spec-
ulation, and that the model's testable implications are unchanged. Section III
below presents results of testing the implications of the model developed
here against the zero-risk-premium alternative.
II. Wholesale Power Trading, Equilibrium Hedging,

and the Forward Risk Premium
We begin by assessing real-time trading in the wholesale spot market while
taking into account previously selected forward positions. We then work back
to assess optimal strategies and market-clearing conditions in the forward
market, given that each producer and retailer will behave optimally in the
spot market.
A. The Wholesale Spot Market

In the wholesale spot market, producers sell to retailers, who in turn dis-
tribute power to their customers. Let Pw denote the wholesale spot power
price, Qw denote the quantity sold by producer i in the wholesale spot mar-
ket, and QF denote the quantity that producer i has previously agreed to
deliver (purchase if negative) in the forward market at the fixed forward
price PF. The ex post profit of producer i is given as
ITPiPWQP + PFQF - F - - (Q)C (2)

c
where each producer's physical production, Qpi, is the sum of its spot and
forwardsales, QW + QFi-
Retailers simply buy in the real-time wholesale markets the difference
between realized retail demand and their previous forward purchases. Let-
ting QF denote the quantity sold (purchased if negative) forward by retailer
j, and PR denote the fixed retail price per unit, the ex post profits of each
retailer j are given by
tRj PRQRj + PFQ -

PW(QRj + QF). (3)
The profit-maximizing quantity sold in the spot market by producer i is
QPi = QPi,(4
where x = 1/(c - 1). Equating the total physical production of producers to

total retail demand, and using the fact that forward contracts are in zero net
supply, the market-clearing wholesale price can be expressed as
QD c-1
Pw=a cl t N ' (5)
where QD = INR1 QRj denotes total system retail demand. Note that if c > 2,
as would be expected if marginal productive efficiency decreases with output
or if capacity constraints are potentially binding, then this expression can
readily account for some of the important characteristics of observed power
prices. These include disproportionately large increases in spot prices when
demand rises, and positive skewness in the probability distribution of spot
prices.
Using expressions (4) and (5), each producer's sales in the wholesale spot
market can be written as
(6)
Pi Np QPi
B. The Demand for Forward Positions

We next step back in time to determine the optimal forward positions
taken by retailers and producers and the equilibrium forward price. It is
useful to consider the interaction between profits in the absence of any hedge
positions and wholesale prices. Restating (2) and (3) and using expression
(6) for Qw, define the "but-for-hedging" profits of producers and retailers as
(QD \ a (QD\\c
PPi-Pw Np
( ) F c 7Np
and
PR, PR QRj -
PWQRj, (8)
respectively.
Define A/2 to be the coefficient on the variance of profit in the objective
functions of power marketers and producers. If A is zero, then volatility risk
is not relevant to corporate objectives, while A greater than zero implies that
volatility risk is viewed negatively. The optimal forward position when the
objective function is linear in expected profit and the variance of profit is
known (e.g., Anderson and Danthine (1980) and Hirshleifer and Subraman-
yam (1993)) to be
QF =PF-E (PW)+ COV(P{PR}i,PW) (9)

9
PRi=A Var(PW) Var(Pw)
The optimal forward position contains two components. The first term on
the right side of (9) reflects the position taken in response to the bias in the
forward price as compared to the expected spot price. The second term is the
quantity sold forward to minimize the variance of profits. In evaluating (9),
it is useful to further consider the covariance between "but-for-hedging" prof-
its and the wholesale price. Forward hedging can reduce risk precisely be-
cause this covariance is nonzero. Using expressions (7) and (8) along with
expression (5) for Pw derived above (to reflect that firms will optimize in the
spot markets), these covariance terms can be written as
1 1x1I rP+
= - Pw) - ca Cov(PN
CoV(Ppi,Pw) Cov(PW + ,Pw) (10)
and
COV (PRj, PW) = PR COV(QRj, PW) - COV(PWQRj,PW) (11)
Expressions (10) and (11) reveal the four types of risk that power forward
positions can potentially hedge. The first term on the right side of (11) re-
flects that, despite fixed retail prices, retail revenues covary positively with
wholesale prices when the quantity demanded locally is positively correlated
with wholesale price. Because the wholesale price is an increasing function
of system demand, this correlation is necessarily positive on average, imply-
ing that the industry's retail revenue is risky. Second, retailers bear risk in
their costs of acquiring power on the wholesale markets, reflected by the

second term on the right side of expression (11). Third, producers bear risk
in their sales revenue, reflected by the first term on the right side of ex-
pression (10). Finally, producers bear risk in production costs, reflected by
the second term on the right side of expression (10). Each of these risks will
affect the forward positions that individual firms take. Note though, that
producer's revenues and retailer's costs are zero-sum, implying that these
risks are diversifiable within the system. The equilibrium forward price will
depend only on the risks borne by the industry as a whole: variability in
retail revenues and in production costs.
C. The Equilibrium Forward Price

The forward price that yields a zero net supply of forward contracts is
shown in Appendix A to be
PF= E(PW) - NcX [CPRCov(P`,Pw) - Cov(Pj ,

,PW)], (12)
where N = (NR + NP)/A is a measure that reflects the number of firms in

the industry and the degree to which they are concerned with risk. The
forward price converges to the expected spot price if the number of firms in
the industry approaches infinity, or if risk is irrelevant to firms' objectives
(A =0).15 With finite N, the equilibrium forward price differs from the ex-
pected wholesale price by the difference between two risk-related terms that
reflect the net hedging pressure of producers and retailers. The forward
price will be less than the expected spot price if the first term in brackets,
which reflects retail revenue risk, is greater than the second term, which
reflects production cost risk.
To gain additional intuition about the characteristics of the forward risk
premium, it is useful to restate equation (12) in terms of the central mo-
ments of the distribution of wholesale spot prices. Appendix A shows that
when the functions PX and px?l are approximated using second-order Tay-
lor series expansions, the equilibrium forward price can be restated as
PF= EE(PW)+ a Var(PW) + ySkew(Pw), (13)
where Var(Pw) and Skew(PW) denote the variance and skewness of the whole-
sale spot price, respectively, and where
15 We assume that producers and retailers are equally concerned with risk. If retailers and
producers view risk differently, then the term N in expression (12) is restated as N = (Np/Ap +
NR/AR), where Ap and AR are the coefficients on the variance of profits in the objective func-
tions of power producers and retailers, respectively. Thus the sign of the forward premium is
unaffected by differential degrees of concern with risk, but the magnitude of the premium will
typically be altered.
Np(x + 1)
Ncax ([E(Pw)]x - PR[E(Pw)]x-1)
and
e- 2NcaX (x[E(Pw)]xl1 - (x - 1)PR[E (PW)]x-2).
To induce risk-averse retailers to enter the industry, the fixed retail price
must exceed the expected wholesale spot price, implying that a and the sec-
ond term on the right side of (13) are strictly negative. If the distribution of
spot prices is not skewed, the forward price is downward biased relative to
the expected spot price. The downward bias in the forward price in this case
reflects retailers' net hedging demand. The profits of power retailers are
positively exposed, on average, because more retail power is sold when the
wholesale price is high. This positive revenue exposure leads to a net de-
mand to sell (to create an offsetting exposure) in the forward market to
reduce risk. The downward bias in the forward price stimulates an offsetting
demand for forward purchases, so that the market can clear. Ceteris pari-
bus, the magnitude of the bias increases with the expected variance of prices.
Recall that c - 2 and x = 1/(c - 1), implying that 0 ? x ? 1. The term y
is therefore positive, implying that, ceteris paribus, the forward price in-
creases with the skewness of spot prices. The distribution of wholesale power
prices will be positively skewed if marginal production costs are convex (c > 2)
or if the demand distribution itself is positively skewed. Positive skewness
in wholesale prices reflects the possibility of large upward spikes in mar-
ginal production costs which, given fixed retail prices, reduce industry prof-
itability. The industry as a whole would prefer to hedge against production
cost spikes by purchasing power at fixed forward prices. As a consequence,
the equilibrium forward price must be increased to induce offsetting forward
sales to allow the market to clear.
To illustrate the sign and economic determinants of the equilibrium for-
ward risk premium as expressed in equation (12), we conduct a series of
simulations. In interpreting the simulation results, note that the magnitude
of the simulated premium can be made arbitrarily large or small by varying
the number of firms or the risk relevance parameter, A. The simulation is
useful in terms of clarifying the determinants of the sign of the forward bias.
Initially, we assume that power demand is distributed normally with mean
E(QD) - 100. We consider demand standard deviations (SD) ranging from
SD = 1 to SD = 40. The numbers of firms are set at NR NP = 20, and cost
functions ranging from quadratic (c = 2) to quintic (c 5) are considered.
To maintain comparability across cost structures, the variable cost param-
eter is set as a = 30(Np/100)(cl), which ensures that the wholesale price
conditional on demand of 100 is $30, regardless of c. The retail price is
set at 1.2 times the expected wholesale price conditional on c. Finally, to
keep risk premia comparable across c, the risk-relevance parameter is set as
*25-30
*20-25
_10-15
I5
.-0-5
D-15-10
-15-
(C)
~~~~~~~~~~~~2
P_D_ In Co Ce RKk
Figure 2. Bias in forward power price as a percentage of the expected spot price, as
a function of demand risk and production cost convexity. Demand is normally distrib-
uted with mean E(QD) = 100 and standard deviations as indicated on the figure. The numbers
of firms are set at NR = NP = 20. Variable costs for each producer are given by (a/c)(QD)c. The
retail price is set as PR = 1.2E(PwIc). Tb render average wholesale prices comparable across c,
we set a = 30(Np/E(QD))(c-l). To render risk premia comparable across c the risk parameter is
set as A = 0.8/2c.
A = 0.8/c2. For each value of c and SD, we randomly generate 1,000 demand
realizations. For each realization, the spot price is computed according to
expression (5), and from the 1,000 realizations, the covariance measures and
the forward price in (12) are computed.
Figure 2 displays the bias in the forward price implied by expression (12),
as a percentage of the expected spot price, for differing degrees of cost func-
tion convexity and levels of demand risk. For quadratic costs (c = 2), the bias
is always negative, with the forward price less than the expected delivery
date spot price, and increases in absolute magnitude as demand risk in-
creases. Quadratic costs imply a linear supply schedule and, given normally
distributed demand, that the skewness of wholesale prices is zero. As a con-
sequence, only the variance of prices is relevant, and the equilibrium for-
ward price decreases as demand becomes more variable.
With convex marginal costs (c > 2) and normally distributed demand, the
distribution of spot power prices becomes positively skewed, and the skew-
ness increases rapidly with demand variability. Figure 2 shows that the bias
in equilibrium forward prices is then negative for low levels of demand risk,
but reaches a minimum and becomes positive at higher levels of demand
30-~~~~~~~~~~~~~~~~~U04
U20-30
M~~~~~~~~~~~~~~~~~10-
20
-2 0 -20-1
-10 \ _ ~~~~~~~~~~~~~~~0124
Demand
Standwd Deviation 'nnd D
Figure 3. Bias in forward power price as a percentage of the expected spot price, for
varying mean demand and demand risk. Demand is normally distributed with mean and
standard deviations as indicated on the figure. The numbers of firms are set at NR = NP = 20.
Variable costs for each producer are given by (a/c)(QD)c, where a = 30(Np/E(QD))c-l with
c fixed at 4. The retail price is set as PR = 1.2E(PwIc). The risk relevance parameter is set at
A = 0.8/2c.
risk. The minimum is reached sooner and the equilibrium bias is larger if
the production cost function is more convex. This reflects that retail revenue
risk dominates when demand variance is small, but that the production cost
risks increase with skewness and dominate when demand risk is large.
Mean power demand varies substantially across times of day and across
seasons. To investigate the implications of the model with respect to varia-
tion in expected demand, we conduct simulations in which E(QD) is varied
from 75 to 125, while the standard deviation of demand is varied from 1 to
40. For brevity, and in light of the empirical evidence presented in Sec-
tion III below, the simulations reported are limited to cost convexity of c = 4.
The simulation results displayed in Figure 3 show that the model implies
that the equilibrium forward power price increases monotonically relative to
forecast spot prices when expected demand increases. The sensitivity of the
forward power price to mean demand is greater when the standard devia-
tion of demand is larger, and the largest forward bias is obtained when both
expected demand and demand variability are greatest. The intuition is that
either higher mean demand for given demand risk, or higher demand risk
for given mean demand, imply an increased likelihood of price spikes, in-
creasing the skewness of the wholesale price distribution.
D. Testable Hypotheses
The model developed here makes the following testable predictions regard-
ing the forward premium (i.e., the bias in the forward as a predictor of the
delivery-date spot) in power prices.
HYPOTHESIS 1: The equilibrium forward premium decreases in the antici-
pated variance of wholesale prices, ceteris paribus.
HYPOTHESIS 2:The equilibrium forward premium increases in the anticipated
skewness of wholesale prices, ceteris paribus.
Though our primary testable implications are stated in terms of power prices,
the underlying state variable is power demand. The simulations conducted
above demonstrate that the model also supports the following implications.
HYPOTHESIS 3: The equilibrium forward premium is convex, initially decreas-
ing and then increasing, in the variability of power demand, ceteris paribus.
HYPOTHESIS4: The equilibrium forward premium increases in expected power
demand, ceteris paribus.
The equilibrium premium in forward power prices is likely to vary in sign
and magnitude on a seasonal as well as a geographic basis. The probability
distribution of spot power prices during the temperate climates of spring
and fall is likely to be characterized by lower mean demand, relatively low
volatility, and low skewness. In contrast, the distribution of spot prices dur-
ing winter and summer is likely to involve higher mean demand, more price
variability, and greater price skewness. Further, the model predicts higher
power forward prices (relative to also higher average spot prices) in geo-
graphic regions where the power system typically operates nearer to system
capacity.
E. Optimal Forward Positions

We next present expressions for producers' and retailers' optimal forward
positions. These expressions are useful in evaluating which firms have a
comparative advantage in selling into the forward market as compared to
buying forward or abstaining from forward transactions.
E. 1. Optimal Forward Positions for Producers

In Appendix B, we show that the optimal forward sale by a power pro-
ducer is given, to a second-order approximation, by
QF E(QD) PREM (x[E(Pw)]Wx (Skew(Pw) (14)

Pi A Var (Pw) + 2ax AVar (Pw) (4
Np
where PREM is the excess of the forward price defined by (12) over the
expected spot price.
1.4- ~~ ~ ~ ~ 1.-.
1.2~~~~314.
31.3-1.4
R~od
0.8~~~~~~0.-
00.6-0.0
Figure 4.Producers optima forward osition elative 0.7~~~~~~3...

oepected produowtNidg .De
(a/)(D)C wth fxedatfou. he etil rie i st a P =1.1 OftmurH.i.12-3
Figure 4. Producer's optimal forward position relative to expected production. De-

mand is normally distributed with mean and standard deviation as indicated. The numbers
of firms are set at NR = Np= 20. Variable production costs for each producer are given by
(a/c)(QD)c, with c fixed at four. The retail price is set as PR = 1.2E(Pwlc).
A starting point for considering the power producer's optimal forward po-
sition is its expected physical production, E(QD)/Np. The optimal producer
forward sale will equal expected output only if the forward price is an un-
biased predictor of the expected spot price and the distribution of wholesale
prices is symmetric. Power producers optimally respond to a positive pre-
mium in the forward price by increasing forward sales and vice versa. Fi-
nally, the optimal producerforward sale increases with the skewness of prices,
which serves to hedge its wholesale revenue stream.
Figure 4 illustrates the implications of expression (14), with c = 4 and
normally distributed demand. The optimal forward positions displayed in
Figure 3 have been scaled by the producer's expected output, and hence can
be interpreted as its optimal hedge ratio. When expected demand and de-
mand risk are low, the producer responds to the resulting downward bias in
the forward price by reducing forward sales, and its hedge ratio is less than
one. However, as either demand risk or expected demand increases, the pro-
ducer optimally increases its forward sales, in response to both the increas-
ing equilibrium forward price and to hedge its revenue stream.
E.2. Optimal Forward Positions for Power Retailers

In this model, producers are homogeneous, reflecting the fact that each
shares the same production technology and sells into the same wholesale
market. Retailers, in contrast, can be heterogeneous if the probability dis-

tribution of demand load varies across firms. To investigate the potential
role of cross-sectional variation in local demand characteristics for optimal
retailer hedge positions, consider the ordinary least squares regression of
local demand on the average of system demand:
QD
QRip? Pi +,Pi N + Si, (15)

NR
where pi = Cov(QRi QD)/NRVar(QD). This "power demand beta" measures

the sensitivity of retailer i demand to average system-wide demand. We as-
sume that the errors f-rom this regression are independent of total system
demand, QD 16
Using (15), Appendix B shows that the optimal retailer forward position
can be expressed, to a second-order approximation, as
PREM ( Skew (PW)i

R (Q
A) AVar(Pw)
? ? f Var(Pw) 1)' (16)
where
Z = (PRx[E(Pw)]1 - x[E(Pw)]x),
(= X(X 1) [E(Pw)]X-2 - ( [E(PW)] ),

2 2 L W I]
and
(a ) NS
Note that since x is less than or equal to one, the term Y is strictly negative.
Expression (16) shows that power retailers faced with different power
demand betas will optimally take differing forward positions. However, de-
mand load characteristics may not vary across retailers in long-run equi-
librium. In the absence of contracting costs, risk-averse retailers have
incentives to enter contracts where each agrees to service a portion of the
demand load in every geographic region, thereby diversifying demand risk
across retailers, and pushing each retailer's power demand beta towards
16
The regression errors are, by construction, uncorrelated with system demand. Our stron-
ger assumption assures that the regression errors are also uncorrelated with the wholesale
price, which is a nonlinear function of demand. In the absence of this assumption expression
(16) contains two nuisance terms involving correlations between the regression errors and func-
tions of the wholesale price.
one. The optimal forward positions taken by such fully diversified retailers
can be assessed by setting the power beta coefficient in (16) equal to one.
We do not present simulations of the resulting forward positions in this
case for two reasons. First, the position that a power retailer with a power
demand beta of 1.0 would take to reduce its risk is simply the opposite of
the position taken by a representative power producer (Figure 4). Second,
contracting costs and remnant regulation of access to local markets, though
not formally included in our model, are likely to prevent actual power
retailers from immediately sharing demand risk. It will likely be years
before most retail markets are shared among power marketing companies
that operate on a system-wide basis. The analysis of optimal hedging for
retailers with heterogeneous demand is likely to be relevant to actual re-
tailers in the interim.
The first term on the right side of (16) is the forecast demand load. How-
ever, the retailer optimally purchases forward a quantity that matches its
expected demand only when the premium in the forward price is zero and
local demand is uncorrelated with system demand. The second term on the
right side of (16) reflects the response to the bias in the forward price. The
third term on the right side of (16) reflects systematic demand risk. If de-
mand risk is not systematic (,3 = 0), the third term is zero. Within the third
term, Z is positive when the retail price exceeds the expected spot price. A
positive value of Z reflects retail revenue risk, which the retailer hedges by
increasing its quantity sold forward (i.e., decreasing the quantity purchased
forward), if the demand risk is systematic. In contrast, any positive skew-
ness in the price distribution increases the importance of risk in the costs of
purchasing power, and leads to a reduction in the optimal quantity sold
forward (i.e., an increase in forward purchases), if the risk is systematic.
Figures 5 and 6 present the results of simulations, again assuming nor-
mally distributed demand and c = 4, that illustrate the implications of ex-
pression (16) for the optimal forward position of power retailers. Figure 5
presents optimal hedge ratios (forward positions relative to expected de-
mand) for a retailer whose demand risk is nonsystematic, that is, its power
beta is zero. Figure 6 presents optimal hedge ratios for a retailer with a high
degree of systematic power demand risk, as indicated by a power beta of 2.0.
The retailer with nonsystematic demand risk chooses a hedge ratio that
differs from (negative) one only as a response to the bias in the forward
price. As a consequence, its purchases forward an amount greater (less than)
its expected sales if the forward price is downward (upward) biased.
The retailer with systematic demand risk optimally deviates further from
a hedge ratio of (negative) one. When demand risk is small and expected
demand is low, this retailer is more concerned with revenue risk, reducing
its forward purchase, and in extreme cases may even sell forward. As ex-
pected demand or demand risk increases, the retailer with systematic de-
mand risk rapidly increases its forward purchases, reflecting an increased
desire to hedge purchase costs against the possibility of wholesale price
spikes.
/ _ . *.o.-o.8
o dgRoRuda
S for _btWW .-1-0.e
_Nk -1.1- < ^-1.1-1
Dumumd
High Oi-1.2--1.1
-1_3* -1.3X-1.2
-1A~~~~~~~~~~~~~~~~~~~~~~~~~~ 141.
. _ ~~~~~~~~~~~~~~~103
_
Figure 5. Optimal forward sale relative to expected demand for a retailer whose de-
mand risk is nonsystematic. Demand is normally distributed with mean and standard de-
viation as indicated. The numbers of firms are set at NR = NP = 20. Variable production costs
for each producer are given by (a/c)(QD)c, with c fixed at 4. The retail price is set as P=
1.2E(PwIc).
III. Preliminary Empirical Evidence on the Pricing

of Power Forward Contracts
A. Issues in Testing Forward Pricing Hypotheses in Power Markets
Empirical testing of the hypotheses developed here is constrained by the
fact that the markets are new and data is scarce. The newness and unique-
ness of the wholesale power markets raise the possibility that observed for-
ward prices will reflect the inexperience of industry participants, and may
differ from the pricing structure that will be observed in a longer-run equi-
librium. A learning phenomenon of this type was documented by Figlewski
(1984), who reported that market prices for stock index futures initially de-
viated significantly from theoretical values, but converged toward predicted
values after a few months of trading. The pricing structure is also likely to
change if financial contracts that facilitate the sharing of power price risk
with outside investors are developed.
There is substantial empirical literature that tests forward and futures
pricing theories in stock, bond, foreign exchange, and commodity markets.
The research designs employed typically seek to identify the ex ante pre-
mium in the forward price by measuring the ex post differential between
RedoforDuidewiwh M-.
Hedgrie
i -i\ Sysmeu
HodzpRlbk
.k -0.5-0
Dwm rwr~~~~~~~~~~~~Uith
I\-1-0.5
101
Dommnd_Ied \} -
Figure 6. Optimal forward sale relative to expected demand for a retailer whose de-
mand risk is systematic, with power demand beta of two. Demand is normally distrib-
uted with mean and standard deviation as indicated. The numbers of firms are set at NR =
NP = 20. Variable production costs for each producer are given by (a/c)(QD)c, with c fixed at 4.
The retail price is set as PR= 1.2E(PwIc).
forward prices and realized delivery date spot prices, which equals the ex
ante premium plus random noise. A common difficulty, however, is that ran-
dom shocks to asset prices are large compared to any premium in the for-
ward price, so that tests conducted in smaller samples lack statistical power.
For example, Fama and French (1987) conduct tests of whether futures risk
premiums are nonzero using between 9 and 18 years of data from 22 commodity
markets. They conclude (p. 73) that "the large variances of r.ealizedpremiums
mean that average premiums that seem economically large are usually insuf-
ficient to infer that expected premiums are nonzero"and "the evidence is not
strong enough to resolve the long-standing controversy about the existence of
nonzero expected premiums." Difficulties related to a lack of statistical power
are compounded in the case of electricity markets because of the short time
series of available data and due to the unusually large variability of realized
prices. As a consequence, the results reported here should be viewed as pre-
liminary; to obtain definitive statistical results regarding the pricing of power
forward contracts may well require numerous years of additional data.
B. Research Design and Data Description

In addition to the traditional approach of comparing forward prices to
subsequently realized spot prices, we adopt an alternative empirical ap-
proach with potential to improve statistical power in the small sample. We
exploit the fact that spot electricity prices are determined by demand for
consumption and supply from production at each point in time.17 Specifi-
cally, we use realized daily demand (load) and price data to estimate param-
eters of the power production function, and then obtain estimates of expected
spot prices by calendar month using the fitted cost curve and realized load
data. Forward premia can then be estimated simply as the difference be-
tween observed forward prices and cost-based estimates of expected delivery
date spot prices.
We examine power pricing for the Pennsylvania, New Jersey, Maryland
(PJM), and the California Power Exchange (CALPX) markets. We focus on
these two markets because they are the first markets to implement central-
ized dispatch and real-time market pricing of power, and because daily load,
spot, and forward price data are available for these markets.18 Each market
is governed by an independent system operator (ISO), who assigns genera-
tion to meet power demand. The ISO uses supply schedules submitted by
producers to dispatch generation assets to minimize the cost of meeting power
demand while maintaining system reliability. The available PJM data cover
the period April 1997 through July 2000. For CALPX, spot price data is
available from April 1998 through July 2000.
Figures 7 and 8 display monthly average spot prices and power demand
(load) for PJM and CALPX, respectively. The PJM price spike during July
1999 is apparent, as is a notable spike in CALPX power prices during sum-
mer 2000. Some distinctions in the two price spikes are also apparent, how-
ever. First, the PJM price spike quickly reverted and affected prices during
only a single month, while the CALPX price spike persisted until the end of
our sample period. Second, the PJM price spike was accompanied by the
highest observed demand, and is potentially attributable to convexity in the
production cost function. The CALPX price spike occurred even though de-
mand levels were actually less than in corresponding months of the previous
two summers, and must therefore be attributed to either a leftward shift of
the competitive supply function, or to the exercise of market power.19 We do
not attempt to distinguish between these explanations, but observe that nei-
17 Prices for storable commodities, in contrast, are determined not only by supply from cur-
rent production and demand for current consumption, but also by supply from or demand for
inventory.
15 Spot price and load data are obtained from the web sites http://www.caiso.com and http://
www.pjm.com, which also contain additional information about the markets. Standardized fu-
tures contracts for delivery at several locations, including Palo Verde (Arizona), California-
Oregon Border, Cinergy (U.S. Midwest), Entergy (South-central U.S.) and PJM, are traded on
the New York Mercantile Exchange and the Chicago Board of Trade. However, trading activity
in the futures has been thin, and load data is not readily available for these markets.
19 The U.S. Department of Energy (on the web page http://www.eia.doe.gov/cneaf/electricity/
california/subsequentevents.html) has attributed the "high level of planned and unplanned (power
plant) outages" in California to "the extended use of power plants during the previous excep-
tionally hot summer months" and that producers "had used their allotted emission allowances."
Joskow and Kahn (2001, p. 30) in contrast attribute reduction in output to "the withholding of
supplies from the market."
180 45000
160 40000
140---7_ 35000
120 - - 30000
100 - - - _ 25000
*C 80 ___ _ _ 20000 .
60 _ __ --_ __ ____ _ 15000
40 / - - - - 10000
20 - _ _ _ _ _ _ - - _ _ _ _ _ - 5000
0- 0
Dec-96 Jun-97 Jan-98 Jul-98 Feb-99 Aug-99 Mar-00 Oct-00
Date
Figure 7. Monthly average electricity demand (load) and realized wholesale spot prices
in PJM. The data cover April 1997 to July 2000.
ther is allowed for in our model. Although we report evidence for both mar-
kets, these considerations suggest that the results from PJM potentially
provide more information as to the validity of our model of forward pricing
in competitive electricity markets.
We begin by estimating the parameters of the power production function,
using daily price and load data. Our model implies that spot prices are given
by equation (5). Taking logarithms, parameters of the production cost func-
tion can be estimated by the following linear regression:
ln(Pt) = a + (c - 1)ln(Qt + dt) + et, (17)
where Pt is the daily average on-peak spot price, Qt is daily average load,
and a and c are parameters to be estimated. We use indicator variables
denoted dt to designate each calendar month to account for the fact that the
power supply function varies seasonally, as producers schedule planned main-
tenance outages during periods of low expected power demand. We also em-
pirically accommodate the CALPX summer 2000 data by use of additional
indicator variables for June and July of 2000. In this specification, c esti-
mates the degree of convexity in the cost function. A value of c greater than
two implies that the spot prices will be positively skewed, even if the power
demand distribution is symmetric.
250 .........-----.- ..........

. .
........................ ..... 35000~
Load 7~~~~~~~~~~~~->
~~~~30000
/ / --., / ~~~~~~~~~~~~~~
150 _ _ _
20000
Spot Price
5000
0 ~~~~~~~~~~~~
Jan-98 Apr-98 Jul-98 Nov-98 Feb-99 May-99 Aug-99 Dec-99 Mar-00 Jun-00 Oct-00
Figure 8. Monthly average electricity demand (load) and realized wholesale spot prices
A-pyplied to daily PJM data from April 1997 through July 2000, the esti-
mated value for c is 4.80 (t-statistic =37.9), indicating a significant degree
of convexity in the cost function. The regression R-squared is approximately
60 percent, indicating that variation in load and the simple cost function in
expression (1) capture a substantial portion of the variation in observed spot
prices. For CALPX over the period April 1998 to J-uly 2000, the estimated
value for c is 5.81 (t-statistic 23.3), and the model explains almost 70 per-
cent of the variation in spot power prices.
The indicator variables used to account for the structural break in prices
in the CALPX market that occurred during summer 2000 are positive and
statistically significant. One potential reason for the change in the pricing
function during this period is rising prices for natural gas, which is often the
marginal input fuel in the California market. To explore this possibility, we
obtain data on the nearby natural gas futures price for delivery at the South-
ern California border.20 There is an upward trend in gas prices over the
sample period, with the mean price increasing from $2.30/MMBtu in August
1998 to $3.88/MMBtu in July 2000. We again estimate (17), but use the
20 We thank the referee for suggesting that we control for fuel prices and Alexander Eyde-
land of Mirant Corporation for the natural gas pricing data. The natural gas futures prices are
only available starting in August 1998. We It-herefore lose the first four months of our sample
when we control for fuel prices.
natural log of the ratio of the electricity price to that of the nearby gas
futures price as the dependent variable. The resulting estimate of c is 5.30
(t-statistic = 17.23), and the model explains about 60 percent of the varia-
tion in spot power prices. More importantly, the summer 2000 indicator vari-
ables remain significantly positive, which indicates that rising fuel prices do
not fully explain the structural shift in CALPX prices during summer 2000.
We report our remaining results for CALPX using the pricing function that
does not control for fuel prices. This pricing function fits the data slightly
better and allows us to use a longer sample period. The results obtained
after controlling for variation in gas prices are qualitatively similar, and can
be obtained from the authors on request.
Having estimated the cost function based on daily realizations of price and
load, we next estimate expected spot prices by month, using a bootstrap ap-
proach to generate average fitted values from (17). We generate a series of
daily load outcomes for each calendar month by randomly drawing (with
replacement) 2,000 observations from the sample of actual load outcomes
observed in that calendar month, during any year of the sample. This pro-
cedure yields a bootstrapped empirical distribution of power demand, under
the assumption that the distribution of power demand varies across calen-
dar months, but is stationary across years. From each daily load observa-
tion, we compute the fitted spot price from (17), and then estimate the expected
spot price for each calendar month as the mean of the 2,000 fitted values.21
The fitted prices for CALPX do not, however, include the effect of the sum-
mer 2000 indicator variable, which occurred for reasons not modeled and
that presumably surprised industry participants.
Table I displays both the realized average on-peak spot power price by
month, as well as the resulting cost-based estimates of expected spot prices
by calendar month in both markets. Both the realized and expected spot
prices exhibit clear seasonalities, with average spot prices highest in the
summer months. Note that average realized spot prices are higher than our
cost-based estimates of expected spot prices during the summer months, in-
dicative of the significant spikes in prices that have occurred in both mar-
kets that are not accounted for by our estimated production function.
The model implies that the volatility and skewness in spot prices will be
largest in the summer months when expected demand is high, and smallest
in the spring and fall. Table II reports expected spot prices, as well as the
standard deviation and the standardized (by dividing by the standard devi-
ation cubed) skewness in unexpected spot prices. The standard deviation
and standardized skewness of unexpected spot prices are estimated on a
monthly basis, using differences between actual daily prices and the cost-
based estimates of expected prices for the calendar month. Results reported
21
We also repeated the estimations of expected spot prices when drawing load observations
from hourly, rather than daily, data. The resulting estimated spot prices were similar to those
reported. Additionally, the inferences regarding the bias in forward prices were unchanged
under this methodology.
Table I
Monthly Average On-Peak Spot Power Prices
in the PJM and CALPX Power Markets
Actual spot prices are equally weighted means of daily prices in dollars per megawatt hour
observed during the indicated calendar months. Fitted spot prices by month are average fitted
values of the estimated cost function (17), obtained based on 2,000 random draws from the
calendar months from the sample distribution of loads. The PJM data covers April 1997 through
July 2000. The CALPX data covers April 1998 through July 2000.
Market Location
PJM CALPX
Actual Spot Fitted Spot Actual Spot Fitted Spot
Month Prices Prices Prices Prices
January 25.91 25.49 30.08 27.37

February 22.54 20.94 27.71 22.89
March 24.08 23.77 26.97 21.12
April 25.46 25.30 29.39 27.59
May 30.89 26.81 37.96 23.12
June 38.23 33.03 79.36 28.07
July 71.40 45.80 73.05 36.82
August 36.66 30.93 47.37 45.53
September 27.89 28.44 41.45 38.09
October 27.08 26.13 42.43 40.98
November 23.57 23.18 37.40 37.35
December 20.72 21.78 32.97 32.19
are averages across the months in each season, with winter defined as the
months December, January, and February, and the other seasons defined
similarly using consecutive three-month periods.
Expected PJM spot prices are highest in the summer, while CALPX prices
are actually slightly higher during the fall. For PJM, expected prices are
similar across the other three seasons, while in CALPX, fall prices are some-
what higher than winter or spring prices. The standard deviation of spot
prices peaks during the summer in each market. As would be expected given
convexity in the cost function, the standardized skewness coefficient is pos-
itive and greatest during the spring and summer months.
Since the standardized skewness coefficient is the third moment of spot
power prices normalized by the cube of the standard deviation in spot power
prices, the seasonal increase in standardized skewness implies that unstan-
dardized price skewness increases during the summer more rapidly than
price variance. The model therefore implies an upward bias in power for-
ward prices for summer delivery. The model predicts that the bias in the
forward price should be small or negative when price variability is moderate
and price skewness is low. In light of the estimates reported in Table II, we
anticipate small or negative bias in power forward prices for spring and fall
delivery. We test whether the data is consistent with these broad predictions
by examining seasonal variation in the bias of forward prices.
Table II
Seasonal Average Estimates of Electricity Spot Price Parameters
Expected spot prices in dollars per megawatt hour are computed on a monthly basis as average
fitted values from text equation (17). Standard deviations and standardized skewness are com-
puted on a monthly basis using deviations of daily prices from the expected monthly prices.
Results reported are averaged across the three calendar months per season and across years.
Winter is defined as months December, January, and February. The other seasons are defined
using consecutive three-month periods. The PJM data covers April 1997 through July 2000. The
CALPX data covers April 1998 through July 2000.
PJM CALPX
Standard Coefficient Standard Coefficient

Expected Deviation of of Skewness Expected Deviation of of Skewness
Season Spot Price Spot Price of Spot Price Spot Price Spot Price of Spot Price
Winter 23.02 7.10 0.18 27.36 7.55 0.65

Spring 26.03 10.87 0.72 24.72 13.84 1.19
Summer 37.63 51.04 1.87 35.22 59.09 1.30
Fall 25.88 7.85 0.23 39.01 16.95 0.94
C. Evidence on Forward Power Pricing

Daily observations on one-month-ahead power forward contracts are ob-
tained from Bloomberg and from the CALPX Web site. Each forward con-
tract calls for the continuous delivery of power throughout the delivery month,
rather than delivery at any particular time during the delivery month.22
Bloomberg reports forward prices for PJM delivery from April 1997, but
does not report forward prices for CALPX delivery. CALPX began trading
block forward contracts in August 1999. We use CALPX block forward data
from that date on, but augment it with Bloomberg data on forward prices for
Palo Verde, Arizona, delivery from April 1998 through July 1999. Power
delivered to Palo Verde is often routed into the California market. Using the
Palo Verde forward prices undoubtedly introduces additional measurement
error into the analysis, but allows examination of a longer sample period.
Figures 9 and 10 display on a calendar month basis the average bias in
the one-month-forward price for the PJM and CALPX markets, respectively.
Results are reported both when the forward price is compared to the cost-
based estimate of the expected delivery date price and when the forward
price is compared to the average realized price during the delivery month. In
each market, the bias in the forward power price is relatively small for de-
livery months outside of summer. Relatively large biases occur during sum-
mer months. In the PJM market, the summer bias in the forward price is
positive regardless of benchmark, but is greater when forward prices are
22
More specifically, the contract calls for delivery during each weekday, on-peak, hour of the
delivery month.
30
Bias relative to fitted spot prices
25 /_
20 _ _ _ _ _ _ - - - --
15- -
10~~~~~~~~~~~~~~~~
0 '_
Jan Feua p y Jn Jot Aug Sep Nov Dec
- /' - ~~~~~~~~~~Bias
retative to reatized spotprices__
-10
Calendar Month
Figure 9. The bias in one-month-forward power prices as predictors of subsequently

realized spot prices, and relative to cost-based estimates of expected spot prices. Price
data are for power delivery at PJM from April 1997 to July 2000, and have been averaged
across all days in the indicated calendar month.
compared to cost-based estimates of expected spot prices than when com-

pared to average realized spot prices. For CALPX, the sign of the bias de-
pends on the benchmark. Forward prices for July delivery exceed cost-based
estimates of expected spot prices by as much as $40/MWh, or over 100 per-
cent. However, forward prices for June delivery are less than average real-
ized spot prices by a similar magnitude. This difference in inference is
attributable to the impact of the large and sustained price spike in summer
2000 CALPX prices. It will likely require many more summers before we can
assess whether these data points reflect price skewness that should have
been anticipated and priced in the forward market, or represent an influ-
ential but nonforecastable event.23
23
Although not reported, there is also some evidence of a learning effect in the PJM market.
In the summer of 1997, the one-month forward price was below the expected spot price. The
first highly publicized spikes in wholesale power prices occurred during June of 1998. Sub-
sequently, one-month forward power prices for summer delivery lay considerably above ex-
pected spot prices, while forward prices for non-summer delivery were little changed relative to
1997. This is consistent with the reasoning that market participants learned of the degree to
which summer power prices are positively skewed, and that the consequences of price skewness
at times of high mean demand then became embedded in forward prices.
Bias relative to fitted spot peices
40
20
m Jan Feb Mae r p May Jun / Jul Aug Sep \ Oc No tec
-20
\ Bias relative to realized spot prices
-40
-60
Calendar Month
Figure 10. The bias in one-month-forward power prices as predictors of subsequently

realized spot prices, and relative to cost-based estimates of expected spot prices. Spot
price data are for power delivery at CALPX from April 1998 to July 2000, and have been
averaged across all days in the indicated calendar month. Forward data from April 1998 to
June 1999 are for Palo Verde, Arizona, delivery, while forward data from July 1999 to July 2000
are for CALPX delivery.
Table III reports the average bias in the forward price in each market by
season, where the bias is computed as the difference between forward prices
and both expected spot prices and realized spot prices. Each seasonal mean
is computed from monthly observations. Also reported are simple t-statistics
for the hypothesis that each seasonal mean is zero. In assessing statistical
significance, it should be kept in mind that each seasonal mean is estimated
from between 6 and 11 monthly observations.
When the forward price is compared to cost-based estimates of the for-
ward price, a positive bias in summer prices is observed. The bias exceeds
$20/MWh in each market, and is statistically significant (t-statistics of 2.25
and 1.84 for PJM and CALPX, respectively). A significant positive bias is
also observed for winter PJM forward prices while a significant negative
bias is observed for fall CALPX prices. The picture is somewhat different
when forward prices are compared to average realized spot prices. The bias
in PJM summer prices remains positive but is reduced in magnitude by
more than half and is no longer statistically significant. The bias in CALPX
Table III
Average Bias in Forward Power Prices, by Season
Reported are mean differences between one-month-forward prices relative to both expected and
realized delivery month spot prices. Expected spot prices are obtained as fitted values from text
equation (17). In PJM, forward prices are the average forward prices for the month prior to
delivery, as reported by Bloomberg. For the period July 1999 through July 2000, CALPX for-
ward prices are average closing prices reported for one-month-block-forward contracts trans-
acted through the CALPX, as reported by Bloomberg. In lieu of earlier CALPX forward data,
prices for the period April 1998 through June 1999 are for delivery at Palo Verde, Arizona, as
reported by Bloomberg. Units are dollars per megawatt hour. Winter is defined as months
December, January, and February, and subsequent seasons are defined using consecutive three-
month periods. For PJM, the data cover the period April 1997 through July 2000. For CALPX,
the data cover the period from April 1998 through July 2000. t-statistics in parentheses.
PJM CALPX
Mean Bias Mean Bias Mean Bias Mean Bias

Relative to Relative to Relative to Relative to
Expected Realized Expected Realized
Season Spot Price Spot Price Spot Price Spot Price
Winter 3.09 2.78 -1.48 -1.90

(6.24) (1.86) (-1.26) (-1.62)
Spring -1.10 -2.73 -1.69 -4.90
(-0.98) (-2.13) (-0.73) (-1.85)
Summer 21.38 8.62 20.07 -12.84
(2.25) (0.62) (1.84) (-1.02)
Fall 0.24 -0.02 -5.44 -6.82
(0.18) (0.01) (-2.01) (-2.87)
summer prices becomes negative and statistically insignificant. There re-

mains evidence of a positive bias in winter PJM forward prices, and there is
evidence of a negative bias in forward prices for spring delivery at PJM as
well as for spring and fall delivery at CALPX.
The evidence of a positive premium in summer forward power prices is
consistent with the implications of the model developed here, and with the
results reported by Pirrong and Jermakyan (1999), who also find that PJM
forward prices for summer delivery contain a substantial risk premium. A
negative bias in power forwards for delivery during the temperate months of
spring and fall is also consistent with the model's predictions.
Finally, we provide some more specific evidence on the relationship be-
tween the bias in forward prices and variations in the demand for power
across seasons. In each market, we estimate the following regression:
PREMit = ao + a,MEANit + cv2STDit + cv3VARit +? it, (18)

Table IV
Regression of Monthly Bias in Forward Power Price
on Mean Demand, Standard Deviation of Demand,
and Variance of Demand, by Month
Reported are coefficient estimates obtained when estimating the following regression:
PREMgt - ao + aiMEANit + a2STDit + a3VARit + Tit,
where PREMit is the difference between the one-month-forward price for delivery in month t
and the cost-based estimate of the expected spot price in month t for market i, MEANit is mean
daily load for month t in market i, STDit is the standard deviation of daily load during month
t in market i, and VARit is the square of STDit. Units for PREM are dollars per megawatt hour.
For the PJM market, the data cover the period from April 1997 through July 2000. For the
CALPX market, the data cover the period from April 1998 through July 2000.
a,1 at2 a{3
PJM
Coefficient 49.03 39.31 -188.68
t-statistic (3.46) (0.37) (-0.96)
CALPX
Coefficient 37.81 145.87 -582.41
t-statistic (2.08) (0.51) (-0.59)
where PREMit is the difference between the one-month-forward price for

delivery in month t and the cost-based estimate of expected spot price in
month t for market i, MEANit is the average load for month t in market i,
STDit is the standard deviation of daily market i load during month t, and
VARit is the square of STDit. The theory predicts that the forward premium
should increase with mean demand (a1 > 0). It also predicts that, ceteris
paribus, the forward premium should be convex, initially decreasing (a2 < 0)
and then increasing (a3 > 0) in demand risk.
The results of estimating (18) are reported in Table IV. In the PJM mar-
ket, the coefficient estimate on mean load is positive and statistically sig-
nificant (t-statistic = 3.46). For CALPX, the coefficient estimate on mean
load is also positive and statistically. significant (t-statistic = 2.08). These
results are consistent with the model implication that the premium in for-
ward power prices for summer delivery is associated with high mean power
demand. Given the convexity of the cost function, high mean demand, in
turn, gives rise to positive skewness in spot power prices. The coefficient
estimates on the standard deviation and variance of load are not statisti-
cally significant. In view of the small sample size and the presence of sub-
stantial multicollinearity among the predictive variables, however, we view
these results as providing a degree of support for the model presented here.
Definitive evidence will not be available until considerable time passes and
the sample size is increased.
IV. Conclusions
Forward contracts for electricity cannot be priced by typical "cost-of-carry"

relationships, because power is not storable. We take an equilibrium approach,
and obtain the implication that the forward power price will generally be a
biased forecast of the future spot price, with the forward premium decreased
by the anticipated variance of wholesale spot prices and increased by the
anticipated skewness of wholesale spot prices. Simulations confirm that the
model implies that forward prices will exceed expected spot prices when
either expected demand or demand volatility are high, due to the positive
skewness induced in the spot power price distribution.
Though we focus on the pricing of forward contracts, there is anecdotal
evidence that positive price skewness (or price spikes) affect the pricing of
electricity options as well. For example, Krapels (2000, p. 24) asserts that
[i]t is common knowledge, however, that traders in many OTC electric-

ity options markets have become so fearful of being physically "net short"
(having agreed to deliver electricity in the future at an earlier agreed-
upon price) when one of the price spikes occurs that they place ex-
tremely high volatility assumptions into the pricing of OTC electricity
call options.
We also analyze producers' and retailers' optimal forward positions to gain

insights into optimal risk bearing, showing that optimal positions depend on
statistical properties of power demand and spot prices, including forecast
demand and measures of "power demand betas" and price skewness. Finally,
we provide some preliminary empirical evidence on wholesale power prices.
Consistent with the implications of the model, we document a positive bias
in forward power prices for summertime delivery, while the bias in forward
prices for spring and fall delivery is zero or negative. Regression analysis
also indicates that the forward premium increases with mean demand, as
predicted.
The summertime forward premia documented here for the power markets,
though estimated imprecisely, appear to be large relative to those that have
been documented for other commodities (e.g., Fama and French (1987)). This
suggests that the power markets are not well-integrated with the broader
financial markets, that is, that outside speculators are not a significant
presence in these markets. A lack of integration may be due to informational
setup costs (as discussed by Hirshleifer (1988)) associated with learning about
power markets, or due to the lack of good benchmark price indices on which
to base cash-settled derivative contracts, as suggested by Ong (1996). It will
be of interest to see if mechanisms are developed to facilitate the sharing of
power price risk with outside speculators, and if risk premia decline as a
consequence.
Appendix
A. Forward Pricing
Substituting the expressions for the covariance between the "but-for-
hedging" profits and wholesale spot price from text equations (10) and (11)
into the expressions for the optimal forward positions taken by retailers and
producers from text equation (9), the optimal quantities sold forward can be
rewritten as
F~_PF- E(Pw) 1 ov (P~' 1,PW)

AVar(Pw) ax c[ Var(Pw) (Al)
for producers and
F"P- E(Pw) COV(QRj,,PW)_ COV(PWQRj,,PW) (A2)

A Var(Pw) Var(Pw) Var(Pw)
for retailers. The market clearing forward price can be determined by equat-
ing the sum of the forward positions across retailers and producers to zero
NR
PF-E (Pw) Cov(QD,Pw)_ Cov(PwQD,Pw)
Rj A Var (Pw) Var(Pw)
=1 P1 Var(Pw)
PF- E(PW) Np F Cov(Pj+ 1,Pw) (A)
Np +? ax [--I arPw 0,
AVar(Pw) a c Var(PW)
where we have used the properties of covariances and the fact that
,N]1QRJ QD in writing the expression. Solving (A3) for PF, using the result
that Pw = a(QD/Np)C l, and letting N = (NR + NP)/A, the market clearing
forward price can be written as
PF= E(Pw) - NcpX [CPRCOV(PW, PW) - COV (P`W1, PW)], (A4)
which is text equation (12). Also, define
PREM = PF- E(PW) - Cov(PW ,Pw)], (A5)

Ncax [CPRCov(P,Pw)
- -
to denote the premium in the forward price as compared to the expected

delivery date spot price.
The function PZ (where z is a constant) can be approximated, by using a

Taylor's series expansion around the point E(P), as
z - -z zz(z -1))
PZ-:[E(P)]z (1 2
Z(z-1) (A6)
+ Z(2 - z)[E(P)]Z1P + 2 [E(P)]z-2P2
Substituting expression (A6), letting z x and z = x + 1 in turn, into (A4),

and using the properties of covariances, we obtain
(x + 1)([E(Pw)]x
Npax - PR[E(PW)]X-l)Var(Pw)
? NP A{x + 1) (A7)
VNcaxC 2
X (x[E(Pw)]X-1 - (x - 1)PR[E (PW)]x-2 )Skew(Pw),
which is equivalent to text equation (13).
B. Optimal Forward Positions

Using expression (A6) with z = x + 1 in combination with text equations
(9) and (10) and the definition (A5), we obtain
PREM + x[E(Pw)]x 1 (Skew(Pw)

QF -[E(Pw)]x + (A8)
Pi- ax 2ax
A Var (Pw) Var(Pw)I
for the optimal quantity sold forward by a power produer. Using a second-
order Taylor series expansion of text equation (6) at E(QD) and the proper-
ties of expectations, we obtain the result that
[E(Pw)]x ax E (QD) (A9)
Using (A9) with (AS) gives text (14).

Combining text equations (6), (9), (11), and (15), and using the properties
of covariances along with the assumption that the regression errors in (15)
are independent of system demand, the optimal quantity sold forward by a
power retailer can be expressed as
+
(
QF +PREM j Np
Rj Pi
A Var(Pw) NRaX Var(Pw)) (PRCV(PW,PW) -
CoV(PI+ PW))
(AIO)
where pi is the intercept from the regression equation (15). Using expression
(A6) with (AlO) this can be approximated as
~~RJ PREM ( Skew(PW

QRJ=
-V +A Va (Pw) +j3,(W+Y Var(Pw)])(A1
where
W= (PRx[E(Pw)]x-l - (x + l)[E(Pw)]x),
Y=(P X(X 1) [E(Pw)]x-2 _X(x

+ 1)
[E(Pw)]X-1
2 2WI1
and
Finally, taking the expectation of text equation (15), and using the result
along with (A9) in (All), we obtain
QF*
~~PREM ? Skew (PW)
R + Y Var(Pw) ])
EQ)+A Var(Pw) (A12)
where Z = (PRx[E(Pw)]X-l - x[E(Pw)]x), which is text equation (16).
REFERENCES
Amin, Kaushik, Victor Ng, and Craig Pirrong, 1995, Valuing energy derivatives, Managing
Energy Price Risk (Risk Publications, London).
Anderson, Ronald, and Jean-Pierre Danthine, 1980, Hedging and joint production:Theory and
illustrations, Journal of Finance 35, 487-501.
Anderson, William, 2000, Getting power out of risk, Energy Power Risk Management, March,
10-11.
Bessembinder, Hendrik, 1991, Forward contracts and firm value: Investment incentive and
contracting effects, Journal of Financial and Quantitative Analysis 26, 519-532.
Borenstein, Severin, and James Bushnell, 1999, An empirical analysis of the potential for mar-
ket power in California'selectricity industry, Journal of Industrial Economics 47, 285-323.
Eydeland, Alexander, and Helyette Geman, 1999, Pricing power derivatives Risk, 71-74.
Fama, Eugene, and Kenneth French, 1987, Commodity futures prices: Some evidence on fore-
cast power, premiums, and the theory of storage, Journal of Business 60, 55-73.
Figlewski, Stephen, 1984, Hedging performance and basis risk in stock index futures, Journal
of Finance 39, 657-669.
Froot, Kenneth, 1998, The evolving market for catastrophic event risk, Workingpaper, Harvard
Business School.
Froot, Kenneth, 2001, The market for catastrophe risk: A clinical examination, Journal of Fi-
nancial Economics 60, 529-571.
Froot, Kenneth, David Scharfstein, and Jeremy Stein, 1993, Risk management: Coordinating
corporate investment and financing policies, Journal of Finance 48, 1629-1658.
Gabillion, Jacques, 1995, Analyzing the forward curve, in Managing Energy Price Risk (Risk
Publications, London).
Green, Richard, and David Newbery, 1992, Competition in the British electricity spot market,
Journal of Political Economy 100, 929-953.
Harvey, Scott, William Hogan, and Susan Pope, 1996, Transmission capacity reservations im-
plemented through a spot market with transmission congestion contracts, The Electricity
Journal 42-55.
Hirshleifer, David, 1988, Residual risk, trading costs, and commodity futures risk premia, Re-
view of Financial Studies 1, 173-193.
Hirshleifer, David, and Avanidhar Subramanyam, 1993, Futures versus share contracting as
means of diversifying output risk, The Economic Journal 103, 620-638.
Joskow, Paul, 1997, Restructuring, competition, and regulatory reform in the U.S. electricity
sector, Journal of Economic Perspectives 11, 119-138.
Joskow, Paul, and Edward Kahn, 2001, A quantitative analysis of pricing behavior in Califor-
nia's wholesale electricity market during summer 2000, Working paper, Department of
Economics, Massachusetts Institute of Technology.
Joskow, Paul, and Jean Tirole, 1999, Transmission rights and market power on electric power
networks II: Physical rights, Working paper, Department of Economics, Massachusetts In-
stitute of Technology.
Kawaller, Ira, Paul Koch, and Timothy Koch, 1987, The temporal price relationship between
S&P 500 futures and the S&P 500 index, Journal of Finance 42, 1309-1330.
Krapels, Edward, 2000, Electricity Trading and Hedging (Risk Publications, London).
MacKinlay, Craig, and Krishna Ramaswamy, 1988, Index-futures arbitrage and the behavior of
stock index futures prices, Review of Financial Studies 1, 159-172.
Newbery, David, 1995, Power markets and market power, The Energy Journal 16, 39-66.
Ong, Jerome, 1996, New paradigms for electric power, Derivatives Strategy, December/January.
Pirrong, Craig, and Martin Jermakyan, 1999, Valuing power and weather derivatives on a
mesh using finite difference methods, in Energy Modelling and the Management of Uncer-
tainty (Risk Books, London).
Rolfo, Jacques, 1980, Optimal hedging under price and quantity uncertainty: The case of a
cocoa producer, Journal of Political Economy 88, 100-116.
Routledge, Brian, Duane Seppi, and Chester Spatt, 1999, The spark spread: An equilibrium
model of cross-commodity price relationships in electricity, Working paper, Carnegie Mellon
University.
Routledge, Brian, Duane Seppi, and Chester Spatt, 2000, Equilibrium forward curves for com-
modities, Journal of Finance 55, 1297-1338.
Smith, Clifford, and Ren6 Stulz, 1985, Determinants of firm's hedging policies, Journal of Fi-
nancial and Quantitative Analysis 20, 391-405.
Stulz, Ren6, 1990, Managerial discretion and optimal financing policies, Journal of Financial
Economics 26, 3-28.
U.S. Department of Energy, 2000, Electric Power Annual 1999 Volume II: Executive Summary,
Washington D.C. Document is available at http://www.eia.doe.gov/cneaf/electricity/epav2/
epav2_sum.html.
Wall Street Journal, 2000, New rules, demands put dangerous strain on electricity supply, May
11, Al.
Wolfram, Catherine, 1999, Measuring duopoly power in the British electricity spot market,
American Economic Review 89, 805-826.

The Journal of Finance - 2002 - Bessembinder - Equilibrium Pricing and Optimal Hedging in Electricity Forward Markets

Uploaded by

Copyright:

Available Formats

The Journal of Finance - 2002 - Bessembinder - Equilibrium Pricing and Optimal Hedging in Electricity Forward Markets

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

The Journal of Finance - 2002 - Bessembinder - Equilibrium Pricing and Optimal Hedging in Electricity Forward Markets

Uploaded by

Copyright:

Available Formats

15406261, 2002, 3, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/1540-6261.00463 by Indian Institute Of Technology, Kharagpur Central Library, Wiley Online Library on [23/12/2022].

Equilibrium Pricing and Optimal Hedging

HENDRIK BESSEMBINDER and MICHAEL L. LEMMON*

WHOLESALEPOWERMARKETS,where producers trade electricity among them-

clearing prices are volatile because inventories cannot be used to smooth

there is another important consequence of non-storability: using the spot

5 See Wall Street Journal (2000).

ever, the existence of a nonzero premium in forward power prices as implied

I. Power Production, Wholesale Markets,

next-hour and next-day delivery. In practice, power demand over a short

where F represents fixed costs, Qpi is the output of producer i, and c is a

If the cost parameter c is greater than two, marginal costs increase at an

B. The Demand for Risk Reduction

variance of profit. Optimal hedging strategies when the objective is linear in

C. Forward Market Participation

regional over-the-counter markets are the centers of price dynamics, with

Though we do not formally present an alternative model with costless

II. Wholesale Power Trading, Equilibrium Hedging,

A. The Wholesale Spot Market

ITPiPWQP + PFQF - F - - (Q)C (2)

tRj PRQRj + PFQ -

The profit-maximizing quantity sold in the spot market by producer i is

where x = 1/(c - 1). Equating the total physical production of producers to

Pw=a cl t N ' (5)

B. The Demand for Forward Positions

QF =PF-E (PW)+ COV(P{PR}i,PW) (9)

COV (PRj, PW) = PR COV(QRj, PW) - COV(PWQRj,PW) (11)

their costs of acquiring power on the wholesale markets, reflected by the

C. The Equilibrium Forward Price

PF= E(PW) - NcX [CPRCov(P`,Pw) - Cov(Pj ,

where N = (NR + NP)/A is a measure that reflects the number of firms in

PF= EE(PW)+ a Var(PW) + ySkew(Pw), (13)

e- 2NcaX (x[E(Pw)]xl1 - (x - 1)PR[E (PW)]x-2).

E. Optimal Forward Positions

E. 1. Optimal Forward Positions for Producers

QF E(QD) PREM (x[E(Pw)]Wx (Skew(Pw) (14)

Figure 4.Producers optima forward osition elative 0.7~~~~~~3...

Figure 4. Producer's optimal forward position relative to expected production. De-

E.2. Optimal Forward Positions for Power Retailers

market. Retailers, in contrast, can be heterogeneous if the probability dis-

QRip? Pi +,Pi N + Si, (15)

where pi = Cov(QRi QD)/NRVar(QD). This "power demand beta" measures

PREM ( Skew (PW)i

(= X(X 1) [E(Pw)]X-2 - ( [E(PW)] ),

III. Preliminary Empirical Evidence on the Pricing

B. Research Design and Data Description

60 _ __ --_ __ ____ _ 15000

ln(Pt) = a + (c - 1)ln(Qt + dt) + et, (17)

250 .........-----.- ..........

January 25.91 25.49 30.08 27.37

Standard Coefficient Standard Coefficient

Winter 23.02 7.10 0.18 27.36 7.55 0.65

C. Evidence on Forward Power Pricing

Bias relative to fitted spot prices

Figure 9. The bias in one-month-forward power prices as predictors of subsequently

compared to cost-based estimates of expected spot prices than when com-

Bias relative to fitted spot peices

m Jan Feb Mae r p May Jun / Jul Aug Sep \ Oc No tec

\ Bias relative to realized spot prices

Figure 10. The bias in one-month-forward power prices as predictors of subsequently

60 _ __ --_ __ _ 15000