Part 3

Market Architecture

1 Introduction
2 The Two-Settlement System
3 Standard Market Designs
4 Ancillary Services
5 The Day-Ahead Market in Theory
6 The Real-Time Market in Theory
7 The Day-Ahead Market in Practice
8 The Real-Time Market in Practice
9 The New Unit-Commitment Problem
10 A Market for Operating Reserves

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 Chapter 3-1
Introduction
The conclusion seventeen years later, is essentially the same . . . industries differ one from the
other, and the optimal mix of institutional arrangements for any one of them cannot be decided
on the basis of ideology alone. The “central institutional issue of public utility regulation”
remains . . . finding the best possible mix of inevitably imperfect regulation and inevitably
imperfect competition.
Alfred E. Kahn
The Economics of Regulation
1995

R EAL-TIME TRANSACTIONS REQUIRE CENTRAL COORDINATION;

WEEK-AHEAD TRADES DO NOT. Somewhere in between are dividing lines that
describe the system operator’s diminishing role in forward markets. Where to draw
those lines is the central controversy of power-market design. A related controversy,
not considered in Part 3, is how finely the system operator should define locational
prices. Those who favor a large role for the system operator in one sphere tend to
favor it in others. Thus the controversies of market architecture have a certain
consistency. Although the rhetoric focuses on how centralized a design is, the litmus
test in most of the controversies is the extent of the system operator’s role. This
too may be a distraction. A larger role for the system operator implies a smaller
role for profitable enterprises. One side fears the inefficiency and market-power
abuses of private parties playing social roles. The other side fears the inefficiency
of nonprofit organizations but also covets the central market roles played by the
system operator.
Power markets present unusually acute coordination problems. They are the
only markets that can suffer a catastrophic instability that develops in less than a
second and involves hundreds of private parties interacting through a shared facility.
The extent and speed of the required coordination are unparalleled. Generators
2000 miles apart must be kept synchronized to within a hundredth of a second.
Such considerations require a market that in some respects is tightly controlled in
real time. Historically, this control has extended to areas far from the precarious
real-time interactions. As deregulation brings markets into new areas, it is not
surprising to find the proponents of markets reaching beyond their ability and to
find the traditional system-control structure attempting to perpetuate now unneces-
sary roles for itself. This clash of interests has produced much heat and shed little
light.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-1 Introduction 203

While Part 2 ignores questions of architecture to focus on structure, Part 3

considers alternative designs for the real-time (RT) market and the day-ahead (DA)
market, as well as the relationship between the two. Most of the design questions
revolve around the extent of the system operator’s role.
Chapter Summary 3-1 : After preliminary definitions of forward, future, real-
time, and spot markets, this chapter outlines the controversies over bilateral markets,
power exchanges, and power pools. Bilateral markets provide a private coordinating
mechanism; exchanges provide a public, centrally determined, market price; and
pools provide a price, side payments and instructions on which generators should
start up. The final section provides a brief introduction to locational prices, the
complexity of which plays a role in assessing the need for central coordination.
Section 1: Spot Markets, Forward Markets, and Settlements. Forward
markets are financial markets, while the RT market is a physical market. To the
extent power sold in the DA market is not provided by the seller, the seller must
buy replacement power in the spot market.
Section 2: Architectural Controversies. The most basic controversy is over
the use of a bilateral DA market as opposed to a centralized market run by the
system operator. If the DA market is centralized, the second controversy is over
the use of a power exchange with a single price and simple bids as opposed to a
power pool with multipart bids and “make-whole” side payments. Multipart bids
are used to solve the unit-commitment problem, that is, to decide which generators
to commit (start up). This problem and the problem of dispatching around congested
transmission lines are the two technical problems that underlie the controversies.
Section 3: Simplified Locational Pricing. All markets discussed in Part 3
produce energy prices that are locationally differentiated, but the theory of such
prices is not presented until Part 5. The key properties of these prices are (1) they
are competitive prices, (2) the locational energy-price difference is the price of
transmission, and (3) a single congested line makes the price of energy different
at every location. Because they are competitive prices, any perfectly competitive
market, whether centralized or bilateral, will determine the same locational prices.

3-1.1 SPOT MARKETS, FORWARD MARKETS AND SETTLEMENTS

Trading for the power delivered in any particular minute begins years in advance
and continues until real time, the actual time at which the power flows out of a
generator and into a load. This is accomplished by a sequence of overlapping
markets, the earliest of which are forward markets that trade nonstandard, long-term,
forward contracts. Futures contracts are standardized, exchange-traded, forward

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

204 PART 3 Market Architecture

contracts. Electricity futures typically cover a month of power delivered during

on-peak hours and are sold up to a year or two in advance. Most informal forward
trading stops about one day prior to real time. At that point, the system operator
holds its DA market. This is often followed by an hour-ahead market and an RT
market also conducted by the system operator. All of these markets except the RT
market will be classified as forward markets.
All except the RT market are financial markets in the sense that the delivery
of power is optional and the seller’s only real obligation is financial. If power is
not delivered, the supplier must purchase replacement power or pay liquidated
damages. In many forward markets, including many DA markets, traders need not
own a generator to sell power. The RT market is a physical market, as all trades
correspond to actual power flows. While the term spot market is often used to
include the DA and hour-ahead markets, this book will use it to mean only the RT
market. A customer who buys power in a forward market will receive either
electricity delivered by the seller or financial compensation. This financial compen-
sation is called liquidated damages, meaning the damage to the customer has been
expressed as a liquid, financial sum. Because customers are virtually never discon-
nected when their forward contract falls through, power is delivered and they are
charged for it. This cost defines the liquidated damages. In most cases, a seller who
cannot deliver power from its own generator will purchase replacement power for
its customer. In either case the obligation has been met financially.
The most formal arrangement for purchasing replacement power occurs in the
system operator’s markets. Any power that is sold in the DA market but not
delivered in real time is deemed to be purchased in real time at the spot price of
energy. This is called a two-settlement system and has a number of useful economic
properties which are discussed in Chapter 3-2.

3-1.2 ARCHITECTURAL CONTROVERSIES

Three architectural controversies have plagued the design of power markets. All
three surfaced early and remain in dispute. Each has a decentralized side (listed
first) and a centralized side. These are:
1. Bilateral markets vs. centralized exchanges and pools.
2. Exchanges vs. pools.
3. Zonal pricing vs. nodal pricing.
Because the controversies have often been seen in ideological terms, discussion
has been characterized by black and white assertions. In reality, there are many
trade-offs and only a few clear-cut answers.
Part 3 does not address the third controversy; it is listed for completeness. But
all of the markets discussed in Part 3 are assumed to take place in the context of
locational pricing, so nodal pricing, or something similar, must be imagined in the
background. The complexity of locational pricing figures in the discussion of market
centralization. The first two controversies are addressed repeatedly as different
time frames and different problems are considered.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-1 Introduction 205

Two technical complexities and one problematic simplicity underlie these

controversies. The complexities are (1) transmission limits, and (2) the nonconvex
structure of generation costs. The first interacts with the physical laws of power
flow to produce different costs of delivered power at every point in the system when
even a single line restricts trade. This drives the third controversy, but it also plays
an important role in the first. Can a decentralized bilateral market solve the problem
of optimizing power flow over a grid in which every trade affects the flow of power
on every line?
The second complexity, referred to as the unit-commitment problem, also
involves a simultaneous optimization over all of the market’s generators. Startup
is costly, and the value of committing (starting) a unit of generation depends on
the cost of power produced by many other generators. Some believe a power pool
is needed to collect all the relevant data and make a centralized calculation in order
to determine if the value of starting is worth the cost. Others say a power exchange
provides enough centralization by computing a public market-clearing price.
Suppliers can use this to solve their own unit-commitment problems individually.
The problematic simplicity is the nature of AC power flow. This is most easily
understood by considering the grid at a time when line limits play no role; then
the power grid is like a pool of water. Any generator can put power in and any load
can take it out; no one knows where their power goes or where it comes from. From
a physical perspective it does not matter. But this makes normal private trading
arrangements impossible. Unless there is a centralized accounting of all trades,
any load or generator can steal power from the pool with impunity.
Bilateral vs. Central Markets. The first two controversies concern the role of
the system operator, which some wish to minimize on general principles. Chapter
3-4 considers the ancillary services that the system operator must either provide
or make sure are provided by appropriate markets. While bilateral markets are
reasonably efficient at providing the main service, bulk power supply, Chapter 3-6
argues that they are too slow to provide efficiently the two ancillary services most
crucial to reliability: RT balancing and transmission security.
Chapter 3-5 considers centralization of the DA market and finds the answer
is less obvious. With more time for trading, the slower bilateral design might
perform well enough, but the unit-commitment problem and the need for coordina-
tion with regard to transmission limits tip the balance. Though it may not be
essential, there is a strong case for at least the minimal central coordination provided
by a power exchange run by the system operator.
The decisive factor in all of these decisions is the need for speed. Bilateral
markets are slower than centralized markets. Because of the extreme complexity
of solving the unit-commitment problem and transmission problem simultaneously,
a bilateral market is simply too slow. They are slow mainly because they lack a
transparent market price. Price is the coordinating agent of free markets, and
bilateral markets make it difficult to discover while centralized markets make it
easy. By providing a transparent price, a centralized exchange makes finding an
efficient set of trades much easier and much faster. In power markets, that means
a great deal.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

206 PART 3 Market Architecture

This does not mean bilateral markets should be suppressed. They should be
encouraged as forward markets and allowed to exist beside the centralized DA
market.
Exchanges vs. Pools. Integrated utilities have always solved the unit-commitment
problem centrally, using many parameters to describe each generating unit. Incorrect
unit commitment, starting the wrong set of plants in advance, can lead to two
problems: (1) inefficiency, and (2) reduced reliability. Chapters 3-7, 3-8, and 3-9
consider whether it is worth moving beyond a power exchange to a power pool.
This would reproduce the old approach but with all of the parameters provided
by private parties and with the pool having no direct control over the dispatched
generators.
The disadvantages cited for an exchange are inefficiency and lack of reliability
due to lack of coordination. The disadvantages attributed to a pool are gaming
opportunities, and biases and inefficiencies caused by side payments. The complex-
ity and nontransparency of pools can also lead to design mistakes that are hard to
discover and correct.
None of the efficiency or gaming concerns have received serious quantitative
assessment, and all seem overrated. Either system should be capable of performing
quite well if well designed. Because the startup costs are only about 1% of retail
costs, and a simple exchange can already manage them quite efficiently, a small
increase in bid flexibility would seem to be sufficient. In other words, a little of
the power pool approach may be useful, but elaborate multi-part bids appear to
cause more problems than they solve. While startup insurance may give the system
operator some useful control over generator ramping, this could be obtained with
a more market-oriented approach.

3-1.3 SIMPLIFIED LOCATIONAL PRICING

Because several points concerning centralization re-
quire a partial understanding of the complexity of
Power Exchanges vs. Pools
locational pricing, this section gives a brief overview.
Pools are often associated with nodal pricing and
The properties presented here are explained more fully
exchanges are sometimes thought to require a
single price throughout their region. In practice in Chapters 5-3, 5-4, and 5-5.
exchanges have been associated with zonal prices, Energy prices differ by location for the simple
but there is no theoretical reason for this association. reason that energy is cheaper to produce in some loca-
Exchanges can provide nodal prices more easily tions and transportation (transmission) is limited.
than pools and pools were always run without nodal When a transmission line reaches its limit, it is said
pricing before deregulation. Part 3 assumes that the to be congested, and it is this congestion that keeps
choice of locational pricing will be unaffected by energy prices different in different locations. For this
whether a pool or an exchange is selected. reason locational pricing of energy is also called “con-
gestion pricing.”
Locational prices are just competitive prices, and these are unique. They are
determined by supply and demand and have nothing to do with the architecture
of the market, provided it is a competitive market. This means a purely bilateral
market that is perfectly competitive will trade power at the same locational prices

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-1 Introduction 207

as a perfectly competitive, centralized nodal-pricing market. Of course, a bilateral

market is likely to be less precise with its pricing, but on average it should find the
full set of competitive nodal prices.
Because there is a unique set of locational prices, there is also a unique set of
“congestion” prices, which will also be called transmission prices. Again, these
are determined by competition and supply and demand conditions and have nothing
to do with market architecture, provided the market is perfectly competitive.
If the competitive energy price at X is $20/MWh and at Y is $30/MWh, the price
of transmission from X to Y is $10/MWh. Transmission prices are always equal
to the difference between the corresponding locational energy prices. If this were
not true, it would pay to buy energy at one location and ship it to the other. In that
case arbitrage would change the energy prices until this simple relationship held.
The relationship can be expressed as follows:
PXY = PY ! PX
The price of transmission from X to Y equals the price of energy at Y minus the
price of energy at X. This relationship is the only one used in Part 3, but a few
related facts will provide a broader context.
When power flows from Y to X it exactly cancels (without a trace) an equal
amount of power flowing from X to Y thereby making it possible to send that much
more power from X to Y. Thus a reverse power flow from Y to X (a counterflow)
produces more transmission capacity from X to Y. As a consequence, if the price
of transmission from X to Y is positive, then the price from Y to X is the negative
of this value. This follows from the above formula, as does another consequence:
The cost of transmitting power from X to Y does not depend on the path chosen.
This is not surprising because, although contracts may stipulate a “contract path”
for power, there is no way to influence the actual path taken. Locational prices
reflect this reality by making sure that PXZ + PZY = PXY for any intermediate point
Z.
Not only is it impossible to select the path of a power flow, power takes every
possible path between two points, with more flowing on the easier routes. The
consequence for a network with a single congested line is that every location has
a unique price. In effect there is a price for using the congested line, and every
transaction uses that line to one extent or another. Sending power from X to fifty
different locations will use fifty different amounts of the congested line, so there
will be fifty different transmission prices and fifty different energy prices (plus
the energy price at X). One congested line in PJM produces 2000 different
locational prices. A centralized market will compute these so accurately that the
true locational differences can be seen. A bilateral market finds them imprecisely,
so many observed differences will be mainly due to the haphazard nature of the
bilateral process.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 Chapter 3-2
The Two-Settlement System

We can scarcely avoid the inference that light is the transverse undulations of the same medium
which is the cause of electric and magnetic phenomena.
James Clerk Maxwell
1861

This velocity is so nearly that of light, that it seems we have strong reasons to conclude that light
itself (including radiant heat, and other radiations if any) is an electromagnetic disturbance in
the form of waves propagated through the electromagnetic field according to electromagnetic
laws.1
1864

T HE REAL-TIME PRICE ALWAYS DIFFERS FROM THE DAY-AHEAD

PRICE. WHICH IS IN CONTROL? Day-ahead (DA) prices, and especially earlier
prices, differ significantly from the corresponding real-time (RT) price. The
differences are due to misestimations made before traders know all the details of
the RT conditions. In a competitive market the RT prices are true marginal cost
prices, and the forward prices are just estimates, sometimes very rough estimates.
With most trade occurring in the forward markets, does this imply that only a small
proportion of generation is subject to the correct incentives? Not under a proper
two-settlement system. The purpose of the RT market is to correct the prediction
errors of the past. If the transaction costs in this market are minimized so that
profitable trade is maximized, the RT price will be accurate and will control actual
production. Past mistakes have financial impacts but will not cause inefficiency
which is a purely physical phenomenon.1
Contracts for differences (CFDs) insulate bilateral trades from all risks of spot
price fluctuations while allowing the inevitable inefficiencies of forward trading
to be corrected by accurate RT price signals. Both the two-settlement system and
CFDs allow efficient re-contracting—a standard economic solution to the problems
of decentralized forward trading. Advocates of bilateral trading have often failed
to recognize this point and have opposed the very mechanisms that make decentral-

1. Maxwell developed the mathematics of electromagnetic fields, later used to design AC motors,
transformers, and power lines. He predicted the possibility of electromagnetic waves and calculated their
theoretical velocity from laboratory measurements on electric and magnetic fields. At first his suggestion
that light is electromagnetic was dismissed as a “not wholly tenable hypothesis.”

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-2 The Two-Settlement System 209

ized trading efficient. This chapter assumes the existence of a centralized RT market
in which all generators and loads must participate.
Chapter Summary 3-2 : If a generator sells its output in the DA market, the
two-settlement system lets it respond efficiently to the spot price without any risk
from the volatility of that price. It can only profit from an unexpected spot price,
and never suffer a loss. If a generator sells its power to a load in a bilateral contract
months in advance, a CFD will let them profit efficiently from an unexpected spot
price. If they trade over lines that may be congested, purchasing an FTR (financial
transmission right) will provide the same guarantee with respect to transmission
prices. In this way forward trades that prove inefficient in real time because of
unexpected circumstances can be corrected without risk to the traders.
Section 1: The Two-Settlement System. If the system operator runs a DA
and an RT market, generators should be paid for power sold in the DA market at
the DA price, regardless of whether or not they produce the power. In addition,
any RT deviation from the quantity sold a day ahead should be paid for at the RT
price.
A CFD requires the load to pay the generator the difference between the contract
price and the spot price whether it is positive or negative. This allows either party
to deviate profitably from the contract, when the opportunity arises, without
affecting the other. If the spot price differs at the two locations, this hedge is not
complete.
Section 2: Ex-Post Prices: The Trader’s Complaint. Spot prices that differ
by location impose transmission costs on traders. These cannot be avoided by the
use of CFDs, and they make trade risky. Some markets in transmission rights exist
but are generally limited and illiquid. Design of such markets is continuing. A
financial transmission right (FTR) from generator to load can perfectly hedge a
bilateral trade that faces congestion charges. Since trade is always allowed in the
RT market, an FTR is as good a guarantee as a physical right.

3-2.1 THE TWO-SETTLEMENT SYSTEM

If a supplier sells most or all of its power in the forward markets, the RT price may
appear to have little chance of affecting the production decisions of suppliers. In
a properly implemented two-settlement system the opposite is true. In real time,
the supplier will behave as if it were selling its entire output in the RT market, even
though, in the forward market, it acts as if that were its final sale.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

210 PART 3 Market Architecture

Separation of Real-Time from Forward Transactions

Say a supplier sells Q1 to the system operator in the DA market for a price of P1.
If this amount of power is delivered to the RT market, the settlement in the DA
market will hold without modification. But what if none is delivered, or more than
Q1 is delivered? In either case the DA settlement should still hold, but there should
be an additional settlement in the RT market. If no power is delivered to the RT
market, the supplier is treated as if it had delivered the amount promised in the DA
market, Q1 , but purchased that amount from the RT market to cover its promised
delivery. Consequently the supplier is still paid P1 for Q1 but is also charged P0 ,
the RT price, for the purchase of Q1 . In general, if a supplier sells Q1 in the DA
market and then delivers Q0 to the RT market, it will be paid:
Supplier is paid: Q1 × P1 + (Q0 ! Q1 ) × P0 (3-2.1)
This is called a “two-settlement system.” If a customer contracts for Q1 and then
takes only Q0 in real time, it is charged exactly the amount that its supplier is paid.

Result 3-2.1 A Two-Settlement System Preserves Real-Time Incentives

When the RT market is settled by pricing deviations from forward contracts at the
RT price, suppliers and customers each have the same performance incentives in
real time as if they had traded all of their power in the RT market.

The incentives of this settlement rule are revealed by rearranging the terms as
follows:
Supplier is paid: Q1 × (P1 ! P0 ) + Q0 × P0 (3-2.2)
When real time arrives, P1 and Q1 have been determined in the day-ahead (DA)
market. Assuming the market is competitive, suppliers will also take P0 as given,
so by real time, the entire first term will be viewed as a “sunk” cost or an assured
revenue. This leaves the second term as the only one that can provide an RT
incentive for generator behavior, and this term pays the generator the RT price for
every megawatt produced. Consequently the generator will behave exactly as if
it is selling all of its product in the RT market. This can be proven by considering
the supplier’s profit, which is revenue minus cost, and the profit it would have had
if it traded only in the RT market.
Table 3-2.1 Profit With and Without a Day-Ahead Trade
Actual Short-Run Profit: SRB F = RF + Q0 × P0 ! Cost(Q0 )
Only-Real-Time Short-Run Profit: SRB0 = Q0 × P0 ! Cost(Q0 )

The only difference between the two is the fixed revenue, RF = Q1 × (P1 ! P0 ), so
the value of Q0 that maximizes one will maximize the other.
This result means that no matter what trades, Q1 , have taken place in the DA
market, or any other forward market, profit-maximizing suppliers will pursue the
same RT strategies as if no prior trades had taken place. Consequently, if the RT
market is competitive and therefore efficient, this efficiency will not be undermined
by forward contracts. Put another way, if mistakes are made in forward markets

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-2 The Two-Settlement System 211

they will affect revenues but not efficiency because the RT market will induce least-
cost operation regardless of these mistakes. The above argument also applies to
loads.

Separation from Imports and Exports

If a supplier sells and schedules Q1 for export in a forward market for a price of
P1 , and delivers this much power locally to the RT market, the exporter will owe
nothing and be paid nothing by the local DA and RT markets. The exporter will
be paid only by the external purchaser. But what if no power is supplied or more
than Q1 is supplied? As before, the supplier will be paid (Q0 ! Q1 ) × P0 by the RT
market, where Q0 is its RT supply. If Q0 is zero, the supplier is paid !Q1 × P0 , which
means it is charged Q1 × P0 . This charge assumes that it has purchased in the RT
market the power that it exported. Of course the transaction could be cancelled,
in which case Q1 would be adjusted to zero, and the RT payment formula would
continue to apply.

Separation from Bilateral Markets

If a trader has arranged to buy Q1 at price PG from a generator and sell it to a load
at price PL , how will the participation of the generator and load in the RT market
be handled? This transaction does not make the trader a participant in the RT
market, yet the generators and loads must participate. Say the generator injects
Q1 and the load withdraws Q1 as the contract demands. The generator will be paid
P0 × Q1 by the RT market which is different from what was specified in the bilat-
eral contract with the traders. To compensate for this, the trader must specify that
the generator will pay the trader P0 × Q1 and the trader will pay the load the same
amount. This is over and above the payments specified for the original purchase
and sale of Q1 by the trader. The entire transaction works like this (Table 3-2.2):
Table 3-2.2 Details of a Bilateral Trade with Adjustments
Trade RT Market Adjustment
Generator is paid: P G × Q1 + P 0 × Q1 !P0 × Q1
Load pays: P L × Q1 + P 0 × Q1 !P0 × Q1
Trader’s net income: (PL ! PG ) × Q1 + P 0 × Q1 ! P 0 × Q1 = 0

The first term for both generator and load is the payment specified by the original
trade. The second term is the result of each participating in the RT market, and the
third term is the adjustment term specified by the trader to keep the bilateral trade
separate from the RT market. The adjustment exactly cancels the RT settlements
of both generator and load, and for the trader, the two adjustments cancel each other.
Contracts for Differences. Combining the trade and the adjustment for the
generator defines a bilateral trade of a special type called a contract for differ-
ences, a CFD. The detailed trades shown above contain two of these plus the trades
with the RT market. The following table shows the two contracts for differences

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

212 PART 3 Market Architecture

used in the above bilateral trade as well as a CFD written directly between a
generator and a load.
Table 3-2.3 Contracts for Differences
Trader pays generator: (PG ! P0 ) × Q1
Load pays trader: (PL ! P0 ) × Q1
Direct CFD: Load pays generator (PC ! P0 ) × Q1

Each bilateral trade is the contract quantity times the difference between the contract
price and the RT price. If the generator contracts directly with the load, there is
only one contract price, PC , and the load pays the generator (PC ! P0 ) × Q1 .
By writing the bilateral contracts, shown in Table 3-2.3, with the generator and
the load, a trader can implement the detailed bilateral contract displayed in Table
3-2.2. The RT market and adjustment terms cancel, and the effect is just as if the
RT market did not exist. The same is true for a trade between a generator and a
load that is implemented with a CFD. These conclusions assume the trade takes
place as specified in the contract.

Contracts-for-Differences Result #1
Result 3-2.2 A Contract for Differences Insulates Traders from Spot Price Volatility
Bilateral trades implemented through contracts for differences completely insulate
traders from the market price provided (1) the traders produce and consume the
amounts contracted for, and (2) the market price is the same at the generator’s
location as at the load’s location.

When CFDs are used, the only effect of the spot market is to provide a convenient
remedy for deviations from the contract position. Such deviations only affect the
party who deviates.
Remarkably, while insulating the bilateral trade from the spot market, the CFD,
together with the two-settlement system, also insulates the traders’ use of the spot
market from the effects of the bilateral trade. The argument used in the beginning
of this section can be used again to show that both customer and supplier will
behave as if they were trading in the RT market because their incentives to deviate
from the contract quantity are determined by the RT price. Such deviations are
desirable and can only make them better off while doing no harm to their trading
partners. CFDs give us the best of both worlds. Traders are protected from the
volatility of spot prices, and the efficiency of the RT market is protected from the
inefficiency of forward bilateral contracting.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-2 The Two-Settlement System 213

Contracts-for-Differences Result #2
Result 3-2.3 Contracts for Differences Preserve Real-Time Incentives
Bilateral traders using contracts for differences feel the full incentive of RT
prices. Because they could ignore this incentive, any deviation from their contract
can only be profitable.

Locational Price Differences

So far, spot prices have been assumed equal at all locations in the market so both
generator and load see the same price, but this is often not the case. As a conse-
quence, the CFD can be written in different ways. To analyze this situation, assume
that generator and load have signed a CFD without the benefit of a middleman.
There are two ways to write the CFD:
Table 3-2.4 Two Possible CFDs
Effect of CFD Load’s Payment to Generator
Load pays for transmission: (PC ! PG0 ) × Q1
Generator pays for transmission: (PC ! PL0 ) × Q1

PG0 is the spot price at the generator’s bus, PL0 is the spot price at the load’s bus,
and PC is the contract price. In both cases, the CFD specifies the payment by loads
to generators. In the first case, because the generator’s RT price is used, the genera-
tor is insulated from locational price differences, while in the second case, load
is insulated. To see this, the full transaction, including the RT market, must be taken
into account. Consider the case in which the generator’s spot price is used and in
which each player produces or consumes as specified in the contract.
Table 3-2.5 Settlements with the Generation-Centric CFD
CFD Spot Market
Generator is paid: (PC ! PG0 ) × Q1 + PG0 × Q1
Load pays: (PC ! PG0 ) × Q1 + PL0 × Q1

The first term is the CFD settlement and the second is the spot market settlement.
These settlements can be simplified as follows:
Table 3-2.6 Algebraic Simplification to Reveal Transmission Charge
Trade Transmission Charge
Generator is paid: P C × Q1 0
Load pays: P C × Q1 + (PL0 ! PG0 ) × Q1

Both will pay or receive the contract price, but the load must also pay the locational
price difference times the quantity traded. Typically the spot price will be higher
at the load’s location, so this charge will be positive. The charge is termed a
congestion charge or a transmission charge, and it arises from the scarcity of

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

214 PART 3 Market Architecture

transmission. It is like a transportation or delivery charge but is usually more

volatile.
Congestion charges are covered in Part 5, but for present purposes it is only
necessary to understand that they are based entirely on scarcity.2 They are typically
zero because there is plenty of transmission capacity most of the time. When it is
scarce, competition for transmission can send the price quite high. To some extent
these prices are predictable, but they contain a significant random component that
can be problematic for traders (see Section 3-1.3).
The uncertainty in the congestion price (PL0 ! PG0 ) can be hedged by buying
energy forwards in the two locations or by buying transmission rights between the
two locations. If the CFD is arranged so that the load pays the congestion charge,
the load may want to buy a hedge. If the CFD is set up the other way, the generator
may want to buy it. If the trade is arranged by a trader, the trader will probably
accept the transmission charge and may want to hedge it.

3-2.2 EX-POST PRICES: THE TRADER’S COMPLAINT3

Power traders can write CFDs and thereby completely insulate themselves from
the spot price (and the DA price). But it is much more difficult for them to insulate
themselves from locational price differences, that is, transmission prices. If they
have used a CFD to execute their bilateral trade and they trade as planned, the spot
price, no matter how high, will have no effect. Without a transmission right, the
spot transmission price must be paid in full.
Why is it so easy to insulate a bilateral trade from the spot energy price and
so difficult to insulate it from the spot transmission price? A buyer and seller,
considered as a unit, are unaffected by the energy price because together their net
position is zero. As a unit, however, they always take a net position in the transmis-
sion market; they consume transmission from generator to load.4 Because they take
a nonzero net position in the transmission market, they are affected by the price
of transmission.
The complaint of traders is that the transmission price is “ex-post”—it is
established after they commit to a trade instead of being posted ahead of time. They
would prefer to check the price of transmission, arrange a trade based on that price,
then call up the transmission provider and purchase the desired quantity at the price
they were quoted. United Parcel Service works this way, and their prices have two
convenient features. First, they are posted in advance, and second, they are good
no matter how many packages are shipped. Real-time transmission prices have
neither feature. Access charges, used to cover the fixed costs of transmission, do
have these features, but they are not the focus of the trader’s complaint which

2. This ignores the charge for losses, which is almost never above 10% and is far more predictable.
3. These “ex-post prices” are not necessarily those that result from the system of “ex-post pricing”
discussed in the preface. True “ex-post pricing” is said to use a quantity optimization procedure that has
not been publically specified but is rumored to be central to the PJM and NYISO pools.
4. If they trade against the prevailing flow, they create a counterflow and their consumption of
transmission service is negative, but they have still taken a position in the transmission market.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-2 The Two-Settlement System 215

concerns only congestion charges. Two approaches to resolving this problem need
consideration: can (1) RT price setting or (2) forward markets for transmission be
improved?

Better Real-time Price Setting?

The system operator could post the RT congestion prices in advance. Because
congestion prices fluctuate somewhat predictably, this would be a time-of-use
transmission charge. The bulk of bilateral trading takes place months in advance,
and these charges would need to be set annually, or preferably even further in
advance. But congestion prices, as observed in systems that compute them continu-
ously, have only a small component of predictability. They are susceptible to
weather, generation outages, transmission outages, and other factors. When the
time-of-use price proved to be too high, valuable transmission would go unused.
At other times, demand for transmission would outstrip supply, and traders would
find none available at the posted price. Curtailments would occur in real time, an
outcome that would be very inefficient and extremely unpopular with traders.
To be of any use, time-of-use pricing must be coupled with a reservation system.
FERC’s pro-forma tariff provides such a system although it is based on contract
paths rather than on actual power flows and thus conflicts with the laws of physics.
With a system of interconnection-wide coordination, such as NERC is now devising,
a consistent reservation system would be possible. This would avoid most RT
curtailment, but traders would still find that the posted price would not guarantee
availability of reservations. Without a real market for transmission rights, nonprice
rationing would fail to select the highest-value users and lead to an inefficient
dispatch. At other times, trades would be curtailed unnecessarily.
Currently, no estimate of the inefficiency of such administratively set prices
is available. The benefits of such a system are even more difficult to determine.
Facilitating bilateral trading, especially long-term trading, is desirable, but the
benefits are hard to quantify. In any case, market-based prices is generally consid-
ered preferable to regulated prices.

Better Forward Markets for Transmission

A second approach to alleviating the trader’s complaint is to implement a forward
market in transmission rights. This has the same effect as the posted price and
reservations, but the price is determined by the supply and demand for transmission
instead of being administratively determined. PJM sells such rights in its DA market
as described in Chapter 3-3. Although this is a help, by the time traders see the DA
price of transmission, it is too late to take advantage of it. Because it is a rough
guide to RT congestion prices, it is some help in deciding whether to sign a last-
minute bilateral contract for power and take a chance on the RT transmission price.
A continuous market is needed with a slowly changing price that traders can observe
before they arrange a trade. Afterwards they can purchase transmission at a price
close to the observed price. So far, such markets have not been developed because
of technical and practical problems, but new proposals are currently being debated.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

216 PART 3 Market Architecture

PJM and the NYISO also sell firm point-to-point financial transmission rights
between any two points in their systems. Unfortunately these markets are quite
illiquid and a trader may have to wait for a quarterly auction to purchase the desired
right, although purchase in a secondary market is sometimes possible.
If well constructed, financial transmission rights provide the security of physical
rights without some of the drawbacks. Consider how a financial right is used to
hedge a bilateral trade in the RT market. A financial transmission right, FTR, from
G to L for Q1 MW is defined to pay5
FTR from G to L for q pays: (PL0 ! PG0 ) × Q1 .
Using this FTR, reconsider the last bilateral trade of the previous section in which
load ended up paying a transmission charge.
Assuming the load has purchased this FTR and the same generation-centric
CFD is used, the settlement will be as follows (Table 3-2.7):
Table 3-2.7 FTR Settlements with the Generation-Centric CFD
CFD Spot Market FTR
Generator is paid: (PC ! PG0 ) × Q1 + PG0 × Q1
Load pays: (PC ! PG0 ) × Q1 + PL0 × Q1 !(PL0 ! PG0 ) × Q1

Note that the FTR payment is included with a minus sign because “load pays” a
negative amount—it is paid this much by the FTR. This time when the settlements
are simplified, the load pays PC × Q1 to the generator and neither pays a transmis-
sion charge. Of course the FTR had to be purchased and that cost is not shown in
the settlement, but once it is purchased, the parties are immunized completely
against the RT transmission charge. By using both a CFD and the FTR they are
protected from fluctuations in the general level of spot prices and from locational
differences in the spot price.
If the trade is made, there will be no transmission charge, but what guarantees
that the parties will be allowed to trade when they have no physical reservation?
Generally, RT locational energy markets allow traders to do as they wish, but they
must pay for their injections and withdrawals at the RT price. In this case no special
action is needed by the parties; they simply make the trade and accept the charges.
Provided their trade matches their FTR, the net charges will be covered exactly.
If the RT market requires bids, then the generator submits the lowest possible bid
(in PJM this would be !$1,000/MWh), and the load submits the highest possible
bid. Unless there is some physical problem with the network (in which case not
even a physical reservation can guarantee the trade), they will be dispatched.
Because they own the FTR they cannot be harmed by any price that results from
their extreme bids. The net result is an assurance that they will be able to complete
the trade and be completely unaffected by the RT prices.

5. This is the definition of a transmission congestion contract (TCC) as found in Hogan (1992). Many
variations on this have been proposed and used.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 Chapter 3-3
Day-Ahead Market Designs

From a long view of the history of mankind—seen from, say, ten thousand years from
now—there can be little doubt that the most significant event of the 19th century will be judged
as Maxwell's discovery of the laws of electrodynamics.
Richard Feynman

One scientific epoch ended and another began with James Clerk Maxwell.1
Albert Einstein

C ENTRAL DAY-AHEAD MARKETS CAN BE DESCRIBED AS AUCTIONS.

The most obvious design sets energy prices based on simple energy-price bids. A
different approach turns the system operator into a transportation-service provider
who knows nothing about the price of energy but sells point-to-point transmission
services to energy traders.1
Either of these approaches presents generators with a difficult problem. Some
generators must engage in a costly startup process (commitment) in order to produce
at all. Consequently, when offering to sell power a day in advance, a generator needs
to know if it will sell enough power at a price high enough to make commitment
worthwhile. Some day-ahead (DA) auctions require complex bids which describe
all of a generator’s costs and constraints and solve this problem for the generators.
If the system operator determines that a unit should commit, it ensures that all its
costs will be covered provided the unit commits and produces according to the
accepted bid. Such insurance payments are called “side payments,” and their effect
on long-run investment decisions is considered in Sections 3-9.3 and 3-9.3.
Chapter Summary 3-3 : Day-ahead markets run by system operators are run
as auctions. Although some trade energy, some sell transmission, and some solve
the unit-commitment problem, they all use the same philosophy for choosing which
bids to accept and for setting prices. Four archetypical markets are summarized:
(1) a power exchange, (2) a transmission-rights market, (3) a power pool, and (4)
PJM’s DA market which mixes all three.

1. The universe is governed by four forces and matter is made of their associated particles. The electron
and photon are the carriers of the electromagnetic force, and the other three forces are gravity, the weak
force and the nuclear force. Maxwell discovered the mathematics of the electromagnetic force, Einstein
the mathematics of the gravitational force, and Feynman the laws of the weak force which he unified with
the electromagnetic force.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

218 PART 3 Market Architecture

Section 1: Defining Day-Ahead Auctions. In DA auctions, bids are selected

to maximize total surplus, gross consumer surplus minus the cost of production.
Locational energy prices are set equal to the marginal change in total surplus when
free power is injected at the various locations. Transmission is priced at the marginal
change in total surplus when a counterflow is introduced on a transmission path.
Section 2: Four Day-Ahead Market Designs. Each auction is specified by
three sets of conditions: bidding rules, bid acceptance rules, and settlement rules.
Market 1 is a power exchange that uses one-part bids. Market 2 is a “bilateral”
transmission-rights market. Market 3 is a power pool which uses multipart bids
for unit commitment. Market 4 combines the features of the other three as options.
Section 3: Overview of the Day-Ahead Design Controversy. Forward
markets are bilateral, and real-time (RT) markets are centralized. The DA market
can be designed either way, and this causes a great deal of controversy. The
“centralized nodal pricing” approach specifies an energy market with potentially
different prices at every node (bus) and specifies that the auction should solve the
unit-commitment problem as well. This requires complex bids. The bilateral
approach specifies that energy trades take place between two private parties and
not between the exchange and individual private parties. To trade energy, the private
parties require the use of the transmission system, so the system operator is asked
to sell transmission.

3-3.1 HOW DAY-AHEAD AUCTIONS DETERMINE QUANTITY AND PRICE

Day-ahead markets run by system operators take the form of either exchanges or
pools and are operated as auctions. The process of selecting the winning bids is
often complicated by transmission and generation constraints which can require
the use of enormously complex calculations and sophisticated mathematics. The
mathematics is often presented as a way of explaining the auction. This is unneces-
sary, frequently confusing, and often less precise than an approach that explains
the purpose rather than the mechanics.

A Simplified Description of Auctions

All auctions solve some mathematical problem. Bids are submitted, some function
of the bids is maximized or minimized, and the solution determines which bids
are accepted. For example, say bids are submitted for the purchase of 100 tulip
bulbs. Each bid states a number of bulbs and a total price. The auction problem
might be to maximize the sum of the accepted bid prices. The solution defines a
set of accepted bids. There is one more step—settlement. The winners must pay
or be paid and must provide or receive the goods being auctioned. Settlement is
not determined entirely by the solution to the acceptance problem but also uses

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-3 Day-Ahead Market Designs 219

separate price-determination rules. In this example the rules could specify that each
accepted bid would pay as bid and receive the number of bulbs bid for, or it might
specify that accepted bids would pay a price per bulb equal to the lowest price per
bulb of any accepted bid.

Summary An Auction’s Three Stages: Bidding, Acceptance, and Settlement

An auction takes place in three stages
1. Bids are submitted.
2. Some bids are accepted and prices determined.
3. Accepted bids are settled at the determined prices.
The optimization problem that determines which bids are accepted does not
automatically determine prices. The pricing rule must be specified separately.

The description of DA markets often focuses on presenting the auction problem,

which bids to accept and what prices to set, in a format that is convenient for
solution using linear programming techniques. The solution technique is of little
interest from an economic perspective, so presenting the problem in this way adds
little value and often much complexity. For example, the auction problem might
be to minimize the total price of accepted bids of generators while purchasing
enough power to supply a 1000-MW load at location X and taking into account
losses. This problem can be presented as a linear programming problem that
approximates power loss equations. But the necessary equations, while invaluable
to the computer programmer, add relatively little to understanding the economics.
Much of the complexity of electricity auctions is embodied in the constraints
placed on bid acceptance by the physics of the system. Given a source (a generator)
and a sink (a load), the path of the power flow is usually uncontrollable. Conse-
quently, to avoid overloading lines, certain combinations of bids cannot be accepted.
Losses pose similar restrictions on bid acceptance.
Instead of specifying physical constraints in abstract mathematical detail, the
following four market summaries will simply name the constraints. The focus can
then shift to the bidding rules, the quantity being optimized in the auction problem,
and the settlement rules. Settlement rules in particular often include overlooked
inducements and penalties that are crucial to the functioning of a market. The shift
in focus to economic aspects, and the standardization of the descriptive format
allows easier and more meaningful comparisons of these archetypical designs for
DA markets.

Accepting Bids: Determining Quantities

The auction problem determines what bids are accepted, or partially accepted, and
this determines through simple accounting the quantities bought and sold. Prices
are determined by a separate set of settlement rules. Quantities determine the
efficiency of a particular outcome, but prices provide the incentives that determine
what bids are submitted and thereby help control the outcome.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

220 PART 3 Market Architecture

All four auction designs considered in this chapter maximize total surplus as
defined by the bids. These designs encourage truthful bidding, and in a competitive
market, bids will reflect true costs and benefits, and the auctions will maximize
actual total surplus.
Total surplus is the sum of (net) consumer and producer surplus, but
it is also the gross consumer surplus minus production costs. If a con-
sumer offers to buy 100 MWh at up to $5,000/MWh, and the bid is
accepted, the gross consumer surplus is $500,000. If the market price
is $50/MWh, and the customer’s cost is $5,000, the net consumer surplus
is $495,000. Similarly, if a generator offers to sell 100 MW at $20/MWh,
its cost is presumed to be $2,000. If the market price is again $50/MWh,
its producer surplus will be 100 × ($50 ! $20), or $3,000. Total surplus
is $498,000. Writing this calculation more generally reveals that price
played no role in determining total surplus:
Total surplus = Q × (V ! P) + Q × (P ! C) = Q × (V ! C),
where Q is the quantity traded, V is the customer’s gross surplus, C is
the producer’s cost, and P is the market price. If many bids are involved, P cancels
out of each trade. The problem of maximizing total surplus can be solved independ-
ently of price determination.
In an unconstrained system, total surplus can be maximized by turning the
demand bids into a demand curve and the supply bids into a supply curve and
finding the point of intersection. This gives both the market price and a complete
list of the accepted supply and demand bids. Unfortunately transmission constraints
and constraints on generator output (e.g., ramp-rate limits) can make this selection
of bids infeasible. In this case it is necessary to try other selections until a set of
bids is found that maximizes total surplus and is feasible. This arduous process
is handled by advanced mathematics and quick computers, but all that matters is
finding the set of bids that maximizes total surplus, and they can almost always
be found.

The Efficient-Auction Result

Result 3-3.1 A Single-Price Day-Ahead Auction Is Efficient
Ignoring nonconvex costs and market power, single-price DA auctions are
designed to maximize the sum of producer and consumer surplus. Having done
this they determine locational prices that support this pattern of production and
consumption. Efficiency depends on honest bidding which results from single-
price settlement and competition between bidders.

Determining the Market Price

The price determined by supply and demand is the highest of all accepted supply
bids or the lowest price of an accepted demand bid. It depends on whether the
intersection of the two curves occurs in the middle of a demand bid or in the middle

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-3 Day-Ahead Market Designs 221

Figure 3-3.1
Changes in total surplus
from the addition of
1 kW of free supply will
be called the marginal
surplus.

of a supply bid. When the demand curve is vertical, the intersection is always in
the middle of a supply bid, and the price is set to the supply-bid price.
Whichever curve is vertical at the point of intersection has an ambiguous
marginal cost or value (see Chapter 1-6). If the demand curve has a horizontal
segment at $200 that intersects a vertical part of the supply curve that runs from
$180 to $220, then the marginal cost of supply is undefined but is in-between the
left-hand marginal cost of $180 and the right-hand marginal cost of $220. (See
Chapter 1-6.) Consequently it causes no problem to say that the market price equals
both the marginal cost of supply and the marginal value of demand.
While the intersection of supply and demand is a convenient method of deter-
mining price in an unconstrained market with a single price, it is cumbersome to
apply to a constrained market with many prices. An alternative approach sets price
equal to marginal surplus. It determines the same price as the intersection of supply
and demand and is easier to use in a constrained system.

Definition Marginal Surplus

Marginal surplus is the change in total surplus with an unit increase in costless
supply. (Total surplus is the sum of consumer and producer surplus as well as the
difference between the supply and demand curves up to the quantity traded.)

First consider the unconstrained markets shown in Figure 3-3.1. The change
in total surplus when a kilowatt is added at no cost to the total supply of power
is the marginal total surplus, which because of its awkward name, will be called
simply the marginal surplus.2 In the case depicted on the left of Figure 3-3.1, the
free unit of supply shifts the entire supply curve left while the amount produced
and consumed remains unchanged. The result is a production cost savings of P
times 1 kW, so the marginal surplus is P. In the case depicted on the right, consump-
tion is limited by the high cost of the next available generator, but if another kilowatt
of free supply were available, consumption would increase by 1 kW. This case is
more easily analyzed by adding the free generation out of merit order as shown.
The value of consumption increases by P times 1 kW, and the cost of production
2. Total surplus should be expressed in $/h. A kilowatt, rather than a megawatt, is used to indicate that
only a “marginal” change is being made. Technically one should use calculus, but this is of no practical
significance.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

222 PART 3 Market Architecture

Figure 3-3.2
Changes in total surplus
from introducing a free
kilowatt of supply at the
consumer and supplier
ends of a transmission
constraint.

stays constant, so total surplus increases by the same amount. In both cases marginal
surplus equals price.
Next consider the constrained market diagrammed in Figure 3-3.2. All suppliers
are on one side of the transmission constraint and all consumers are on the other,
and the free kilowatt of supply can be introduced at either location. At the consumer
end, the kilowatt increases consumption, so the marginal surplus is the marginal
value of consumption, PD. At the supplier end, the kilowatt decreases the cost of
supply, so the marginal surplus is the marginal cost of
Does Price Equal Marginal Cost? supply, PS. Because of the constraint, marginal value
Prices set by the auctions under consideration are and marginal cost are different, and thus the two prices
often called marginal-cost prices. When the demand are different. Because the prices are different, the
curve is vertical at the intersection of supply and system operator captures part of the total surplus, but
demand, this term is accurate. With the current lack in spite of this, total surplus still equals gross consumer
of demand elasticity, it is almost always correct. surplus minus the cost of production.
When the demand curve intersects a vertical If the supply and demand curves were smooth, the
supply curve, a range of marginal costs is deter- results would be the same, but in an unconstrained
mined, as explained in Chapter 1-6. Then price is set
market, consumer value and production cost would
to the marginal value of demand which is within the
range of ambiguity of marginal cost.
jointly determine marginal surplus. If there were more
constraints and more distinct locations, there would
Marginal surplus is defined in both cases and
never contradicts marginal cost. Price always equals be more prices. If demand were vertical, marginal cost
marginal surplus and is always equal to marginal would determine marginal surplus at every location.
cost or within its range of ambiguity. This technique for computing price handles all neces-
sary complexities.
Setting the market price equal to the marginal surplus is justified because it gives
the competitive price and thus induces efficient behavior. It also clears the market,
which means all accepted bids will voluntarily comply with the settlement, and
all rejected bids will suffer no loss given the settlement price. Trade is voluntary
at these prices.
A fundamental difference between market-based and auction-based price
determination is that in markets marginal cost is set equal to price, while in auctions,
price is set equal to marginal cost.3 (See Section 1-5.1) This causes no problem

3. This is not always strictly correct. The logic of auction price determination runs from optimization of
trade through determination of marginal costs which then define prices. But the computer programs that
solve these problems sometimes make use of the market’s logic to find their solution.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-3 Day-Ahead Market Designs 223

as long as marginal costs are correctly determined by the auction, which they are
because the auction problem is specified as the maximization of total surplus. This
minimizes production costs given consumption and maximizes consumer value
given production. This determines the same efficient pattern of production as a
competitive market and thus the same marginal costs.

3-3.2 SUMMARIES OF FOUR DAY-AHEAD MARKETS

This section summarizes four archetypical DA markets, from simple to complex.
Each is a “locational” market, and these locations may be either single buses or
zones containing several buses.4
Market 1: A Power Exchange. A power exchange is a centralized market that
does not use (“make-whole”) side payments. Market 1 is a classic exchange in that
it employs one-part bids—its bids consist of only a supply or demand curve. Califor-
nia’s Power Exchange was a standard power exchange in its operation but had
imposed on it a peculiar relationship to other “scheduling coordinators” and to the
CA ISO. Such relationships have nothing to do with the definition of an exchange
market. Alberta’s power exchange uses two-part bids and serves as the completely
centralized RT spot market. Both are exchanges because they do not use generator-
specific side payments and both are typical in that their bid formats are simple.

Definition Power Exchange

An exchange market is a centralized market that does not use side payments. At
any given time and location it pays the same price to any generator selling power.
It can use multiple rounds of bidding or multipart bids to determine this price and
can implement full nodal pricing. Typically, its bids are much simpler than those
used by pools, but a centralized market using 20-part bids but not making side
payments would still be an exchange.

In some respects Market 1 is the simplest DA market. Participants do not search

for trading partners and do not have to consider many prices in many locations.
Each trader simply trades with the exchange at the trader’s location. The system
operator’s job is simple because it ignores the unit-commitment problem. One
difficulty, discussed in Chapter 3-9, is that competitive suppliers will not always
find it most profitable to bid their marginal costs.
Market 2: A Transmission-Rights Market. A transmission market is equally
simple for the system operator but requires a complex pre-market step for market
participants. Buyers and sellers must find each other and make provisional energy
trades that are contingent on the outcome of the DA transmission market, or they
must buy transmission on speculation. As with an energy market, most transactions

4. If zones are used, the transmission constraints will represent the market less accurately, and a more
conservative representation of constraints may be required. This affects only the details of the constraint
specification and not the specification of the markets. However, zones and other complications may require
settlement rules, not discussed here, to handle special circumstances, for example, intrazonal congestion.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

224 PART 3 Market Architecture

can be handled in longer-term markets. The hallmark of this market is that the
system operator does not know the price of energy, only the price of transmission.
Market 3: A Power Pool. “Power pool” is a term used to describe an organiza-
tion of regulated utilities that trade power. Such pools did not use nodal pricing,
but tight pools did solve the unit-commitment problem centrally. Joskow (2000b)
has said of PJM,
. . . it is not very different from the power-pool dispatch and
operating mechanisms that were used when PJM was a large
traditional power pool relying on central economic dispatch
based on marginal cost pricing principles.
“Pool” and “poolco” were initially popular terms to describe this type of market,
but proposals with this name were tagged as communist by the California
bilateralists; hence those names were dropped in favor of “nodal pricing” and
various related terms.5 Unfortunately these terms refer to the transmission-pricing
half of the proposals and not to the unit-commitment part, so there is no common
term remaining for a market that mimics an old-fashioned tight power pool. Given
the circumstances, it seems best to continue using the old term, power pool, in
analogy to power exchange, with the understanding that in the new context of
market design, it refers to a market that does what a tight power pool did under
regulation. In the present work, “power pool” will mean an auction market that
uses side payments and multipart bids to solve the unit-commitment problem
centrally; the extent of locational pricing will not be implied. As a corollary, “nodal
pricing” will mean nodal pricing and will not imply centralized unit commitment
or side payments.

Definition Power Pool

A pool is a centralized market that uses “make-whole” side payments to, in effect,
pay different prices to different suppliers at the same time and location. These
payments are only made when an accepted supplier would lose money at its as-
bid costs given the pool price. A pool can implement full nodal pricing. It
typically uses multipart bids which cover all important aspects of a generator’s
operating costs and physical constraints, but a centralized market using two-part
bids and making side payments is a pool.

Market 4: PJM’s Day-Ahead Market (2001). Market 4 is modeled on PJM’s

current market and includes all of the types of bids allowed in the previous three
markets. This is the most complex market from the system operator’s point of view,
but, like market 3, it can be quite simple for suppliers if they simply bid competi-
tively. This requires only that they bid their costs.

5. The origins of the English pool can be traced in a memo from Larry Ruff to Stephen Littlechild in June
of 1989. It is available with other references at www.stoft.com. The theory of a power pool market is
described in Schweppe et al. (1988).

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-3 Day-Ahead Market Designs 225

Table 3-3.1 Notation Used in Auction Market Summaries

Term or Symbol Definition
PS(Q ), PD(Q ) Supply and demand bids. Supply bids must be nondecreasing.
QD Completely inelastic demand bid.
PT, QT Transmission bid.
P E , QE , Simple energy bid (no supply or demand curve allowed).
Q1 , Q1T Day-ahead accepted bid energy and transmission quantities.
P1 , P1T Day-ahead locational energy prices and transmission prices determined by the auction.
Q0 , Q0T Actual RT transactions.
P0 , P0T Real-time locational energy and transmission prices.
X, Y Two different locations. Used to define transmission bids and prices.
Uplift A charge to load to cover unattributed costs, in this case startup insurance.

Conventions and Notation for Describing Auctions

Bidding, Bid Acceptance, and Settlement. The market summaries describe

market operations in terms of the three auction stages. First, restrictions on bidding
are described that can have important consequences for the acceptance problem,
and for market efficiency. For example, allowing 24 hourly bids instead of a single
bid for the day can increase market power. Second, bid acceptance and price
determination are lumped together because they are computed together, although
conceptually they are distinct. Third, settlement includes penalties for noncompli-
ance with commitments made in the auction. These typically involve RT market
prices and are crucial to the functioning of the market.
Supply and Demand Curves. Usually supply curves are represented by either
piecewise linear functions (connect the dots with straight lines) or step functions.
Typically they allow the bidder to specify about ten sloped lines or horizontal steps,
but all that matters is that bidders can submit a fairly accurate approximation to
their actual supply and demand curves. Supply-curve bids must always be
nondecreasing.
Locational Prices. Every supply and demand bid at the same location is paid
or charged the same price. Transmission prices and quantities are each associated
with two locations. The following summaries all assume that demand prices are
the same as supply prices and vary by location, but for political reasons, markets
must usually charge loads a uniform price.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

226 PART 3 Market Architecture

Market 1: A Power Exchange

Bidding Restrictions:
Supply: PS(Q) 24 hourly bids
Demand: PD(Q) or QD 24 hourly bids

Bid Acceptance and Price Determination:

Acceptance problem: Maximize total surplus of accepted bids, {Q1 }
Constraints: Transmission limits
Price determination: P1 = marginal surplus at each location

Settlement Rules:
Supply: Pay: Q1 × P1 + (Q0 ! Q1 ) × P0
Demand: Charge: Q1 × P1 + (Q0 ! Q1 ) × P0

Notes: P1 is the DA price appropriate to the location of the accepted bid, Q1 . Q0 is the
quantity actually produced or consumed, and P0 is the RT locational price.

Comments:
Because generators cannot bid their startup costs, it is generally believed they
need to submit different price bids in different hours. Loads, whose usage is
largely unrelated to price, must do the same. The set {Q1 } represents the set of
accepted bid quantities, one for each supplier and each demander, a different
one in every hour. Acceptances may be for partial quantities.
The auction first finds the set of supply and demand bids which, if
accepted, would maximize total surplus to all market participants. Then market
price is determined at each location by the marginal surplus of additional
supply. This can be computed by making another kilowatt-hour available at no
cost, recomputing the optimal dispatch and finding the increase in total
surplus. That value is the price per kilowatt-hour assigned to that location. The
increase in value can come from either more consumption or from reduced
production costs. A kilowatt-hour is used to mimic a “marginal” change.
A DA market is a forward market, and the forward price holds if suppliers
deliver and customers take delivery of the DA quantity. Participants may not
make or take delivery of the exact quantities accepted, so strategy in the DA
auction depends on the penalty for not fulfilling the DA contracts. The NYISO,
for instance, confiscates the RT payments specified when Q0 ! Q1 > 0.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-3 Day-Ahead Market Designs 227

Market 2: A Transmission-Rights Market

(the “bilateral” approach)

Bidding Restrictions:
Demand: QT @ up to PT, from X to Y 24 hourly bids

Bid Acceptance and Price Determination:

Acceptance problem: Maximize total surplus = G(P T

× Q1T ) with 0 < Q1T < QT
Constraints: Transmission limits
Price determination: PT from X to Y = the marginal surplus of a 1-kW increase in
the transmission limit from X to Y

Settlement Rules:
Demand: Charge: Q1T × P1T + (Q0T ! Q1T ) × P0T

Notes: P1T is the DA price from X to Y. Q0T is the quantity actually produced or consumed, and
P0T is the RT price from X to Y.

Comments:
P1T is a price for transmission, not energy. If there are 10 locations there will be 90
pairs of locations and consequently 90 transmission prices. These can be computed
by subtracting pairs of energy prices. Adding the same constant to the energy
prices leaves the differences and thus the transmission prices unchanged.
Consequently the ten energy prices cannot be computed from the 90 transmission
prices.
The total surplus from the transmission sold in the auction is the sum of the
accepted quantities times the respective bid prices. Bid acceptance is required to
maximize total surplus. The price of transmission is set to the marginal surplus of
increasing transmission capability on the path in question. If the path is not
constrained the price is zero.
The price of transmission, PT, is the marginal surplus of increasing the
transmission limit from X to Y. This limit may be complicated, but an additional
free kilowatt injected at Y and withdrawn at X is always equivalent to raising that
limit by 1 kW.
Allowing fixed-quantity bids complicates the auction problem dramatically, so
it may be best not to allow them. If a partially accepted bid cannot be used, it could
be sold, or it could be returned to the system operator for resale in the RT market.
In this case the purchaser would earn Q1T × P0T from the RT market. On average,
arbitrage between the day ahead and RT markets should keep P0T close to P0T .

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

228 PART 3 Market Architecture

Market 3: A Power Pool

Bidding Restrictions:
Supply: PS(Q), startup cost, ramp-rate limit, etc. One bid per day
Demand: PD(Q) or QD 24 hourly bids

Bid Acceptance and Price Determination:

Acceptance problem: Maximize total surplus of accepted bids, {Q1 }
Constraints: Transmission limits, ramp-rate limits, etc.
Price determination: P1 = marginal surplus at each location
(Computed with accepted generators committed)

Settlement Rules:
Supply: Pay: Q1 × P1 + (Q0 ! Q1 ) × P0
Pay: Make-whole side payments for generators with
accepted bids.
Demand: Charge: Q1 × P1 + (Q0 ! Q1 ) × P0 + uplift

Notes: P1 is the DA price appropriate to the location of the accepted bid, Q1 . Q0 is the
quantity actually produced or consumed, and P0 is the RT locational price.

Comments:
In PJM, startup insurance is provided to generators who are scheduled to start
up in the DA market and who do start up and “follow PJM’s dispatch.”
Following dispatch amounts to starting up when directed to and keeping
output, Q, within 10% of the value that would make PS(Q) equal the RT price.
Startup insurance pays for the difference between as-bid costs and the
supplier’s revenue from DA and RT operations. As-bid costs include energy
costs, startup costs, and no-load costs.
Most generators that start up do not receive insurance payments as they
make enough short-run profits. The total cost of this insurance is less than 1%
of the cost of wholesale power. “Uplift” includes the cost of startup insurance
in this simplified market and several other charges in real markets.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-3 Day-Ahead Market Designs 229

Market 4: PJM’s Day-Ahead Market (2001)*

Bidding Restrictions:
Supply (w/UC): PS(Q), startup cost, ramp-rate limit, etc. One bid per day
Demand: PD(Q) or QD. 24 hourly bids
Transmission: QT @ up to PT, from X to Y 24 hourly bids
E E
Pure energy: Q @ up to P (supply or demand) 24 hourly bids

Bid Acceptance and Price Determination:

Acceptance problem: Maximize total surplus of accepted bids, {Q1 }, plus
the total surplus of transmission, G (PT × Q1T )
Constraints: Transmission limits, ramp-rate limits, etc.
Price determination: P1 = marginal surplus at each location
(Computed with accepted generators committed)
PT from X to Y = (P1 at Y) ! (P1 at X)

Settlement Rules:
Energy supply: Pay: Q1 × P1 + (Q0 ! Q1 ) × P0 +
Make-whole side payments for generators with
accepted bids
Energy demand: Charge: Q1 × P1 + (Q0 ! Q1 ) × P0 + uplift
Transmission demand: Charge: Q1T × P1T + (Q0T ! Q1T ) × P0T

Notes: See previous notes.

Comments:
“Ramp-rate limit” is meant as a proxy for this and various other constraints on the
operation of generators, such as minimum down time. Startup cost serves as a
proxy for other costs that are not captured in the supply function PS(Q), such as
no-load cost.
The pure energy bids, also called “virtual” bids because they can be made
without owning generation or load, are more restricted in format than the energy
bids made by actual load and generation.

* This description is still quite simplified as it leaves out PJM’s daily capacity
market, various other markets, and near-markets for ancillary services and the
accompanying uplift charges.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

230 PART 3 Market Architecture

3-3.3 OVERVIEW OF THE DAY-AHEAD DESIGN CONTROVERSY

The central debate in power market design continues to be between the advocates
of “centralized nodal pricing” and uncoordinated “bilateral” trading. It is generally
agreed that RT operation should be centralized and the forward markets beyond
a week should be bilateral and decentralized. At some point, as real time ap-
proaches, the market structure needs to switch from uncoordinated to coordinated,
and at the last second the market is almost completely replaced by the “command
and control” of engineers and automatic protective circuitry.
The controversy focuses most intensely on the DA market with one extreme
claiming it should be highly centralized and the other claiming it should be com-
pletely uncoordinated. Two compromise approaches received little attention until
recently. The market can be semicoordinated, or it can contain all possibilities and
allow the participants choose to their form of participation. PJM is evolving toward
the latter. England and California have compromised by alternating between the
extremes. As the English market rejects its centralized approach in favor of mini-
mum coordination, California is headed away from its uncoordinated approach
toward maximum centralization.6

Nodal Pricing with Central Unit Commitment

Nodal pricing refers to computing a different price at every node, or bus, of the
network. This system computes hundreds or thousands of different prices when
one or more of the system’s lines is “congested” (meaning they are fully used, and
it would be beneficial if they could carry more power). The point is not the many
prices as there would be as many in a competitive bilateral market. The point is
these prices are computed by an auctioneer.
The centralized-nodal view encompasses a separate and unrelated tenant. It
advocates a centralized solution to the unit-commitment problem. More recently
this has been advocated as an option instead of a requirement. Centralized unit
commitment, but not nodal pricing, requires complex bids that describe in detail
the costs and limits of generators. In a power exchange, even one that uses nodal
pricing, the auctioneer has no detailed information about generators and makes
no attempt to solve this problem. The four summarized markets allow comparison
of the unit commitment part of this view but not its nodal pricing component.

The Bilateral Approach

The bilateral view generally admits that a centralized DA market is needed but
contends that it should not trade energy. Instead it should facilitate bilateral trades
of energy by selling transmission rights to the traders. This is generally claimed
to be a simplification, and for the auctioneer it is a great simplification relative to
solving the unit-commitment problem. But it is no simpler than a power exchange,

6. In May 2001 the CA ISO proposed a multi-day centralized unit-commitment pool with the claim that
it intends to adopt nodal pricing as soon as practical. This was rejected by FERC. In January 2002, the CA
ISO is again contemplating nodal pricing with centralized unit commitment.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-3 Day-Ahead Market Designs 231

and it makes the job of traders far more complex. They must make energy trades
with dozens of parties in different locations with different prices, instead of just
trading with the exchange at their local price. But the point of the bilateral approach
is not to simplify life for traders; it is to minimize the role of the system operator
and to maximize the role of traders.

The Role of the Four Markets in the Controversy

Market 1: A Power Exchange. The pure energy market is a semi-coordinated
market. It compromises by taking a middle path and is simplest in operation for
both the exchange and for traders. It prices energy nodally but ignores the unit-
commitment problem. This complicates the supplier’s bid strategy which must
indirectly account for startup costs and generator limitations, but the market itself
remains simple. A slight complication in power exchange bidding can help genera-
tors solve the unit commitment problem.
Market 2: A Transmission-Rights Market. The transmission-rights market
represents a pure bilateral approach. This approach maximizes the role of traders
and forces traders and generators to solve the unit-commitment problem on their
own. The job of the system operator is no simpler than in a power exchange because
it must still account for all transmission constraints and arrive at what is equivalent
to a full set of nodal energy prices except that their collective level is indeterminate.
Only the differences between locational energy prices are known to the system
operator.
Market 3: A Power Pool. Generators report as many details of their costs and
limitations as the auctioneer’s computer program can handle. The auctioneer then
computes which generators should be started ahead of time. It finds the optimal
solution, assuming that the bid data and demand forecasts are correct. Because these
programs do not find market clearing prices, and the prices they do find are too
low to induce a few of the needed generators to start up; these generators must
receive “side payments.” Once these are guaranteed, the energy prices clear the
market.
Market 4: PJM’s DA Market (2001). PJM’s market combines all the bid types
found in the previous three markets. Together all of these bids determine one set
of energy prices, and the differences between them determine the transmission
prices. Current practice is for almost all generators in PJM to use the unit-commit-
ment bids for bilateral traders within PJM to use the pure energy bids, and for
bilateral traders who want to send power through PJM to use the transmission bids.
One argument for this approach is that generators should be allowed to let the
system operator solve their unit-commitment problem if that is their preference,
but this does not answer the more important question of whether the system operator
should offer startup insurance to generators.
It should not be concluded that the use of various types of bids in PJM indicate
a clear endorsement of one or another type of DA market by any of the market
participants. Besides the basic rules, there are still many hard-to-quantify pressures
that influence the choices of market participants. Nonetheless, this experiment may
someday provide a useful comparison of the various approaches.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 Chapter 3-4
Ancillary Services

Electric energy thus is the most useful form of energy—and at the same time it is the most
useless. [It] is never used as such . . . [but] is the intermediary.

Charles Proteus Steinmetz

Harper’s
January 1922

P OWER IS THE PRIMARY SERVICE, BUT SIX ANCILLARY SERVICES ARE

NEEDED TO ENSURE RELIABLE, HIGH-QUALITY POWER, EFFICIENTLY
PRODUCED. Usually ancillary services are defined by how they are provided rather
than by the service rendered. This results in a plethora of services and little insight
into their relationship to market design. Defining services by the benefit they
provide and defining them broadly produces a short but comprehensive list. Services
directed at long-term investment are not counted as ancillary to real-time (RT)
power delivery. The six listed services require planning by the system operator,
but economic dispatch is jointly provided by the system operator and the market.
How that task is shared should be the subject of intense debate.
Chapter Summary 3-4 : Of the six ancillary services, the system operator or
its agent must directly provide transmission security and trade enforcement, and
to some extent economic dispatch. The other services, balancing, voltage stability,
and black-start capability can be purchased from a competitive market, but the
system operator must demand and pay for these services.
Section 1: The List of Ancillary Services. Services are defined by the benefit
they provide to the market and its participants, not by their method of provision.
An accurate frequency is required by some motors and particularly by large genera-
tors. Many appliances need a fairly accurate voltage, and together these two services
define the provision of power from the customer’s perspective. Transmission
security and occasionally black-start capability are indirect services needed by the
market to provide the first two. Economic dispatch can include the solution of the
unit commitment problem and often includes efficiently dispatching around conges-
tion constraints. Trade enforcement is required to provide property rights essential
for bilateral trading.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-4 Ancillary Services 233

Section 2: Balancing and Frequency Stability. Frequency is determined

by the supply–demand balance. Frequency stability cannot be provided by individual
system operators, but each is required to balance the real power flows in its control
area taking into account a frequency correction. When every system provides this
balancing service, they collectively provide frequency stability.
Section 3: Voltage Stability. Voltage support is provided passively by capaci-
tor banks and actively by generators. Provision is by the supply of “reactive” power
which is difficult to transmit. This makes it difficult to purchase under competitive
conditions, but long-term contracts that permit competition through entry should
be helpful.
Section 4: Transmission Security. Transmission security can be provided
initially through the control of transmission rights or the operation of a day-ahead
(DA) energy market with locational prices. In real time, a locational balancing
market should be used.
Section 5: Economic Dispatch. The system operator must assist in the provi-
sion of this service by conducting a balancing market and may assist by conducting
a DA power exchange or by providing almost the entire service through a power
pool.
Section 6: Trade Enforcement. Once power is injected into the grid it cannot
be tracked, so ownership is lost. This makes bilateral trading impossible unless
a substitute property right is defined and enforced by the system operator, or in
the case of inter-control-area trades, by a higher authority. This property-right
system requires the metering of all traders, the recording of all trades, and some
power of enforcement to deter or compensate for discrepancies between registered
trades and actual power flows.

3-4.1 THE LIST OF ANCILLARY SERVICES

There are many lists of ancillary services which differ mainly in how they split and
combine the categories of service.1 The present list combines services whenever
they differ only by speed of delivery or method of provision. The purpose is to make
the shortest possible comprehensive list that does not combine grossly dissimilar
services.
The list is organized very roughly by time scale, starting with real-power
balancing, which is of constant concern and sometimes must be provided on the

1. As reported in Order 888, when FERC (1996a) asked for help in defining ancillary services it received
over a dozen different lists. One list had 38 ancillary services just for transmission. NERC presented a list
of 12, though it adopted FERC’s final list of six in its Glossary of Terms (NERC 1996). These six are
covered by, but do not directly correspond to, the six listed in this chapter.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

234 PART 3 Market Architecture

shortest possible time-scale, and ending with the black-start service which is seldom
used.
Ancillary services are those traditionally provided by the system operator in
support of the basic service of generating real power and injecting it into the grid.
Much more is needed to ensure that the supply of delivered power is reliable and
of high quality. Some of these services are indirect, but all ancillary services are
concerned with the dispatch, trade, and delivery of power. Looking after investment
in generation (discussed in Part 2) and transmission, while they may involve the
system operator, are not considered ancillary services. Because the role of the
system operator is at the heart of most architectural controversies, understanding
the ancillary services is crucial to evaluating any proposed market designs.

Defining the Ancillary Services

This section defines the ancillary services not by how they are provided but by the
benefits they provide. The following sections describe the provision of these
services.
Frequency and Voltage. The two fundamental characteristics of power delivered
to a customer are frequency and voltage. As long as these remain correct the
customer will have access to the needed power, and it will have the required
characteristics.2 Whenever a customer encounters trouble with the supply of power,
either voltage or frequency will have deviated from its allowed range. In the
extreme, a power outage is defined by zero voltage.
Frequency, measured in Hertz (Hz), is the number of times per second that
the voltage goes from its maximum (positive) value to its minimum (negative) value
and back to its maximum. This is called a “cycle” and frequency was formerly
measured in cycles per second (cps). (When power is used, the frequency of the
current flow is the same as that of voltage, and its alternating direction of flow
explains the term, AC, alternating current.) Frequency affects the operation of
motors, many of which draw more power and run faster at a higher frequency.
Synchronous motors, the type used in AC clocks and phonographs, run at a speed
directly proportional to frequency. But the greatest service of a stable frequency
is provided to large generators—basically synchronous motors operated with a
reverse purpose—which suffer less stress when run at a constant speed.
A 10% increase in voltage will cut the life of an incandescent light bulb approxi-
mately in half. Voltage is electrical pressure, and the more pressure the more
electrical current is forced through appliances. Some are built to be insensitive to
voltage so they can be used both on 110 volt AC in North America and 220 volt
AC in Europe. Low voltage can cause appliances to underperform and, oddly, can
burn out some motors.
Transmission Security. Compared with other networks, the transmission system
is quite fragile.3 Overuse can cause lines to overheat which can cause them to sag

2. The issue of power quality, which includes wave shape and occasional voltage spikes, will be ignored.
3. Too much traffic at one time does not damage a highway, and too many phone calls will not hurt the
phone wires.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-4 Ancillary Services 235

permanently or even to melt. It can also cause complex electrical problems that
interfere with power flow. Because of such problems, most high-voltage power
lines have automatic protective circuitry that can take them out of service almost
instantly for their own protection. But protecting one line can endanger another
and can cut off service to customers. The ancillary service of transmission security
keeps the grid operating.
In other lists of ancillary services, this service is buried within the scheduling
and dispatch services. Those are the methods of providing transmission security,
but they also provide other services.
Voltage support is needed both as a direct customer service and to provide
transmission security. This complicates its classification, but within the present
classification of ancillary services, reactive power supply is viewed as helping to
provide two distinct services: voltage stability for customers, and transmission
security. (In the case of voltage support, the method of provision is as useful a
method of classification as the benefit of the service, but consistency dictates
classification by benefit.)
Economic Dispatch. Economic dispatch refers to using the right generators in
the right amounts at the right times in order to minimize the total cost of production.
Like transmission security, this service is provided by what are elsewhere
referred to as scheduling and dispatch services. From an economic viewpoint,
economic dispatch and transmission security are only minimally related. Economic
dispatch often has nothing to do with power lines as no line limits are binding, while
transmission security has nothing to do with the marginal cost of generators. Rolling
both into scheduling and dispatch makes some sense in a regulated world but is
of little help when designing a market.
Trading Enforcement. Trading enforcement is an expanded version of part (b)
of NERC’s sixth ancillary service (NERC, 1996) which is its interpretation of the
FERC’s first ancillary service required by Order 888 (FERC, 1996a):
Scheduling, System Control, and Dispatch Service — Provides
for a) scheduling, b) confirming and implementing an inter-
change schedule with other Control Areas, including intermedi-
ary Control Areas providing transmission service, and c) ensur-
ing operational security during the interchange transaction.
[Emphasis added]
An expansion of this definition is required because trade is no longer just between
control areas but also between private parties within a control area.
Although this service also requires scheduling, its purpose is different from
either economic dispatch or the provision of transmission security. This is an
accounting and enforcement service. Power trades are inherently insecure as origin
and ownership cannot be identified once power has entered the grid. Trading
requires property rights which become problematic when ownership cannot be
established. As a substitute for ownership identification, all injections and withdraw-
als from the grid are metered and recorded by system operators. They also record
all trades. By matching trades with physical inputs and outputs, the honesty of the
traders can be established, and deviations from recorded trading intentions can be

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

236 PART 3 Market Architecture

dealt with appropriately. Trading enforcement comprises three steps: (1) recording
of trades, (2) metering of power flows between all traders and the common grid,
and (3) settling accounts with traders who deviate.
Note that this is not physical enforcement of RT bilateral trading whose absence
constitutes the second demand-side flaw (see Section 1-1.5). This enforcement
occurs much later, but if it were implemented with extremely high penalties (not
recommended) it would have almost the same effect as RT enforcement.
The Black-Start Capability. In the worst system failure, either a large part of
or an entire interconnection could be shut down. Unfortunately most generators
must be “plugged in” to get started; they require power from the grid in order to
start producing power. To start the grid, generators possessing black-start capability
are needed. Starting a power grid requires the final ancillary service, the black-start
service.
Overview of the List of Services. This completes
Ancillary Services the list of basic tasks that must be carried out in whole
1. Real-power balancing (frequency stability) or in part by the system operator (or its agents) to
2. Voltage stability (for customers) support the basic service of RT power production.
3. Transmission security Frequency and voltage stability are the only two ser-
4. Economic dispatch vices provided directly to customers. Transmission
5. Financial trade enforcement security (and on rare occasions the black-start service)
6. Black start are provided to the wholesale market to aid in fre-
quency and voltage stability. They are named sepa-
rately because their role is indirect. Economic dispatch is a service provided to
generators, not to end users. Most services must be paid for—this one simply
reduces the price consumers pay for power. Trade enforcement is a service to
bilateral traders. Without it bilateral trade would be impossible.

3-4.2 REAL-POWER BALANCING AND FREQUENCY STABILITY

Frequency stability and power balancing are two sides of the same service. The
end-use service is a stable, accurate frequency, but no system operator can provide
that because it is mainly controlled by the other systems in the interconnection.
Instead all system operators are required to balance their power inflows and
outflows according to NERC’s ACE formula. When all system operators provide
this balancing service to the interconnection, the result is a stable and accurate
frequency. The actual service provided by the system operator is balancing, but
the service provided to customers as a result of collective balancing efforts is
frequency stability. “Real” power is ordinary electrical power, but it is given its
full name here to distinguish it from “reactive” power which is used to provide
the voltage stability service.
Balancing is provided by a number of different methods which are normally
classified as separate ancillary services. Four of the six ancillary services listed
by NERC and FERC fall into this category:

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-4 Ancillary Services 237

1. Regulation and Frequency Response Service

2. Energy Imbalance Service
3. Operating Reserve: Spinning Reserve Service
4. Operating Reserve: Supplemental Reserve Service
Some system operators further classify supplemental reserves into 10-minute, 30-
minute, and replacement reserves. Also direct-control load management and
interruptible demand, as defined by NERC, would be classified as spinning reserve
by some jurisdictions but might legitimately be separated by others.
Frequency instability is caused by a mismatch of supply and demand, as is
explained in Section 1-4.3. When total supply in an interconnection is greater than
demand the frequency is too high, and when it is less than demand frequency is
too low.4 The comparison must be made for the entire interconnection because
frequency is uniform throughout this region as all generators are synchronized.
Within an interconnection, one market cannot have a high frequency and another
a low frequency.
Supply and demand fall out of balance for three basic reasons. First, load varies
in a predictable pattern throughout the day. This requires active “load following”
on the part of the system operator and the generators it controls. Second, there are
unpredictable, constant small fluctuations in most loads and some generators, such
as wind-power generators. Third, there are generator and line outages.
Corresponding to these are three basic approaches to keeping the system
balanced. Small random fluctuations are handled by a service called “regulation,”
which utilizes generators that receive a control signal directly from the system
operator. The process is called automatic generation control, AGC. Predictable
daily fluctuations, like “the morning ramp,” are handled by scheduling generation
and, in an unregulated system, by the balancing market. Unexpected generation
and transmission outages are handled by operating reserves of various types starting
with 10-minute spinning reserves.
The Market’s Role in Provision. Regulation and operating reserves are
generally purchased by the system operator in a market that it organizes and in
which it is the sole source of demand. Sometimes there is an attempt to have
generators “self-provide” operating reserves as if they might have their own need
for such reserves. This makes little sense except as a means of preventing the system
operator from scrutinizing the quality of the reserves provided or of capturing the
service of administering the market for operating reserves.
Tracking the more predictable and larger-scale fluctuations in demand can best
be handled by a DA market, an hour-ahead market, and an RT market. The architec-
ture of this group of markets is the subject of much of Part 3.

4. The supply-demand balance for a given control area within the interconnection is defined by ACE,
which measures a combination of frequency and inadvertent interchange (see Section 1-4.3). This
ancillary service is more properly defined as controlling ACE, but this translates into balancing supply and
demand locally and results in controlling frequency globally.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

238 PART 3 Market Architecture

3-4.3 VOLTAGE STABILITY FOR CUSTOMERS

Voltage tends to drop as power flows from generator to load, and the more power
that flows the more the voltage tends to drop. Voltage drop also depends on the
type of load. Those involving electromagnetism, like motors and some fluorescent-
light ballasts, tend to use a lot of “reactive” power which causes the voltage to drop.
If this voltage drop is not corrected, this affects all other loads in the vicinity.
Usually the voltage drop on the transmission line is corrected by a transformer that
automatically adjusts the voltage before the power is distributed, but this process
is not perfect and customer voltage fluctuates as a consequence.
Voltage drop can be counteracted by the injection of reactive power (explained
in Part 5). Reactive power is quite different from real power except that its descrip-
tion shares very similar mathematics. For instance, it can be supplied by capacitors
which are entirely passive and consume no fuel. It can also be supplied by genera-
tors, generally at very low cost. The cost to a generator of supplying reactive power
is mainly an opportunity cost because, while supplying reactive power consumes
essentially no fuel, it does reduce the generator’s ability to produce real power.
This relationship is nonlinear. Synchronous condensers are another source of
reactive power. These are essential generators that are run electrically. They take
real power from the grid and return reactive power.
The stabilization of voltage for consumers interacts in a complex way with
voltage support for transmission security. Consequently, the customer service and
the voltage-support part of the transmission service should be considered together.
However, this book will not examine either service in detail.
The Market’s Role in Provision. Some have proposed setting up a market in
reactive power in parallel with the market in real power. The price of reactive power
would be adjusted to clear the reactive power market.
Because of the externalities associated with the consumption of reactive power,
a bilateral market is out of the question. Users of reactive power would purchase
far too little. Unfortunately, reactive power is difficult to transmit. In effect it suffers
losses that are very roughly ten times greater than the losses of real power. As a
result, an RT reactive power exchange would have severe difficulties with market
power. The system operator should, however, be able to purchase long-term
contracts for the provision of reactive power provided time is allowed for competi-
tive entry. In other words, market power may be exercised over contracts sold before
competitors could enter. In the mean time, the system operator may need to exercise
monopsony power or regulate reactive power in order to secure it at a reasonable
price in the near term.

3-4.4 TRANSMISSION SECURITY

Transmission line limits are a complex matter. First there is a thermal limit. If the
total real and reactive power on a line exceeds a certain threshold for too many
minutes the line will be permanently damaged. The most likely cause of a line

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-4 Ancillary Services 239

exceeding its thermal limit is another line going out of service unexpectedly.5 Then
the power flow on the system re-routes itself instantly, with the most flow being
transferred to the nearest lines. In the (almost) worst case, if there are two identical
parallel lines between two regions, and these are the only lines between the regions,
if one goes out of service the other receives all of its power flow. If each line could
handle 100 MW, then to protect each of them, their combined security limit would
be 100 MW. The only benefit of having two lines instead of one would be the
increased reliability.
When security limits are computed accurately, they change frequently. For
example, suppose line L1 and L2 both have a 100 MW thermal limit and L1 would
dump 50% of its power onto line L2 if L1 were to go out of service. When L1 has
100 MW flowing on it, the security limit on line L2 is its 100-MW thermal limit
minus the 50 MW of potential spillover from an outage on L1. The security limit
of L1 is then 50 MW. If, however, L1 is carrying only 20 MW, then L2 has a
security limit of 90 MW.
Besides thermal limits there are other basic physical limits that may serve as
the basis for security limits. These are called voltage limits and stability limits. Both
are less stable than thermal limits which vary only a small amount with ambient
temperature.
To provide transmission security, the system operator must first compute these
limits (and recompute some of them frequently) and then ensure that the dispatch
of generation, given the existing load, is a security constrained dispatch. This
is a set of generation output levels which, given the load, does not cause more power
to flow on any line than its security limit allows.
The Market’s Role in Provision. Transmission security is a service that must
be provided by the system operator. It can be provided by selling transmission rights
or by controlling the acceptance of energy bids in a DA energy auction. Neither
method is foolproof as both operate at some interval before real time. In the mean
time, generators may go out of service, lines may go out, loads may increase or
decrease. Any of these can change the flow of power in such a way that a security
limit is violated. Fortunately, in most systems, no security limit is threatened in
most hours. In PJM, for example, no security limit is binding about 3'4 of the time.
When a security limit is breached in real time, the system operator needs to
induce more generator in one location and less in another. The problem is not total
production; it is only a matter of where production occurs. One method of dealing
with the problem is to look for a bilateral trade that is causing power flow on the
path with the security violation and disallow that trade. The load will then have
to find another source of power. This is the current NERC approach to line-loading
relief. It is a centralized, nonmarket approach and, as would be expected, is quite
inefficient.
NERC is moving toward an RT balancing market. Essentially this amounts to
lowering the RT price of power in the region where less generation is wanted and
raising the RT price of power in the region where more is wanted. This is a service

5. It is the square root of the sum of the squares of real and reactive power that is limited by the thermal
limit, but in most cases this is quite similar to a simple limit on real power.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

240 PART 3 Market Architecture

that should be provided by using a market, but the market is only the tool used by
the system operator. The demand for security originates with the system operator,
and the system operator ensures its provision. The core of the service is a highly
complex coordination problem whose solution must be centrally provided.

3-4.5 ECONOMIC DISPATCH

Economic dispatch, as defined above, is a matter of using the cheapest generation.
There are two parts to this problem: (1) deciding which generators to start up, the
unit-commitment problem; and (2) deciding how much to use each generator that
is running. However, most generators that are started are used at full output except
when they are in transition or when one is avoiding the cost of a night-time shut-
down and restart by staying on at a low output level over night.
If the RT dispatch is economic, the DA market doesn’t matter. Because RT
supply and demand conditions are not known precisely until real time, RT economic
dispatch is quite important. Because many generators are slow and expensive to
start, the DA market plays an important role in making an efficient RT dispatch
possible.
The Market’s Role in Provision. There are three basic approaches to economic
dispatch: (1) bilateral trading, (2) a centralized DA power exchange, and (3) a
centralized DA power pool. In the first approach, suppliers and power traders solve
the entire problem. In the second, the system operator helps by running a DA
exchange. The exchange provides a public price which is either used by the ex-
change to select the generators that should start up or could be used by suppliers
to self-select during several rounds of bidding. In the third approach, the power
pool would require generators to submit complex bids and the system operator
would optimize the dispatch using all available bid information. That optimization
would be used to control the actual dispatch through prices and start-up insurance.
In this approach the system operator provides a full unit-commitment service. In
all approaches, the suppliers and the market price jointly determine a dispatch,
which if the market is competitive, will be very nearly least cost.
If the system operator relies on a competitive balancing market, the market will
provide an economic dispatch. If the system operator chooses such tools as pro
rata curtailment or bilateral curtailment for line-loading relief, they will interfere
with economic dispatch. There is a debate over how much efficiency can be
improved by RT centralized unit commitment.

3-4.6 TRADE ENFORCEMENT

When control area A sells power across control areas B and C to control area D,
how does anyone know whether A really provided the power it sold? Area D cannot
measure A’s output and does not really care whether A produced the power or not
because it can take power out of the grid in any case. The more it uses, the more
flows into its area. If the power was not produced, then frequency will decline and

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-4 Ancillary Services 241

all control areas in the interconnect will be responsible for increasing supply to
correct the problem.
The origin of this problem is the fact that power cannot be tracked. Once a
generator injects power it cannot keep track of it, so it loses ownership of that
particular “piece” of power. The solution to the problem is to give the supplier of
the power a right to withdraw the same amount of power and let the supplier
reassign that right at will, then enforce the rights to withdraw power. For example,
when area A sells power to area D, it does so by injecting power and thereby
obtaining the right to withdraw power which it then assigns to D. This solves the
problem of tracking where the power actually flows, but it requires that injections
and withdrawals be metered and compared with the rights to withdraw power that
are granted and exercised.
This is an abstract description of the principles on which the actual accounting
system is based. In reality flows are measured between adjoining control areas,
and each control area has a net scheduled interchange that is publically known.
Interchange is flow on the lines connecting control areas (interties), and every
interconnecting line is metered by both control areas. Each gets credit for outflow
so they have opposite interests in terms of reporting the flow on the line. One would
like to err in one direction, and one would prefer to err in the other direction. This
prevents collusion and keeps both honest.
When A sells power to D, this is publically recorded as an increase in A’s
scheduled net interchange and a decrease in D’s scheduled net interchange. The
schedules of B and C do not change because any power that flows into them as
a result of this trade also flows out. With this system of recording transactions, plus
the metering of the interties, bilateral trades between control areas can be enforced.
This system creates the appropriate set of property rights.
An analogous system is needed for private bilateral trades within or between
control areas. For simplicity consider internal trades. If no one were watching, A
could sell power to B within the control area and then fail to produce. B would take
the power, and no one would be the wiser. The power taken would be power
produced by some other generator somewhere in the interconnection, but there is
no physical way to check on who produced the power. Again a system of trade
enforcement is needed. No matter how the details of such a system are arranged
it requires at least the following three conditions:
1. All trades must be registered with the trade enforcement authority.
2. All traders must have their connection to the grid metered by the authority.
3. The trading authority must have the power to charge traders for discrepan-
cies.

Result 3-4.1 Strictly Bilateral Power Trading Requires Centralized Coordination

Without the central registration and measurement of all bilateral trades and
penalties for discrepancies, bilateral power trading would be impossible.

Power Losses. One service which is frequently labeled an ancillary service, but
has not been mentioned so far, is the replacement of power losses. When the system
operator balances the system, it necessarily computes a power flow from which

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

242 PART 3 Market Architecture

losses can be approximated. Any discrepancy between actual losses and the
computed value that is included as part of demand will be reflected in the control-
area’s ACE which the system operator is obliged to correct. In fact if losses are
not computed, ACE measurements could be relied on as a method of determining
losses.
Thus losses are just another demand on the system and are handled through
normal balancing. The only additional service required of the system operator is
to assign the cost of losses to users. The simplest approach is to charge loads for
them in the uplift. An economic approach, discussed in Chapters 5-7 and 5-8 is
to compute locational loss prices and charge generators. This calculation and its
inclusion in billing are the real services associated with losses, but these are simply
part of normal economic dispatch.
The Market’s Role in Provision of Trade Enforcement. Trade enforcement
for bilateral trading must be provided by the system operator. Individual traders
have no incentive to provide this service for their own trades and no power to
provide it for other trades. No other private party has the power to enforce trades.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 Chapter 3-5
The Day-Ahead Market in Theory

Economic questions involve thousands of complicated factors which contribute to a certain result.
It takes a lot of brain power and a lot of scientific data to solve these questions.

Thomas Edison
1914

T HE DAY-AHEAD MARKET IS THE FORWARD MARKET WITH THE

GREATEST PHYSICAL IMPLICATIONS. By providing financial certainty, it can
remove the risk of incurring startup expenses. The more efficient the market, the
more accurate the startup decisions and the lower the cost of power. Even without
the unit-commitment problem, reducing financial risk would reduce the cost of
capital.
As explained in Chapter 3-3, the day-ahead (DA) market can utilize one of three
basic architectures or a combination. Bilateral markets, exchanges and pools can
each provide hedging and unit commitment. The controversy over the choice of
architecture is driven by concerns over the shortcomings of private markets and
nonprofit system operators in performing the coordination functions associated
with unit commitment. Hedging is also an issue as pools claim to provide it more
completely than exchanges.
Some theory of market clearing—when it is possible and when not—helps to
provide a framework for evaluating the various designs. “Nonconvex” production
costs are the key to this theory, and while conceptually arcane, the focus of current
controversy and volumes of market rules attest to their impact on market design.
Chapter Summary 3-5 : Nonconvex generation costs violate an assumption
of perfect competition, but the magnitude of the resulting problems is unknown.
The pool approach is designed to minimize these problems, but its pricing ignores
investment incentives. The bilateral approach faces formidable coordination
problems in the DA market. It may be less efficient and provide less reliability than
a centralized approach. Side-payments made by a DA pool do not increase
reliability.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

244 PART 3 Market Architecture

Section 1: Equilibrium Without a Clearing Price. Startup and no-load costs

are “nonconvex” and violate the assumptions of perfect competition. In particular,
they often prevent the existence of a market-clearing price. Nonetheless, bilateral
markets, exchanges, and pools all have equilibria and equilibrium prices. Bilateral
and exchange prices approximate competitive prices as best they can. Consequently,
they are approximately right for short-run supply and investment. Pool prices are
optimized for short-run supply but ignore investment incentives.
Section 2: Difficulties with Bilateral Day-Ahead Markets. In a bilateral
market, generators must self commit and, in doing so, they do not consider reliabil-
ity. When DA load can be served at almost the same price by quick-start or slow-
start generators, the market may select too many quick-start generators just by
chance. Then the slow-start generators will be unavailable in real time. The uncer-
tainties of the bilateral process cause decreased reliability. Bilateral markets also
have more difficulty developing locational prices and can be expected to be less
efficient in markets with significant congestion. The magnitudes of these effects
are unknown.
Section 3: Settlement, Hedging, and Reliability. Selling power in the DA
market provides a hedge against the risks of the volatile real-time (RT) price. For
the hedge to work, the supplier must commit its generator and be prepared to
produce according to its DA contract. Thus DA contracts provide some inducement
to commit units as a way of reducing risk. If the DA contract is with a pool and
therefore includes a “make-whole” side payment, this leads to no increase in the
incentive to commit because that payment is not contingent on RT performance.
Section 4: Other Design Considerations. These include transaction costs,
facilitation of market monitoring, provision of publicly known prices, and nondis-
criminatory access. Penalties for avoiding the DA market should not be used in
an attempt to obtain more accurate unit-commitment data.

3-5.1 EQUILIBRIUM WITHOUT A CLEARING PRICE

The theory of perfect competition, explained in Parts 1 and 2, assumes that there
is a price that “clears the market” when suppliers take that price as given. In this
case a competitive market is efficient, but if production costs are nonconvex, there
may be no market clearing price, and when there is none, economic theory does
not guarantee efficiency. Startup costs and no-load costs are “nonconvex,” and
some constraints on generation output cause similar problems.
Although these nonconvex costs cause deviations from perfect competition,
the deviations may be inconsequential. The debate over the power pool approach
concerns the magnitude of the resulting inefficiency. This inefficiency can take

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-5 The Day-Ahead Market in Theory 245

three forms. Nonconvex costs can cause (1) an inefficient but reliable dispatch,
(2) decreased reliability, or (3) financial risks for generators. The justification
claimed for the complexity of the pool approach is that it minimizes these three
inefficiencies.

Why There May Be No Clearing Price

A market clearing price is a single price that causes supply to equal
demand. Suppose power is needed for just two hours and exactly 100
MW is needed. Suppose there are a number of competing generators,
so they offer to supply at a competitive price. Their marginal cost is
$20/MWh up to their capacity of 200 MW, and their startup costs are
$30/MW started. Ignoring demand limitations, at a price of $35/MWh
a generator can earn a scarcity rent of $15/MWh for two hours on each
MW of its capacity for a total scarcity rent of $30/MW.1 This would
exactly cover its start up cost. At a lower price it would not voluntarily
agree to sell any power. At a price of $35/MWh or more it would offer
to supply its full 200 MW.
Below $35/MWh supply is zero, while at or above $35 supply is 200
MW or greater. There is no price at which supply equals 100 MW.
Technically, there is no price at which a supplier could profitably sell 100 MW
but could not increase its profit by selling more. Thus there is no price at which
supply equals the 100 MW level of demand. There is no market clearing price. (See
Section 3-9.2 for another example.) The root of the problem is that production costs
are not convex. Convex production costs have the property that twice as much
output always costs at least twice as much to produce. In this case 100 MW for
two hours would cost $50/MWh, while 200 MW for two hours would cost
$35/MWh. Twice as much is cheaper per unit, so the production cost function is
not convex.

A Bilateral Equilibrium
Market clearing is often defined narrowly in the context of perfect competition,
but sometimes it is defined more broadly to mean that no profitable trades remain
to be made. To avoid confusion this situation will be referred to as an equilibrium
but not as market clearing. This definition will be used with bilateral markets. In
the above example, a bilateral market could reach an equilibrium by arranging a
trade in which 100 MW of power was sold for two hours at a price of $35/MWh.
Having made that trade, demand would be satisfied and no further trade could
profitably take place. In this context “profitably” takes account of the consumer’s
value as well as the supplier’s profit.
Because of bargaining problems, it is not obvious that a bilateral market would
reach this equilibrium. In a more complex and realistic market, the frictions of
bilateral trade inherent in the costly process of gathering information and arranging

1. Scarcity rent (see Section 1-6.6) is revenue minus costs that vary with output. After startup and no-load
costs are subtracted, short-run profits remain.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

246 PART 3 Market Architecture

trades could also prevent the market from reaching equilibrium, especially given
the limited trading time of a DA market.
Even if equilibrium is reached there is no guarantee that it will be efficient (least
cost) in a market with nonconvex costs. But even if inefficient, in a large market,
this inefficiency may prove to be too small to matter. The problems caused by
nonconvex costs in a real power market with a peak load of 10 GW and a 1 GW
interconnection to a larger outside market may be entirely negligible. This is a
matter for empirical research or detailed theoretical calculations.

An Exchange Equilibrium
A bilateral market can reach an equilibrium by trading at more than one price, but
an exchange is limited to a single price. With nonconvex costs, there may be no
price that clears an exchange market. In spite of this the exchange will always have
an equilibrium. Exchanges are auctions and, as described in Chapter 3, have a
definite set of rules that determine what bids will be accepted and how they will
be settled. The Nash theorem states that such games always have at least one Nash
equilibrium.2
An equilibrium of an exchange market will be defined to be a Nash equilibrium,
a situation in which no player could do better by bidding differently, provided other
players maintain their equilibrium bids. Using this definition, the Nash theorem
guarantees the market will have at least one equilibrium. Examples of this are given
in Chapter 9.
Although the exchange’s equilibrium price cannot clear the market, it will tend
to come as close as possible in the sense of minimizing the gap between competitive
supply and competitive demand. For example, the market may clear except for a
single generator that sells less than its full output while wishing it could sell it all
at the market price.

A Pool Equilibrium
Like an exchange, the equilibrium of a power pool is defined as any Nash equilib-
rium of the pool auction. Like an exchange and a bilateral market, a pool determines
an equilibrium set of quantities traded. Unlike an exchange, a pool does not deter-
mine a single market price. Instead, it determines a nominal market price, the pool
price, and a set of supplier-specific side payments which, in effect, create a different
price for each supplier. In practice many suppliers will receive no side payment
and so this group will be paid the same price, the nominal market price. But, by
design, the pool allows every supplier to be paid a different price.
The average of the individual prices in a pool approximates the price in an
exchange, but the nominal pool price is lower and could be much lower. In practice
it appears to be lower by (very roughly) 3%. As an example consider a market with
generators having marginal costs of $20.00/MWh, $20.10/MWh, $20.20/MWh
2. Auction rules specify what bids are acceptable including the fact that prices must be rounded to the
penny and quantities to the MW. These rules and limitations imply the auction is a finite game (players
have a finite though large number of strategies available). In solving such games, it appears that the
finiteness of the strategy set is not required to produce a Nash equilibrium.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-5 The Day-Ahead Market in Theory 247

and so on. Assume they have no startup costs but have a no-load cost that averages
$10/MWh when they produce at full capacity. Assume the generators are small
relative to the market size so that the integer (lumpiness) problem is of little concern.
A pool price is defined by system marginal cost which is $25/MWh if the 51st
generator is needed. This generator would receive a side payment of $10/MWh
to cover its no-load cost. An exchange would need to set a single market price of
$35/MWh to induce this supplier to produce. Thus the nominal pool price can be
dramatically lower than the price that comes closest to clearing the market and
dramatically lower than the market price of a power exchange. The difference is
mainly, but not entirely, made up by the side payments.

The Efficiency Of Equilibrium Prices

Bilateral prices and exchange prices approximate competitive prices in the sense
that they come about as close to clearing the market as possible. Pool prices are
based on a different philosophy. They are designed to solve one problem optimally
while ignoring two others. Competitive prices, when they exist, play three useful
roles: (1) they induce least cost supply given demand, (2) they induce efficient
consumption, and (3) they induce efficient investment in generation capacity.
Bilateral markets and exchanges provide prices that approximately achieve all three
efficiencies while pool prices succeed perfectly at (1) while ignoring (2) and (3).3
The demand-side efficiency of pool prices is discussed in Chapter 3-8, while long-
run efficiency is discussed in Chapters 3-8 and 3-9.
Notice that, in the above example, the generator with a $20/MWh marginal cost
would lose money if paid only the nominal pool price of $25/MWh, so it would
be given a side payment of $5/MWh. Its total cost at full output, $30/MWh, would
then just be covered. Had this been an exchange or bilateral market, the “almost-
clearing” price would have been $35/MWh. With regard to short-run supply, this
makes no difference since, given demand, the pool always induces an efficient
dispatch. But with $5/MWh less revenue to cover fixed costs, long-run incentives
will be quite different. A pool does not concern itself with these incentives when
setting price and its philosophy of price setting does not produce efficient long-run
incentives as a side effect.
Again, the real-world implication may be minimal and empirical research or
detailed theoretical analysis is needed. The more difficult question is whether the
modest increase in efficiency attributed to optimizing short-run supply exceeds
the modest loss of efficiency from ignoring the long run.

3-5.2 DIFFICULTIES WITH BILATERAL DAY-AHEAD MARKETS

Markets should be encouraged as long as they do not inhibit the provision of
ancillary services by the system operator. These services are needed in real time,
3. Pool prices claim to send efficient signals to consumers but ignore some implications of nonconvex
production costs. When the optimal dispatch occurs at a production level for which marginal cost is not
defined, pool prices send the wrong signal to consumers. This circumstance occurs with a significant
probability. Chapter 3-8 gives an example which proves this point.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

248 PART 3 Market Architecture

not a day ahead, so the DA market can be problematic only to the extent it interferes
with RT operations. Trades made in the DA market can be rearranged in real time
as needed—with one exception. If a generator has not been started and is slow to
start, it may be unavailable when needed in real time. Thus, proper unit commitment
is sometimes required for balancing and calls for special attention when evaluating
DA market designs. The ancillary service of transmission security also impinges
on the design of the DA market by adding complexity to its operation.

Balancing and Unit Commitment

Real power balancing mainly refers to small, but often quick, adjustments of supply
to keep it equal to demand. If demand cannot be balanced because of a shortage
of supply, load must be shed. This is the most extreme result of imbalance and the
most important problem that could be caused by the DA market. The possibility
of this failure is greatest in a bilateral market.
Example of a Unit-Commitment Failure. Consider a bilateral market with
20 identical slow-start, midload generators, each with a different owner, only 10
of which are needed on a particular day. With 20 identical generators and only 10
needed, competition will be stiff and expected profits will be zero, but depending
on the number that commit, actual profits will be positive or negative. Before trading
begins, each generator knows that it has a 50/50 chance of selling its power in the
DA market, but it is indifferent as to whether it does or not because in either case
its expected profit is zero.
On the first round of trading, because there is no coordination in a bilateral
market, an average of 10 generators will sell power, but because of randomness,
5.8% of the time six or fewer will be committed.4 When this happens there is a good
chance that an unexpected RT event will cause a shortage and prices will be very
high. There is even a chance of a backout because there may not be enough quick-
start generators to make up for the lack of committed slow-start generators.
Normally, a bilateral market would fix this type of problem. Some customers
would realize they might be the ones demanding extra power in a tight market, or
some generators would realize they might have a forced outage and need to buy
extremely expensive replacement power. But the re-contracting process also takes
place without coordination. Just as the outcome of the initial round is partially
random, so is the outcome of re-contracting.
With enough time and with low-enough re-contracting costs, a bilateral market
will reach the optimal solution. But with only a few hours of trading it may not
even learn what collective mistakes were made in the first round. The problem is
not that bilateral markets have an inefficient equilibrium; the problem is that without
coordination and with a very limited trading time, they have a great deal of trouble
finding the equilibrium. This is especially true when the optimal set of trades
depends on the distribution of generation technologies selected and not just on the
character of individual trades. As a consequence, bilateral trading in a constricted
time frame produces results that are more random with regard to reliability than
4. This assumes that there are enough quick-start generators to substitute for most of the slow-start
generators that have nearly the same average costs including startup costs.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-5 The Day-Ahead Market in Theory 249

does a centralized market. The optimal solution to the reliability problem is not
a randomized level of commitment, so bilateral markets provide less reliability.
The magnitude of the problem deserves investigation, but for the present,
randomness of commitment in a bilateral market must be taken seriously. Power
markets that rely on bilateral DA markets usually do take it seriously and provide
some non-market arrangement to help ensure sufficient commitment. One mitigating
factor is that days with serious reliability problems are usually extremely profitable,
and any hint of such a problem is likely to produce ample commitment.

Result 3-5.1 A Bilateral DA Market Decreases Reliability

Because of information problems, a bilateral market, while getting essentially the
right answer on average, will produce a more random outcome than a centralized
market with respect to the level generating units committed in advance. This
randomness reduces the efficiency of the market and its reliability. The magnitude
of this effect has not been determined.

Transmission Security
The DA market cannot easily interfere with transmission security, but transmission
limits complicate the DA market. If locational pricing is needed because of transmis-
sion congestion, it will complicate all three types of market. The power pool and
power exchange will be complicated by having to compute them directly. The
bilateral market will be complicated by having locational prices induced by the
cost of purchasing transmission rights. The latter is the more complex process.
To ensure transmission security either physical rights must be centrally allocated
and sold, or security must be provided through RT pricing of transmission. The
latter produces volatile congestion prices that the bilateral trader will wish to hedge
with financial transmission rights. Financial rights are simpler in practice, but
physical rights will more easily illustrate the trading difficulties of a bilateral market.
The nature of the difficulty is the same under the two systems.
Consider a remote generator that wishes to sell its power. With physical rights,
the generator must buy a transmission right in order to implement a trade. Thus,
two steps are required for a complete trade. If the power is traded first, then the
generator must guess the cost of the transmission right before trading the power
in order to know how much to ask for the power. If the transmission right is
purchased first, the generator must guess the price at which power will be sold in
order to know how much to offer for the transmission right. In either case the
generator is forced to guess market conditions in order to set an appropriate offer
price in one market or the other.
In a centralized energy market, the system operator, in effect, sells transmission
rights along with any trades it arranges. There is no risk to traders because they
only end up purchasing these implicit transmission rights when they are worth
purchasing. In such a market, the competitive generator need not consider market
conditions but only needs to bid its marginal cost. If the market is tight, it will
receive the appropriately high price and earn a profit; if demand is slack, the price

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

250 PART 3 Market Architecture

will be low and it will neither sell its power nor accidentally buy a transmission
right it does not need. In a centralized market the generator needs only to understand
its own cost and has no need to forecast the weather.
Selling transmission rights is a more complex indirect method of determining
the same locational prices that are determined directly by a power exchange or
power pool. It increases the transaction costs of the market and the randomness
of its outcome. The result is not disastrous, but it is not beneficial.
Economic Dispatch
The last relevant ancillary service is economic dispatch. Although the bilateral
market may introduce more randomness in advance of real time than the other two
designs, most of this should be taken care of by re-contracting in the RT market.
Ideally the RT market should facilitate the rearrangement of prior bilateral contracts
at as low a cost as possible. If this is done, then with the exception of commitment
failures, the right dispatch should occur regardless of mistakes in prior markets.
Penalties for trading in the RT market interfere with re-contracting and so are most
harmful when applied to bilateral markets.

3-5.3 SETTLEMENT, HEDGING, AND RELIABILITY

Traders use forward markets, including the DA market, for hedging. The system
operator is interested in it for reliability. By scheduling generation in advance, it
hopes to guarantee its availability when needed in real time. Settlement accounting
provides the basis for analyzing the relevant incentives.
When power is sold in the DA market, the outcome is a quantity sold, Q1 , and
payment, R1. The payment, R1, would be Q1 × P1 in a power exchange and that plus
a side payment in a power pool. It is useful to define DA short-run profits, SRB 1,
to be the profits that would result from fulfilling the DA contract exactly.
SRB 1 = R1 ! MC × Q1 ! SC
DA short-run profit equals DA revenue less marginal cost and startup cost (which
should be thought to represent all nonmarginal costs). If the supplier performs as
specified in the RT market, it will earn exactly the DA profit regardless of the RT
price (See Chapter 3-2).
Depending on the RT price, the supplier may be able to increase its profit by
producing at a level different from the one specified in the DA contract. Assuming
that Q1 is the supplier’s full capacity, it can only earn more by producing less and
then only when the RT price is low. This is investigated with the help of final short-
run profits. These are given by
SRB F = R1 ! MC × Q0 ! SC0 ! P0 × (Q1 ! Q0 )
Final short-run profit includes DA payments, actual production costs and an
adjustment for any shortfall in actual production relative to the DA contract. Solving
for R1 from the DA short-run profit equation and substituting into the above
equation gives
SRB F = SRB 1 + (MC ! P0 ) × (Q1 ! Q0 ) + (SC ! SC0)

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-5 The Day-Ahead Market in Theory 251

Figure 3-5.1
Final profit including DA
contract revenue.

Final profit is DA profit plus any profit that can be made by buying RT power for
less than marginal cost plus any savings from not starting up. SC0 is actual startup
cost which will be either SC or zero.
As shown in Figure 3-5.1, a supplier with a hedging contract in the DA market
can guarantee itself at least SRB 1 if it starts up. If after starting, it finds the RT price
below its marginal cost, it can buy power instead of producing it to satisfy its DA
contract and thereby increase its profit. In this case it will wish it had not incurred
the cost of starting up, SC. But without incurring this cost it cannot protect itself
against the danger of high RT prices. This is shown by the RT short-run profit
function that goes negative for high RT prices.
Because startup is necessary to achieve the risk-reduction of a DA market
hedging contract, suppliers will be motivated to start up. But if they become
convinced RT prices will be very low, they may decide to save the cost of starting
up. It is unlikely that this will cause a reliability problem because whenever such
problems are likely, there is a good chance of high RT prices. Then suppliers with
DA contracts will start because they do not wish to risk not starting.
Notice that the pool’s side payment plays no role in this analysis. A DA contract
with a power exchange provides exactly the same incentives to start up as does
a DA contract with a power pool, provided the two contracts pay the same amount
for the same accepted quantity of power. This is discussed further in Section 3-7.3.
As shown in Example 2D in Section 3-9.2, a power exchange, if its bid structure
is too simple, may cause bidding to be somewhat random, which may result in
inappropriate bids being accepted. These inappropriately accepted bids are more
likely to fail to lead to physical commitment. In example 2D, this improves effi-
ciency and does not decrease reliability.
With flexible bidding a power exchange will accept bids with essentially the
same accuracy as a pool, and so exchange and pool contracts will pay the same.
Because DA side payments do not induce commitment, the two approaches will
have the same reliability properties. The consequences for reliability of less flexible
exchange bids are unknown.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

252 PART 3 Market Architecture

Result 3-5.2 Side-Payments in a Day-Ahead Pool Do Not Increase Reliability

Receiving a side payment, as opposed to being paid the same total amount
entirely through the market price, does not increase a generator’s incentive to
commit. Thus it does not increase reliability.

3-5.4 OTHER DESIGN CONSIDERATIONS

Ancillary services do not give rise to all of the factors that need consideration. Other
factors include transaction costs, facilitation of market monitoring, provision of
publicly known prices, and nondiscriminatory access (see Section 1-8.1). Although
these have been given less consideration both here and in the public debate, some
may be equally important.

Penalties To Promote Centralized Day-Ahead Trading

One design consideration that makes a frequent appearance is a penalty for not
trading, and thus not scheduling, in the DA market. Three reasons have been given
for this proposal: (1) the limitation of market power, (2) increased reliability, and
(3) increased efficiency. PJM requires its generators to bid in the DA market but
sets no limit on the price they can bid, so some bid extremely high thereby circum-
venting the requirement.
About 11% of PJM’s generation does not enter the required multi-part bids with
a low enough price to be accepted in the DA market, even though they generate
in real time. In spite of this, because of arbitrage bids, more power is sold day ahead
than in real time. Most of these arbitrage bids may be speculative bids because
hedgers would use multipart bids.
Because of the bidding inaccuracy in the DA market, PJM does not rely on its
outcome for reliability purposes. The same inaccuracies call into question the
efficiency claims of a pool.
To make an effective penalty, generators must be penalized for the difference
between their contract position and their RT production. There is some cost to
forcing trades into the DA market, but this is difficult to determine. There may also
be costs to last-minute scheduling. If these can be determined, imposing them on
last-minute traders is sensible. Market power considerations are more difficult to
evaluate.
Given PJM’s success at providing reliability without relying on the DA market
and given the doubtfulness of efficiency claims for a pool, penalties do not seem
to be warranted in order to increase the accuracy of the DA pool. Even if generators
are forced into the DA market, penalties cannot guarantee the accuracy of their
bids.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-5 The Day-Ahead Market in Theory 253

Complications and Simplification

Improvements in the basic designs should also be considered. As discussed in
Chapter 3-9, a slight complication of the power exchange’s one-part bids may
capture most of the benefit of a power-pool’s multipart bids. Similarly, greater
control over ramping might usefully be purchased by a system operator with a power
exchange.
Finally, simplicity and transparency of design are great virtues, not so much
because of the direct cost of operating under complex designs, which can be
substantial, but because of the flaws complex designs conceal.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 Chapter 3-6
The Real-Time Market in Theory

It must be done like lighting.

Ben Johnson
Every Man in his Humour
1598

U NLIKE A DAY-AHEAD EXCHANGE, A REAL-TIME EXCHANGE CANNOT

USE BIDS. The real-time (RT) market consists of trades that are not under
contract—power that just shows up, or is taken, in real time and accepts the spot
price. An RT exchange works like a classical Walrasian auction. A price is an-
nounced and suppliers and customers respond. If the market does not clear, a new
price is announced. The difference is that in a power market trade takes place all
the time; there is no waiting to trade until the right price is discovered. Like a
Walrasian auction and unlike a day-ahead (DA) exchange, an RT exchange may
find there is no price that balances supply and demand.1 Consequently, if an
exchange is used, it must be supplemented with another exchange or perhaps an
operating-reserve market in the form of a pool. There are many possibilities, and
little is known about their relative merits.2
Chapter Summary 3-6 : Pure bilateral markets are too slow to handle RT
balancing and transmission security. A centralized market is needed which can
take the pool approach, the exchange approach, or something in between. Pools
have an easier time achieving a supply-demand balance than do exchanges because
they utilize different prices for different generators. This allows them to offer an
option in a forward market that depends on the real-time market price. An RT
exchange can achieve a similar effect but only by employing one or more additional
exchanges. These could be for detrimental generation or operating reserves.
Alternatively an operating-reserve pool could supplement an RT exchange.

1. Although the lack of a market clearing price is indicated by a static analysis, the dynamics of a power
market which are limited by ramping constraints may reverse this conclusion.
2. See Rothwell and Gomez (2002) for descriptions of Norwegian, Spanish, and Argentinian spot markets.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-6 The Real-Time Market in Theory 255

Section 1: Which Trades Are Part of the Real-Time Market? RT trade

is trade that takes place at the RT price and not at a contract price. Thus, deviations
from quantities specified in forward contracts comprise the transactions of the RT
market.
Section 2: Equilibrium Without a Market-Clearing Price. As in the day-
ahead market, nonconvex production costs often prevent the RT market from having
a clearing price. Bilateral markets and pools, with their multiple prices automatically
have an equilibrium, but an RT exchange will often find it impossible to exactly
balance supply and demand. The assistance of a second or third market may be
required. This could be an RT market for decrementing generation below forward
contract levels, or it could simply be the market for operating reserves which is
needed in any case.
Section 3: Why Balancing Markets Are Not Purely Bilateral. Bilateral
markets are slower than centralized markets because traders must find a partner.
This makes it difficult for them to maintain a precise system balance while ensuring
transmission security. A centralized market takes advantage of electricity’s greatest
trading virtue: Electricity does not need bilateral delivery. When a load needs power,
power from any combination of generators will do. This flexibility speeds the
balancing process in a centralized market.

3-6.1 WHICH TRADES ARE PART OF THE REAL-TIME MARKET?

A clear understanding of what constitutes the RT market is required before the
problem of nonconvex costs can be considered. This question is best approached
by first identifying the amount of power traded in real time. For simplicity it will
be assumed that the RT price is set every five minutes and that settlements are
performed on a five minute basis.

Real-time Sales Are Not Under Contract.

If the RT price were increased by $1/MWh from 12:00 till 12:05, how much would
that increase the revenue of suppliers? Equation (3-2.1) provides the answer and
is repeated here for convenience.
Supplier is paid: Q1 × P1 + (Q0 ! Q1 ) × P0 [3-2.1]

This shows that the increase in revenue would be the total amount by which RT
supply exceeded power sold in the DA market times the RT price. Thus only the
deviations from day-ahead contracts are sold at the RT price. If power is also sold
in an hour-ahead market, these contracts would be settled according to the same
principle and again the RT price would only be paid for deviations from the forward
quantity.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

256 PART 3 Market Architecture

All power supplied in real time can be classified as sold under a forward contract
or not sold under a forward contract. If not sold forward, it is sold in the RT market
at the RT price. Thus power sold in the RT market is not sold under financially
binding bids but is instead the power that is not covered by any contract. Instead
it is simply produced and sold at the market price prevailing at the time of produc-
tion. Real-time trades are not based on accepted bids.
First Illustration. A complex illustration of this
Side Payments and Incentives conclusion occurs in PJM’s RT pool. PJM accepts bids
When a generator realizes that it will receive a in the operating-reserve market that takes place just
“make-whole” side payment because the RT price after the close of the DA market, but no quantity is
will have too low an average, the RT price loses its associated with the accepted bid so there is no binding
incentive properties. As shown in Chapter 3-2, the commitment on the part of the generators. But PJM
two settlement system preserves the incentives of does make a binding financial commitment. If a gener-
the RT price under normal forward contracts.
ator performs according to its bid and still loses money
Usually this incentive problem is corrected by a at the RT price, it will receive a “make-whole” side
substitute incentive: If the generator does not “follow
the dispatch,” i.e. follow the RT price, its make-whole
payment. This is what makes PJM’s RT market a pool
payment will be taken away. Typically such regula- instead of an exchange market. Power sold under such
tory incentives have hidden loopholes that do not an agreement is sold at the RT price if the average
occur with a price incentive. price is high enough so that no side-payment is needed.
Ramp-rate limits are said to be under-reported in But if the RT price is low, a generator’s payments are
multipart bids. This would allow a generator to ramp determined by its bid. In other words, the generator
slowly without being seen as “not following dispatch.” has been given an option to sell at the RT price or
Under a price incentive, the generator might well find under a forward pay-as-bid price. If the RT price is
it profitable to ramp more quickly. Side-payments high, the generator sells the power in the RT market
may dampen the responsiveness of generators to
the RT price.
rather than under a forward contract; while if the RT
price is low, it sells the power under a forward contract
Fortunately, side payments play a small role and
their interference with the price incentive is probably accepted in the operating-reserve market. Oddly, the
not a serious matter. generator does not always know which market it is
selling power in until the end of the day, after the
power has been produced.

Result 3-6.1 Real-Time Power Is Not Bought or Sold Under Contract

Power sold in the RT market is power sold at the RT price. Power sold under
contract is sold at the contract price and its cost cannot be affected by the RT
price. All contracts are forward contracts. Deviations in supply or demand from
forward contract specifications, intentional or unintentional, comprise the forward
market trades.

Second Illustration. A second illustration may further clarify the distinction

between the RT market and forward markets. Suppose a generator has offered to
sell 50 MW for $60/MWh for two hours and the system operator needs the power.
Say it accepts the contract at 12:00, sets the RT price to $60/MWh and the supplier
begins to flow power immediately. It appears that power is being traded under
contract in the RT market. But by 12:05 supply and demand conditions will have
changed and the system operator may well feel it needs to raise price to attract more

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-6 The Real-Time Market in Theory 257

supply or reduce it to attract less. Suppose it raises the RT price to $65/MWh. The
power flowing under the 12:00 contract continues to flow but is paid for at the
contract price of $60/MWh, not at the RT price. Five minutes after the start of the
contract, the power sold under the contract is no longer part of the RT market. In
fact it never was; the fact that it was traded at the RT price was merely a coincidence
and, as will be explained, probably a mistake. In some later five-minute interval
the RT price might again equal the contract price but that would again be a coinci-
dence and not an indication that for five minutes the 12:00 contract is again part
of the RT market. The litmus test is this: Would raising the RT price raise the cost
of the power? The answer is no, hence the contract power is not part of the RT
market.
The 12:00 contract is actually a forward contract with an extremely short lead
time. Usually such contracts are agreed to several hours in advance, but sometimes
they are agreed to only a few minutes in advance. In any case, the power delivery
takes place in the future, not at the time of agreement. Consequently the system
operator should evaluate the purchase over that time horizon. If the RT price will
be $60/MWh at the start of the contract but will likely increase to $200/MWh during
its two-hour duration, the system operator should be willing to pay more than the
RT price for the contract power. If the RT price is expected to decline it should
pay less than the RT price at the start of the contract. It is only by coincidence that
the price of power sold in a contract equals the RT price during the first five minutes
of the contract, or during the last five minutes, or at any other time. In any case,
by the time the power starts to flow, the payment for that power is not affected by
the RT price and so the power is not part of the RT market.
Third Illustration. Suppose a contract is written that stipulates a supply of 100
MW for two hours to be paid for at the RT price. Is this an exception? Is this power
under contract and also part of the RT market? There are two possibilities: (1) the
contract may stipulate the usual settlement rule, or (2) it may stipulate a penalty
for deviations from the 100 MW flow. With the usual settlement rule, deviations
are paid for at the cost of the deviation, that is, at the RT price. In this case if only
50 MW is delivered the supplier is first paid for the 100 MW at the RT price and
then it must buy 50 MW of replacement power at the RT price. The net effect is
simply that it is paid only for the amount it delivers and is paid at the RT price.
This is exactly what would happen without a contract. This is a contract without
effect and will be deemed not to be a contract.
If the contract specifies a penalty for deviations that is different from buying
replacement power at the RT price, this constitutes a genuine exception. There is
little reason for such a contract, but power sold in this way is part of the RT market
and is also sold under contract.

What Determines the Real-Time Price

The RT price is determined by total actual (RT) supply and demand. This includes
power traded under forward contracts and power traded in the RT market. Equation
3-2.2, repeated here, makes this clear.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

258 PART 3 Market Architecture

Supplier is paid: Q1 × (P1 ! P0 ) + Q0 × P0 (3-2.2)

Because suppliers cannot alter their forward quantity, Q1 , and, assuming a lack
of market power, they cannot alter (P1 ! P0 ), they optimize their choice of RT
output, Q0 , solely on the basis of the RT price, P0 . (The quantity sold in the RT
market is (Q0 ! Q1 ), not Q0 which is total RT production.)
When the system operator sets an RT price, it must balance total supply and
demand. Suppose more power is needed and there is a generator with a day-ahead
contract and a marginal cost of $65/MWh while the RT price is $60/MWh. In a
competitive pool, that generator would not produce even if it had sold its power
in the day-ahead market. It would earn more by buying RT replacement power than
by generating. Setting the market price to $66/MWh will call forth that supply and
help to balance the market. Thus raising the RT price can balance the market by
increasing the amount delivered under forward contracts, but that does not change
the amount sold under forward contracts; that amount is fixed. So any change in
the amount delivered is a change in the amount sold in the RT market. This is why
the RT price controls all RT power flows even though most of these flows are
forward trades and not RT trades.
Changing the RT price changes the quantities delivered under some forward
contracts as well as quantities delivered in excess of forward contracts. For example,
if some load it taking more power than it has contracted for, raising the RT price
may reduce that demand. This is a purely RT transaction.
The RT price should be set by taking into account the full supply and demand
response. If this is not done, it will be necessary for the system operator to circum-
vent the market in some way to balance supply and demand.

3-6.2 EQUILIBRIUM WITHOUT A MARKET-CLEARING PRICE

Although the unit-commitment problem is less severe in real time than it is a day
ahead, the startup cost problem still complicates the dispatch. Even relatively quick-
start generation has startup costs and these make production costs nonconvex. Also,
gas turbines are usually block loaded, meaning they are run at full capacity once
started. Again this causes a nonconvexity in the production-cost function. As
discussed in Chapter 3-5, such cost functions violate an assumption of perfect
competition and can prevent the existence of a market clearing price. As in the day-
ahead market, the lack of a market-clearing price does not indicate the market lacks
an equilibrium, but the RT character of the market changes the analysis.

Equilibrium Prices for Pools and Bilateral Markets

Nonconvex costs make the task of balancing supply and demand more difficult.
In this regard, a power pool has an advantage because it can control output with
more than the pool price. The possibility of side payments can help with fine tuning.
Thus if it needs a generator to start and produce at half of full output, it can set the
pool price to the generator’s marginal cost. If the generator has a forward contract
that gives it an RT option on a make-whole side payment, it will know that its

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-6 The Real-Time Market in Theory 259

startup cost will be covered even when the pool price is insufficient. Setting the
pool price to the generator’s marginal cost is a signal that it will be selling power
under the forward contract with the make-whole side payment. The power is not
sold just for the RT price. By taking advantage of its forward contracts in this way,
a power pool can achieve an equilibrium in spite of the lack of a market-clearing
price. Essentially an RT pool is a combination of two markets, a forward market
and a real-time market.
A pool has one more degree of latitude not available to an exchange. It can call
up a generator with a forward contract and tell how much to produce. If the genera-
tor obeys, then it retains the protection of it’s side-payment guarantee. If it does
not, then it loses the guarantee. There are cases when an RT pool must make such
phone calls if it is to achieve an efficient dispatch. For example suppose one of
two generators is no longer needed for reliability. If one has a marginal cost of $25,
while the other has a marginal cost of $20, the one with the lower marginal cost
would be more expensive if its no-load cost were $10/MWh at full output. In this
case the efficient dispatch requires it to shut down, but it will not do so until the
pool price is below $20/MWh, its marginal cost. At that price both generators would
shut down and that would reduce production too much. The solution is to phone
the low-marginal-cost generator and tell it to shut down while keeping the pool
price at $25/MWh.

Result 3-6.2 Real-Time Pools Sometimes Require Direct Control of Generation

To achieve an efficient dispatch, it is sometimes necessary for a pool to directly
control a generator’s output because the pool price does not provide the necessary
signal.

Like power pools, bilateral markets have the advantage of trading at many
different prices. Although these tend towards a single market-clearing price, when
that does not exist, there is nothing to prevent deals from being made at whatever
price is required. Given enough time and information, all profitable trades should
be made and the market will be in equilibrium.
The Problem of a Real-Time Exchange
A power exchange has the most difficulty balancing supply and demand because
in real time there is no formal bid acceptance process. True bidding is impossible.
In effect, an RT power exchange is a Walrasian auction. Without some elaboration
of this mechanism, and in the presence of nonconvex costs, the supply-demand
balance may be impossible to achieve. There are several possible responses to this
problem.
First, the problem may not need to be solved. The magnitude of the problem
has not been evaluated and a properly run exchange may come extremely close
to balancing supply and demand. All control areas experience some balancing error
and the result is “inadvertent interchange” between control areas. For the Intercon-
nection as a whole these approximately cancel due to the coordination provided
by NERC rules (which rely on the area control error, ACE, as an indicator). To
balance the market with a single price, the system operator would need to control

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

260 PART 3 Market Architecture

the RT price very cleverly. This could mean computing and announcing the RT
price somewhat in advance in order to assure generators that if they started, they
would be guaranteed a high price for at least some minimum amount of time. Such
pricing dynamics are difficult and presently not understood.
A second approach to balancing is to run an additional exchange which provides
more flexibility. The second exchange could be for decrementing generation.3
Suppose a generator has a marginal cost of $20/MWh and a no-load cost of $5/MWh
at full output. That generator will bid into an exchange at $25/MWh, but at that
price it will produce at full output because that is the only level at which it recovers
its full no-load cost. However it will be willing to buy back power at a price of
$20/MWh, its marginal cost. This can be accomplished in a second exchange that
accepts decremental bids. If bids are accepted, they must have an associated
duration and the exchange takes the form of a forward market. It would also be
possible to have an RT exchange for decremented power. This would apply to all
forward contracts. Any supplier who produced less than specified by its forward
contract would buy power at the RT decremental price, while any supplier produc-
ing more would sell power at the RT incremental price.

A Hybrid Solution
Power pools will inevitably do much of their RT trading in the exchange mode
because imports, load and some native generation will have no option of receiving
side payments. These traders must trade at the pool price. Cost nonconvexities on
either side of the market may make finding the supply-demand balance difficult
because large loads may turn off when a price threshold is reached and generators
may start and immediately ramp to full output. These problems are similar to the
traditional load-following problems faced by system operators. The solution to such
imbalances has always been operating reserves.
PJM runs a pool after the close of the DA pool and considers it to be an operat-
ing reserve market. Generators accepted in this market do not automatically sell
any power under contract, but if they follow the PJM dispatch they are guaranteed
make-whole side payments if they are needed. The make-whole guarantee provides
sufficient inducement for them to allow PJM to control their output. In PJM such
contracts cover a fairly small number of generators. In effect PJM has a small
operating reserve pool that operates alongside a larger RT power exchange.
In general, operating reserves need to be, and are, under the control of the
dispatcher. Consequently an operating reserve market seems to be the appropriate
tool for handling the small discrepancies in balancing caused by the use of an
exchange. This does not imply that the use of both incremental and decremental
RT exchanges is inappropriate.

3. This is just one possibility. Any additional exchanges will provide added flexibility and improve the
supply-demand balance. The design problem is to do the best job with the fewest number of markets. An
additional exchange could be an hour-ahead market or a market for ramping services or operating reserves.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-6 The Real-Time Market in Theory 261

The Efficiency Of Equilibrium Prices

Real-time pool prices suffer from the same defects as day-ahead pool prices. They
are not designed to be optimal signals to short-run demand or long-run supply. See
Sections 3-5.1 and 3-8.3.

3-6.3 WHY REAL-TIME MARKETS ARE NOT PURELY BILATERAL

Most markets operate under a regime of pure bilateral trade, so it is reasonable to
ask why RT electricity markets do not. The answer is related to the provision of
the ancillary services described in Chapter 3-4. Some of these must be provided
by the RT energy market, and a bilateral market is not well suited to their provision.
The critical issue is speed. Bilateral markets are slow to make complex trades.
Centralized markets are quick largely because they can take full advantage of the
homogeneity of electric power.
One ancillary service plays a special role with respect to bilateral trading. Trade
enforcement is required for the very existence of the bilateral market. It provides
the necessary property rights (see Section 3-4.6). This means that the bilateral
market is dependent on centralized accounting, metering, and enforcement. Still,
this does not preclude the bilateral market from handling all real power trades.
Voltage support is generally purchased in long-term markets and any interaction
with the real-power market is minimal, so this ancillary service plays a minor role
in RT trading and will be ignored. The black-start service is procured far in advance
and is irrelevant. This leaves three ancillary services to consider.
1. Real-power balancing.
2. Transmission security.
3. Economic dispatch.
The system operator must ensure that these services are provided but may do so
by utilizing a market. The question for an RT bilateral market is whether, by
stringently enforcing balanced trades, the system operator can induce a bilateral
market to provide these RT services. The first two of these are critical as they are
the basis for reliability. In combination they are extremely difficult for a bilateral
market to provide, which explains why all RT power markets are centralized. In
addition, the bilateral market would be so inefficient at providing frequency stability
(balancing) while complying with transmission security constraints that it would
fail to provide an economic dispatch.

The Balancing Market and Bilateral Trade

Although the primary goal of the balancing market is to maintain the real power
balance, it must do so while respecting transmission constraints and it should do
so at least cost. Real-power balancing means keeping the area-control error, ACE,
near zero. This is done by keeping local supply and demand, net of scheduled
interchange, in balance. NERC’s requirement for ACE is quite stringent.4
4. See Hirst (2001) for a clear explanation of the NERC requirements.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

262 PART 3 Market Architecture

Electricity’s greatest virtue as a tradable commodity is that it does not need

bilateral delivery. There are no different grades of electric power, and often the
point of supply matters very little. If a customer receives generator B’s power
instead of generator A’s power, no one cares. In a power market, this advantage
is amplified by the need for strict RT balancing. When one supplier falters, or one
load takes more than planned, any other supplier or group of suppliers can make
up the difference; there is no need for bilateral trades to stay physically balanced.
Forcing a power market to trade only bilaterally would limit the advantage that
can be gained from this interchangeability of supply. While many commodity
markets have interchangeable products, no other market needs to balance physical
supply and demand minute by minute to prevent physical damage to, or disruption
of, the marketplace.
Random Fluctuations. Load and intermittent generators, such as wind genera-
tors, fluctuate constantly. This makes it difficult to keep a bilateral contract in
balance. A centralized market has a natural advantage in this respect. Consider a
market at a time of day when load is not changing systematically, and suppose there
are 25 bilateral contracts. Say the 25 loads each change randomly up or down by
40 MW. The corresponding generators will need an equal adjustment, for a total
adjustment of 1000 MW. In a centralized market, because random fluctuations tend
to cancel, the total fluctuation in load would, on average, be only 1'5 as much
—
(1 '% 25 ). Thus only 200 MW of adjustment in generation would be needed.
Moreover, the adjustment would be made by whichever generators in the market
could adjust most cheaply.
Forced Outages. An unexpected forced outage of a line or generator causes
frequency-control problems that must be corrected within a matter of minutes. Such
outages cause imbalances in specific bilateral contracts, but sufficiently high
penalties could cause traders to contract for their own private spinning reserves.
This would impose unnecessary transaction costs and complexity on the market.

Transmission Security and Bilateral Trade

When the transmission network is unconstrained (as has been tacitly assumed so
far), the problems of the balancing market are relatively simple. If a bilateral
contract is out of balance due to a lack of generation, the trader responsible for
the imbalance may contract with any other generator in the market to compensate.
When there is a transmission constraint, this is not the case. Then it is necessary
to check with the system operator before completing the new agreement. Every
unscheduled adjustment must be checked in a purely bilateral market.
Consider a line that is operating at its security limit with power flowing from
A to B. If a particular bilateral trade has generation at B and load at A, it creates
a counterflow on the line. This trade decreases the physical flow on the congested
line. If the load is unexpectedly reduced and its generator follows, this will increase
the actual power flow on line A–B. In this case the trader responsible for the trade
must find another generator at the A end of the line and arrange for it to back down
instead of the generator that is part of its bilateral trade and located at the B end.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-6 The Real-Time Market in Theory 263

This type of complexity has proven too great for bilateral markets and as a
consequence the RT markets in fully deregulated systems are all centralized.
Because the system operators know the transmission constraints and can trade with
any generator in the system, they are able to keep the system in balance with a great
deal of accuracy, except in extraordinary circumstances.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 Chapter 3-7
The Day-Ahead Market in Practice

Predicting is pretty risky business, especially about the future.

Mark Twain
(1835–1910)

IS THE DAY-AHEAD POOL PRICE DETERMINED BY ARBITRAGE OR

COMPUTATION? Forward prices are usually determined by arbitrage between
the forward market and the real-time (RT) market. Day-ahead (DA) markets are
forward markets, but DA pools with multipart bids have been promoted for their
ability to determine through computation the optimal dispatch and the efficient
price. Both theories might prove true, or half true, but more likely one is essentially
right and the other wrong.
There is no question that the computation takes place accurately and in a
mechanical sense determines the DA pool price. The result of the computation is
also determined by its inputs. Because the computation itself is a fixed procedure,
while the input changes daily, it may be best to view the pool price as determined
by inputs. Then the question becomes: Do the pool’s input data accurately reflect
the producer’s reality, or do they deliberately misrepresent that data in order to
take advantage of arbitrage opportunities. In the first case, the pool calculation
makes use of good input data to produce a price that reflects the true details of
generation costs. In the second case, the calculation is not a sensible unit-commit-
ment calculation because its inputs are false. The bidders have manipulated the
pool’s computation, and the outcome may be thought of as being determined by
arbitrage.
Unfortunately, it takes only a small percentage of arbitragers to dominate the
outcome. In PJM’s DA market, 11% of the generators that are needed, and eventu-
ally produce power, in the RT market are rejected by the unit commitment calcula-
tion (PJM, 2001, 30).1 This alone proves the calculation is not highly accurate.
Morever, while 3,260 MW of multipart bids representing needed generation are
rejected, 6,169 MW of supply-side arbitrage bids (one-part bids) are accepted on
1. Only 26,771 MW submit DA multipart bids, while 30,031 MW proves to be necessary in real time.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-7 The Day-Ahead Market in Practice 265

average in the DA market. Thus more power is sold in the DA market than is needed
in real time. Many of the accepted arbitrage bids probably represent speculation
as hedgers would be inclined to use multipart bids in order to take advantage of
the insurance provided to accepted bids by “make-whole” side payments. This
circumstantial evidence does not prove that the arbitrage bids are controlling the
DA price, but it does suggest that this is a serious possibility.
Because the accepted DA bids commit 11% less generation on average than
will be needed in real time, PJM re-solves the unit-commitment problem with better
data immediately after the DA market closes. The outcome of this second computa-
tion is used to commit generators for reliability purposes. No power is traded in
this market and PJM views it as a method of providing operating reserves.
Chapter Summary 3-7 : Limited empirical evidence from the ISO markets
suggests that DA pool prices may well be determined primarily by arbitrage. DA
pool prices are not obviously more efficient than exchange prices, and they are
not needed for reliability as is demonstrated by PJM’s lack of reliance on them for
this purpose. The side payments of the DA pool seem useful for reducing the DA
market risk of generators, but this can be accomplished with far simpler bids and
without side payments.
Section 1: Arbitrage vs. Computation. The DA arbitrage price is the expected
RT price. The DA computational price is the pool price that would result from the
most accurate possible set of bids. These two prices determine whether the actual
pool price is primarily the result of computation or arbitrage.
Section 2: Efficiency. Given the extent of missing data in PJM’s DA pool and
the small magnitude of the errors that the pool computation is designed to address,
it seems likely that the computation is ineffective. There is currently no evidence
that DA pools provide a more efficient dispatch than DA exchanges.
Section 3: Reliability. PJM’s DA market does not serve directly as the basis
for reliability. A post-DA-pool calculation plays that role.
Section 4: Risk Management. The DA market provides insurance for some
generators with large startup costs and unpredictable profits. But the risks seem
modest, and a less-elaborate insurance scheme may be appropriate.

3-7.1 ARBITRAGE VS. COMPUTATION

The introduction to this chapter discussed two possibilities: The DA pool price
could be controlled by (1) arbitrage with the RT market, or (2) accurate bids and
computation. This section defines the conceptual difference more precisely and
suggests one observable consequence that might be used to distinguish between
the two possibilities.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

266 PART 3 Market Architecture

Operationalizing the Conceptual Difference

If all generators and loads in the PJM market submitted their bids as accurately
as possible, the DA pool calculation would produce a market price and a dispatch.
This price, PComp , is the benchmark for a “computationally” determined pool price.
Perfect arbitrage would result in the DA price equaling the expected RT price.
A good statistician, which an arbitrager would hire, can determine this price which
is the benchmark for a pool price determined by arbitrage, PArb .
If the actual DA pool price is P1 , the question of whether the pool price is
determined by arbitrage or computation can be operationalized as: Is P1 closer to
PArb or to PComp ? The most difficult step in answering this question would be the
computation of PComp , but PJM has apparently done most of the necessary work.
It has made its best judgement (which is probably excellent) of an accurate bid set.
These estimates could be run through the DA market’s algorithm to determine PComp .
In fact this may already have been done. (The actual computation used for unit
commitment after the close of the DA market is not public information.)

Why Computational and Arbitrage Prices May Differ

If PComp and PArb were the same, there would be no meaning to the question, and
perhaps they are very close. But if they are the same, then there is little to be gained
from computation. A standard market would find the same price and that would
imply much the same dispatch. The main selling point for the pool approach has
been the improved accuracy of the DA dispatch, which is said to improve efficiency
and reliability. This might occur as follows.
Suppose midload plants have a startup cost of $60/MW and a variable cost of
$15/MWh. If peakers have no startup costs and a marginal cost of $30/MWh, the
breakeven point is four hours. If the plant is needed for less time the peaker is
cheaper, but for more time, the midload plant is cheaper. Suppose that on half the
days power is needed for just less than four hours and on the other half for just
a little more. If the price is computationally determined, the accurate input data
will provide better load and generator-availability data than is publicly available.
Also the computation will correctly make complex unit-commitment decisions that
the market cannot predict. The result will be that the DA pool will price power at
$15/MWh on half the days and $30/MW on the other half, and it will usually be
right.
If the pool is usually right, the DA pool is providing a valuable service by
dispatching accurately. The arbitrage price, as defined, is the best feasible public
prediction of the RT price. Since the DA pool price is assumed to anticipate the
RT price, if the RT price could be predicted, PArb would equal P0 and thus PComp.
But the only way to predict the RT price is by replicating the pool’s computation
with good data and this is not feasible. So the arbitrage price, PArb , will get the RT
price right on average but will be wrong every day. The arbitrage price will always
equal the average RT price, $22.50. In this example, if the pool really works, its
price will be a more accurate predictor of the RT price than the arbitrage price.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-7 The Day-Ahead Market in Practice 267

A Simpler Test
Although PJM has the necessary inputs to compute PComp , without those inputs,
it is hard to estimate. But, as seen in the present example, if the DA pool were very
accurate, the pool price would come quite close to the RT price, while the arbitrage
price would be much further away. The average arbitrage price is just as close to
the average RT price, but on any particular day, the actual pool price exactly equals
the RT price. The arbitrage price is always off by $7.50 one way or the other.

Hypothesis The DA Pool Price is More Accurate

Because the DA pool has access to much accurate data that is not publicly
available, and because of its superior computational abilities, the DA pool price
more accurately predicts the RT price than does a DA estimate based only on
public information.

If the DA pool has better information and better computing abilities than the
market at large, then the pool price should be closer to the RT price than DA
predictions based only on public information. This hypothesis could be tested as
follows.
The actual DA pool price, P1, could be compared with the RT price hour by
hour for a year and the mean absolute deviation computed. Then the arbitrage price
could be approximated compared in the same way. (Root-mean-square deviations
could also be used.) If P1 were found to be significantly closer to the RT price than
to the estimated arbitrage price, the hypothesis of a computationally determined
DA pool prices would be accepted.
If P1 is found not to predict the RT price any better than the estimated arbitrage
price, then the possibility that the DA pool price is determined by arbitrage remains
open.

A Glance at the Data

A little calculation shows that using only the hour of the day and no other informa-
tion, the RT price can be predicted with an mean absolute error of $17, compared
with the $12 error achieved by the DA pool price, P1 . It seems plausible that taking
into account day-of-week, season, and weather forecast might reduce that error
to $12. This would favor rejection of the above hypothesis.
If only the 95% of hours with RT prices below $80/MWh are considered, the
simple prediction, based on a constant and the hour of the day, out-performs P1
when predicting the RT price (as measured by standard errors). The standard error
of P1 is $15.99 while the standard error of the estimated arbitrage price is $14.57.
Measured by average absolute errors, the DA price is a slightly better predictor.2

2. To test whether these results were due to some difficulty with DA bids when RT prices exceeded
“physical” marginal costs, all hours with RT prices above $80 were removed. The DA average absolute
error was then $9.25, while the simple prediction absolute error was $10.50. For all hours, the DA standard
error was $29.47 and the standard error of the simple prediction was $45.16.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

268 PART 3 Market Architecture

These data suggest even more strongly that the hypothesis should be rejected,
at least in the 95% of hours with price below $80/MWh. Since the true arbitrage
price, as defined, would be more accurate than the very simplistic predictor used
for this test, the hypothesis would be even more strongly rejected if tested more
rigorously.
Another piece of evidence comes from Hirst (2001) who reports that the
correlation between DA and RT prices was 63% in PJM, 62% in California’s
market, and 36% in NYISO. Note that the California market and the PJM market,
both of which allow arbitrage, score the same, while the NYISO’s DA pool predicts
the RT price less well than California’s arbitrage-based market. This may be because
arbitrage in the NYISO market was more difficult during this period. That would
indicate that arbitrage improves the accuracy of pool pricing and is thus the deter-
mining factor.
Both sets of evidence are inconclusive, but both suggest that the possibility that
pool prices are determined by arbitrage rather than computation (as defined above)
should be taken seriously. If arbitrage is determining the pool’s outcome, then the
computation is largely a charade. This question deserves investigation.

3-7.2 EFFICIENCY
Three primary arguments have been made for the use of DA pools: (1) efficiency,
(2) reliability and control, and (3) risk management for generators. The efficiency
argument asserts that a DA pool will increase the efficiency of the dispatch over
that obtained with a DA power exchange. The case for this has never been made
in writing, but it is still widely believed among those who support the pool approach.
The theory of the efficiency of DA pools begins with the observation that
nonconvex generation costs prevent the existence of a market clearing competitive
price and so invalidates the Efficient-Competition Result (Result 1-5.1). This proves
a power exchange has no claim to perfection, so a pool might do better. The pool
approach is explicitly designed to handle nonconvex costs, but these occur only
at the generating unit level and affect only a portion of the generation costs.
Because the cost nonconvexities are very small, accurate data are required if
they are to be properly taken into account and the dispatch is to be optimized. With
no generating unit ever exceeding 5% of the market and the typical unit closer to
0.5% of the market, it would be surprising if incorrect data for 11% of the genera-
tion did not significantly degrade the ability of a pool to find an optimal dispatch.
In fact, more than 11% of the data in PJM is incorrect from the point of view of
a unit-commitment program because of the large number of arbitrage bids, and the
result is a dispatch that is 11% short of committed units. This must negate the effect
of the detailed optimizing calculations of a pool. Before the efficiency argument
can be given credence, some theoretical or empirical support for it must be ad-
vanced.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-7 The Day-Ahead Market in Practice 269

3-7.3 RELIABILITY AND CONTROL

Reliability has become the primary justification for a DA power pool. The basic
argument has two steps. First a DA power pool will find the efficient dispatch more
accurately. Second, its side payments will assure the selected generators start up.
The first step can be questioned for the reasons explained in the previous section.
At least in PJM, the second step is incorrect. As explained in Section 3-5.3,
because PJM relies strictly on a two-settlement system, DA market side payments
are not contingent on RT performance. Consequently they have no influence on
a generator’s decision to commit or not. DA markets of any variety encourage
commitment because, once a hedging contract is signed, the signing party must
perform in order to be hedged (see Section 3-5.3). Consequently, PJM’s DA pool
has no advantage over a DA power exchange in this regard.
PJM does not use the DA market to assure its reliability. Instead it uses a
separate after-market calculation that interacts with the real-time market. This
calculation may be quite similar to a unit commitment calculation. It is based on
some bids submitted after the close of the DA market as well as on DA market bids.
It could just as well rely entirely on bids submitted outside of the DA market. Thus
PJM’s market demonstrates that a DA pool is not needed for reliability. Perhaps
a multi-part bid calculation with side-payments is needed, but this can be done
outside of the DA market.

Result 3-7.1 A Day-Ahead Power Pool Is Not Required for Reliability

PJM achieves all of the reliability benefit normally attributed to the pool approach
without relying on its DA pool for this purpose.

PJM’s post-market calculation is not a power market—no power is traded—but

it is a type of market. Essentially it is a market in which PJM trades profit insurance
for an implicit guarantee to start up. Generators submit bids which for simplicity
can be taken to consist of a startup cost (SC) and a marginal cost (MC). If PJM
accepts the bid, it guarantees that the generator’s profit, as computed by
Q0 × (P0 ! MC) ! SC + Side Payment
will be nonnegative where Q0 is actual production as a function of time and P0 is
the real-time price. The side payment is an insurance payment because it is made
only when the generator has the misfortune to earn a negative profit. Presumably,
PJM minimizes its expected insurance payments subject to the constraint that
enough generators must be started to satisfy reliability requirements. This arrange-
ment benefits both the generators and PJM. The implicit guarantee of the generator
to start up is believable because it cannot lose money if it starts and has some chance
of earning a positive profit. To claim the insurance it must “follow the PJM dis-
patch” which means keeping its output within about 10% of the quantity indicated
by P0 and its marginal-cost bid, MC. Ramp-rate limits, as indicated by its bid, are
also taken into account.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

270 PART 3 Market Architecture

Alternatively this market may be viewed as a market for spinning reserves. In

fact this is essentially PJM’s view of it. For the accepted generators, it has the effect
of making side payments contingent on RT performance thereby turning part of
PJM’s RT market into an RT pool.

3-7.4 RISK MANAGEMENT

The third justification for a DA pool is that it helps generators manage their risk.
As seen in the previous section, the insurance market also provides risk manage-
ment, but the DA market plays a similar role.
In a power exchange, it is quite possible to bid in such a way that a profit is
expected yet find that money has been lost if the bid is accepted. For example, a
generator with a $30/MWh marginal cost may bid that cost on the belief that the
market price will be above $30 by enough and for long enough that it will recover
its $40/MW startup cost. But its bid might be accepted for four hours at a price
of only $35/MWh. In this case it recovers only half of its startup costs. Regardless
of what happens in the RT market, it has lost money in the DA market. It would
have been better off had it not bid. This cannot happen in a pool if the generator
bids its true costs. In a pool the generator would bid the $40/MW startup cost as
well as its marginal cost, and the bid would probably not be accepted. If it were
accepted at a price of $35/MWh for four hours, it would be accepted with a side
payment of $20/MW, and the generator would break even.
A primary purpose of forward markets is to hedge suppliers and consumers.
When the forward market itself generates risk this partially defeats the purpose.
There is a real advantage in the pool’s reduction of risk for generators. But if this
is the reason for a DA power pool it raises the question of design and complexity.
Unit commitment programs are extremely complex and require many inputs
because they are designed to optimize the dispatch. They were not designed to
minimize market risk. Had they been, their design would have been far simpler.
Two part bids, as are used in the Alberta power exchange would do the job. The
bids in that market include a minimum run time. The actual risk involved in
committing for a day is quite small, both because a day is a small fraction of a year
and because generators can predict prices well enough that they will rarely start
and then find they earn no scarcity rent at all. Adding a second part to the bid should
make a small risk negligible. The big market risks for generators come from annual
changes in load and supply, not from small daily errors. So the effect of a small
remaining DA market risk will be undetectable.3
While risk management argues for two-part bids it does not argue for a pool.
The signature of a pool is side payments. Side payments are not required to remove
the DA market risk. Alberta’s power exchange does not make side payments and
yet removes most of the risk with its two part bid. An exchange sets the market
price high enough to cover the costs of all accepted bids. No bid is accepted that
loses money on its own terms. If two or three part bids are used, the generator can

3. Risks are subadditive because they are proportional to standard deviations. Variances of uncorrelated
events are additive and standard deviations are proportional to the square root of the variance.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-7 The Day-Ahead Market in Practice 271

bid so that any bid that does not lose money on its own terms will, at worst,
breakeven if accepted and fulfilled.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 Chapter 3-8
The Real-Time Market in Practice

Where there is much light, the shadows are deepest.

Goethe
Wilhelm Meister’s Apprenticeship
1771

W HEN MARGINAL-COST PRICES WON’T CLEAR THE MARKET, ARE

THEY STILL THE RIGHT PRICES? When production costs are “nonconvex,”
as are startup and no-load costs, competitive market theory predicts the market may
not clear, marginal-cost prices may not be optimal, and the market may not be
efficient. The sole purpose of the multipart bids used by power pools is to overcome
problems caused by nonconvex costs. The pool approach recommends setting the
market price equal to marginal cost exactly as if there were no problem and then
making side payments to generators who are needed for the optimal dispatch. These
payments cover only the costs that marginal-cost prices fail to cover.
The pool approach recognizes the first failure of marginal-cost pricing and
corrects it with side payments. But true competitive prices do more than minimize
production costs; they send the right signals (1) to consumers and (2) to investors
in new generation. Can a power pool’s combination of marginal-cost prices and
side payments replicate these benefits of competitive prices? Using an example
from a dispute over NYISO’s pricing, this chapter shows that pool pricing fails
both of these tests. Surprisingly, it was NYISO’s position that standard marginal-
cost pricing—pool pricing—was inefficient for at least five reasons.
Pool prices are neither the prices of Adam Smith, nor those of competitive
economics. They are not right for the demand side and they are not right for long-
run investment. Pool prices are right for the centralized solution of the problem
of minimizing short-run production-costs, given an output level that is incorrectly
determined when demand is elastic.
This does not mean the pool approach is a bad idea; it simply means that
adopting it because it gets the prices right would be naive. The problem of
nonconvex costs is difficult and a complex market design, such as a pool, could

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-8 The Real-Time Market in Practice 273

be needed.1 No market design is likely to get the prices exactly right even in theory,
so getting them wrong proves little.
A simpler approach is used as widely and deserves equal attention—the power
exchange approach. Unlike pools, exchanges do not make side payments. As a
consequence exchange prices tend to be slightly higher than marginal cost. The
change in pricing introduced by the NYISO and approved by FERC raised the
market price above marginal cost just to the point where it covers the costs of a
generator that might otherwise have needed a side payment. This is exactly the
philosophy of power exchange pricing. The problems with pool pricing listed by
the NYISO correspond well with the theoretical problems of pool prices, and the
advantages they sought are those offered by a power exchange.
Chapter Summary 3-8 : The real-time (RT) market can use a pool approach
or an exchange approach. Pools use marginal cost prices and side payments. These
prices are too low in cases where another unit of demand would cause a jump in
total cost. Instead, price should be raised to curb demand until demand increases
to the point where the cost increase is warranted. Pool prices also send the wrong
signals for investment. Exchange prices are also suboptimal, but they are simpler
and potentially more accurate with regard to demand and investment.
Section 1: Two Approaches to Balancing-Market Design. The power-pool
approach collects data on generators through multipart bids, computes an optimal
dispatch, and sets the market price equal to marginal cost. This does not clear the
market, so it makes side payments to keep some required generators in the market.
The power-exchange approach typically uses one- to three-part bids, but its distin-
guishing feature is a lack of side payments. It relies on a single price, while a pool
uses prices that are effectively tailored to individual generators.
Section 2: The Marginal-Cost Question As Decided by FERC. FERC was
asked to direct the NYISO to use marginal-cost pricing as specified in its tariff,
and on July 26, 2000 it did. The NYISO objected and proposed to keep its above-
marginal-cost prices. FERC accepted NYISO’s old pricing scheme with little
modification. Non-marginal-cost pricing will be applied in real time while the same
situation will call for marginal-cost pricing in the day-ahead (DA) market.
Section 3: Making Sense of the Marginal-Cost Pricing Charade. Although
NYISO’s prices are not marginal-cost prices, they are probably better. Theory
indicates that marginal-cost prices can be too low for all purposes. Side-payments

1. Because nonconvex costs violate a basic assumption on which the Efficient-Competition Result is
based, they invalidate many of the conclusions previously reached in this book. Fortunately the problems
actually caused by nonconvex costs are small in magnitude, as is explained in Chapters 3-8. So, while there
may be no market clearing price, a power market will have a (Nash) equilibrium price that is extremely
efficient.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

274 PART 3 Market Architecture

fix the short-run supply-side problem but not the demand-side or investment
problems. NYISO’s above-marginal-cost pricing moves toward the power exchange
approach in this circumstance.
Section 4: The Power Exchange Approach. An RT power exchange could
set a market price different from any bid and let the non-bid supply and demand
responses, along with accepted bids, clear the market. Because of the nonconvex-
cost problems that pools focus on, an additional market for ramping services or
decrementing generation is beneficial.

3-8.1 TWO APPROACHES TO BALANCING-MARKET DESIGN

A centralized, RT, energy market must provide three ancillary services: balancing,
transmission security, and efficient dispatch. When system operators worked for
regulated monopolies, they had direct physical control, but in a market, they must
rely on RT prices and contracts for control. This increases the difficulty of providing
the first two services and introduces the problem of market power, which compli-
cates the job of providing an efficient dispatch.
Having ruled out the purely bilateral approach to the RT market, two basic
approaches remain: the exchange approach and the pool approach. The exchange
approach uses simple bids and relies on a single price which is set to clear the
market as well as possible. The pool approach is based on the observation that cost
nonconvexities make clearing the market impossible. It compensates for this flaw
with multipart bids, optimal bid-acceptance procedures, and side payments. The
main cost problems (nonconvexities), are low-operating limits, no-load costs, and
startup costs. The litmus test for a pool is the existence of side payments, which
are not used by exchanges.

The Block-Loading Example

The different natures of the two approaches are best explained by considering how
each would handle the same nonconvex cost problem. Consider a system with only
low-variable-cost steam units ($20/MWh) and high variable cost gas turbines (GT’s)
($40/MWh). The GT’s must be block loaded; that is, they must be dispatched at
their full 100-MW output capacity or not at all. For simplicity, they are assumed
to have no startup costs, and steam units are assumed to have constant variable
costs. Suppose that the total maximum output capacity of all steam units is
20,000 MW, that load has just increased to 20,020 MW, and that the system
operator is required to keep the system perfectly balanced.

The Pool Approach

This problem will first be analyzed according to the pool approach which has three
steps:

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-8 The Real-Time Market in Practice 275

1. Find and implement the optimal dispatch.2

2. Set market price equal to system marginal cost.
3. Make side payments as needed to compensate dispatched generators for
negative profits.
One GT must be dispatched, but it provides too much supply so a steam unit must
be backed off by 80 MW. Then total supply equals load at 20,020 MW. System
marginal cost is popularly defined as “the cost to supply the next increment of load.”
This requires clarification at a point where all generators are fully loaded and
another must be started. But the present case is free of this ambiguity as the GT
has been started and there is an unconstrained generator. One more megawatt of
load, or two more, or 20 more would all be supplied at a cost of $20/MWh hour
by the steam generator that has been backed down. Similarly, one less would save
$20/h.
Miscalculation of Marginal Cost. Because both FERC and NYISO
miscalculate marginal cost in this example it is necessary to consider
their misinterpretation in some detail. The basic notion of marginal cost
is that it is the slope of the total cost curve at a particular operating point.
Thus it has nothing to do with the actual chronology of load arrivals or
generator startups. In the present example the total cost curve has a
constant slope starting well below the operating point and continuing
for 80 MW above that point. Its slope is unambiguous at the operating
point and can be calculated in the usual way as the cost of serving the
next increment of load. This is shown at the left and explained by Figure
3-8.1.
The misinterpretation assumes that the relevant “increment” of load
is the change in load that caused the GT to be started. But this is not the
“next” increment; it is the prior increment. Moreover it misinterprets “next” as
having to do with time, while the phrase “next increment” in the marginal cost
definition actually means “an additional increment.” (Note that when the 20-MW
increment of load caused the GT to start, the incremental cost was actually
$120/MWh, not $40/MWh, because that 20-MW increment raised system cost by
$2,400/h.) The system’s marginal cost when $20/MWh generation is available is
$20/MWh.
Side payments. The third step in the pool approach is to pay generators that have
been dispatched but have lost money. In this example the GT lost $20/MWh and
needs a side payment of $2,000/h.
The Rationale. The rationale for an RT pool is first that it achieves an efficient
dispatch. The required power could not be produced at a lower cost. Second,
consumers are sent the right price signals in the sense that if they choose to consume
five more megawatts, that will cost $100/h to produce and consumers will be

2. As will be demonstrated, the standard pool approach does not find the optimal dispatch when demand
is elastic. Instead it finds a dispatch corresponding to a price that clears the market at marginal cost with
the help of side payments (when this is possible).

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

276 PART 3 Market Architecture

Figure 3-8.1
Marginal cost is
$20/MWh because
marginal power is
provided by a steam unit.

charged $100/h to consume it.3 As a final consideration, the market should send
the correct signals for long-run investment. Marginal-cost prices do this when costs
are convex. When they are nonconvex and side payments are included, economics
makes no guarantees. The hope is that any long-run damage will be less than the
short-run gains from efficient dispatch.

The Exchange Approach

The exchange approach is not as well specified in theory as is the pool approach,
though in practice it may have less variability. Its first rule is to make no side
payments. This means that to induce the GT to start up it must set the market price
to $40/MWh, or higher.4
To prevent oversupply, the exchange approach needs some way to induce a
steam unit to back off. This might be done by accepting a decremental bid. In this
case it would be an offer by the steam unit to pay the system operator perhaps
$19.90/MWh to allow the steam unit to reduce its output while still being paid
according to all prior contracts as if it were producing its full output. If it has sold
its power for $22/MWh and the system operator bought a decremental contract
at $19.90, the steam unit would be paid $22 and would in turn pay the system
operator $19.90 for the megawatt-hours that it sold but did not produce.
Many designs are compatible with the exchange approach, and some will work
better than others. In an exchange approach, there might be several parallel markets,
such as one for incremental energy, one for decremental energy, one for reserves,
and another for ramping or load-following. Each would attempt to set price to clear
the market. Pools use a similar set of markets but handle the problem of backing
down units differently.

3. This assumes that the consumer sees the RT price, which is unlikely, but the job of the RT market is
to produce the right signal; passing it on to consumers is a separate problem.
4. While the market has no “market-clearing price” in the classic economic sense, an exchange will have
an equilibrium price which typically comes much closer to clearing the market than the pool price.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-8 The Real-Time Market in Practice 277

3-8.2 THE MARGINAL-COST QUESTION AS DECIDED BY FERC

On April 24, 2000, the New York State Electric and Gas Corporation (NYSEG)
filed a complaint with FERC that challenged NYISO's fixed block (block-loaded
GT) pricing methodology on the grounds that it violated NYISO's (FERC-approved)
Tariff and theory of locational-based market prices (LBMPs). In effect NYSEG
claimed (among many other things) that NYISO should set price to system mar-
ginal cost (SMC) at all times including when block-loading a GT caused a steam
unit to be backed down.
FERC (2001a) describes its July 26, 2000 decision as follows
. . . the Commission concluded that when less expensive genera-
tion resources are dispatched down for the purpose of accommo-
dating more expensive fixed block resources, the marginal cost
of supplying the next increment of load should be equal to the
bid price of the least expensive unit that has been backed down.
The Commission concluded that in these instances the Services
Tariff required the energy price to reflect the marginal cost of
the backed-down unit, and directed NYISO to revise its proce-
dures for setting prices accordingly.
FERC and NYSEG were right. The NYISO Tariff and NYISO’s theory of pool
pricing require price to be set to SMC, and the cheap unconstrained generator
determines SMC, not the expensive constrained generator.
On August 25, 2000, NYISO filed for a rehearing and proposed a “hybrid”
pricing approach which almost entirely rejected the use of SMC pricing in the RT
market.5 In this filing NYISO made two points that are investigated in the next
section, and several that go beyond that analysis. All of these points are implicit
criticisms of SMC pricing. First it noted that the pool approach (SMC pricing) sends
an inefficiently low price signal to loads.
The Commission’s rule [SMC pricing] would discourage the
development of price-responsive real-time loads in New York,
because the RT prices it would establish would not reflect the
incremental cost of meeting load. The Commission’s rule would
eliminate the incentive of loads to reduce output in real time in
many hours when the incremental cost of meeting load is very
high, because this high cost would not be reflected in RT prices.
(NYISO 2000, 9) (Emphasis added.)
Second it notes:
. . . the new rule [SMC pricing] might: . . . (ii) frequently under-
state RT prices when supplies are tight, thereby sending inaccu-
rate price signals to the market and creating long-term ineffi-
ciencies. (NYISO 2000, 5) (Emphasis added.)

5. SMC pricing was retained for the unusual event that a GT was on because of a minimum run-time
constraint but was otherwise not needed by the RT market.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

278 PART 3 Market Architecture

It did not specify the long-term inefficiencies, but the main long-term concern of
any market is efficient investment. This would coincide with the finding in the next
section that SMC pricing discourages the steam units relative to the GT’s when
the opposite is desired.
Because their analysis considers the full complexity of the market, the NYISO
discovers a number of other problems not covered in the following analysis. SMC
pricing would discourage imports.
The Commission’s fixed block pricing rule [SMC pricing] will
very likely discourage external suppliers from delivering real-
time imports from external resources that are scheduled in the
DA market. This disincentive would arise because the real-time
imbalance price would be artificially driven below the market-
clearing level. (NYISO 2000, 8) (Emphasis added.)
It would discourage loads from participating in the DA market.
First, the new rule [SMC pricing] can be expected to undermine
the willingness of loads to participate in the DA market because
of the high cost of meeting incremental load in real-time would
not be reflected in RT prices, i.e. real-time LBMPs would usually
be set by less expensive steam units, rather than fixed block
GT’s that actually were incremental. (NYISO 2000, 10) (Empha-
sis added)
Coupled with the use of uplift and side payments, it would distort congestion costs.
Second, because the NYCA’s [New York Control Area’s] fixed
block GT units are located principally in New York City or on
Long Island, they often operate only when there is transmission
congestion, and they only set prices in the eastern part of the
state. In this situation, the Commission’s fixed block pricing rule
[SMC pricing] will likely lead to lower prices in the east, but not
in the remainder of the NYCA. Any reduction in real-time LBMPs
below the bid prices of GT’s running to meet eastern loads will
be recovered in the form of uplift charges from customers all over
the state. The net effect will be to shift congestion costs from
eastern New York into state-wide uplift. (NYISO 2000, 10)
It would encourage load-serving entities, which would otherwise not want a high
market price, to exercise market power.
The Commission’s fixed block pricing rule [SMC pricing] gives
certain market participants an incentive to exercise market
power . . . . such an entity would have an incentive to bid ex-
tremely high, knowing that it would receive substantial uplift
payments at the expense of customers throughout the NYCA if
its resources are called upon, while the price it pays for energy
will be set by less expensive resources. (NYISO 2000, 12)
The above analysis by the NYISO appears to be accurate although decidedly
nonrigorous and intuitive. The only significant error in its analysis is required for
bureaucratic reasons. It concludes that

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-8 The Real-Time Market in Practice 279

. . . the Commission could revise its rule to permit fixed-unit GT’s

that must be run to set real-time LBMPs without requiring the
NYISO to amend its tariff because the GT’s bids will truly reflect
“the marginal cost of supplying one more unit of energy.”
(NYISO 2000, 8)
This claim avoids the difficult task of amending the most fundamental element of
the tariff’s philosophy—price will be set equal to “the cost to supply the next
increment of Load at that location (i.e., short-run marginal cost)” (NYISO Tariff).
Note that it claims their price will “truly reflect” the SMC, not equal it. Their next
reference to the cost of the GT calls it the “actual incremental cost of meeting real-
time load;” it does not refer to it as marginal cost. FERC (2001a) accepts this claim
that marginal cost is not marginal cost, forgetting what it knew in July 2000.
We note that the issue being resolved here arose because NYISO's
tariff did not clearly specify how the LBMP would be calculated
when lower-priced units were backed down to accommodate fixed
block units. To avoid such disputes in the future, NYISO's tariff
should specify how it will treat fixed block units in setting the
LBMP.
FERC claims the source of the problem was that NYISO’s tariff was not clear
on how LBMP would be calculated. But the tariff was not at fault. It clearly stated
the price would equal “the cost to supply the next increment of Load at that location
(i.e., short-run marginal cost).” In the present context, this is unambiguous. NYSEG
understood it, and FERC understood it the previous July. All of NYISO’s substan-
tive arguments were aimed at showing the price should not equal marginal cost.
Their argument consists of this:
Sometimes our price, X, is a better price than marginal cost and
sometimes it is not. When it is, then marginal cost must be X
because marginal cost is always the best price. In the DA market,
marginal cost will be the best price, so then marginal cost can
remain marginal cost. But in the RT market, X is better, except
in certain circumstances when FERC’s [correct] interpretation
of marginal cost is best. Then marginal cost can remain marginal
cost in real time as well.
By approving this approach, FERC removed the previous clarity of NYISO’s
tariff. NYISO was right that their price was better than marginal cost, and FERC
was right to approve it, but despite the best efforts of regulators and system opera-
tors, marginal cost will remain “the cost to supply the next increment of load”
whenever the cost function is smooth.

3-8.3 MAKING SENSE OF THE MARGINAL-COST PRICING CHARADE

The New York ISO is supposedly founded on the rock of marginal-cost pricing.
But when one of its members tried to enforce this simple rule, NYISO put up a
vigorous and successful fight against it. In the process it convinced FERC that

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

280 PART 3 Market Architecture

marginal cost does not mean what everyone knows it means. The outcome is a rule
based on the misinterpretation of the tariff, but which by all appearances is an
improvement on SMC pricing.
Could SMC pricing really be as dangerous in this case as claimed by the
NYISO? If SMC pricing is wrong how should price be set? Should it always equal
the average cost of the last unit started? Even more puzzling is the fact that FERC
and the NYISO retained proper SMC pricing for the identical circumstance in the
DA market while replacing it in the RT market. Is there some incompatibility
between SMC pricing and an RT market?
One clue to this puzzle is that the pricing rule NYISO fought for and won is
a step toward the exchange approach. It is designed to eliminate the side payment
to the GT, and it attempts to implement a market-clearing price (see emphasis in
the third quote above from page 8 of NYISO’s filing). But to understand the
underlying problem a theoretical analysis is required.

Analysis of the Block-Loading Example

In the example of Section 3-8.1, when load reaches 20,020 MW the system operator
has a problem. At a market price less than $40/MWh, the new generator will
produce nothing and demand will exceed supply. At a market price of $40/MWh
or more, the next generator will produce 100 MW and supply will exceed demand.
There is no price that clears the market. This problem, characteristic of nonconvex
costs, makes the exchange approach difficult. When generalized to similar problems
caused by the other cost nonconvexities, it is the motivation for the pool approach.
The Pool Approach. As explained in Section 3-8.2, the pool approach dispatches
one GT at full output and backs down 80 MW of steam output. It sets the price
at system marginal cost, $20/MWh, and it pays the GT $2,000/h as a side payment
to make up for dispatching it with a market price below its variable cost.
In competitive markets, economics predicts that a market price equal to marginal
cost will provide optimal incentives for suppliers and customers in both the short
run and long run alike. These economic conclusions are based on the assumption
that costs are convex, so it is not obvious that SMC pricing will prove optimal for
solving the problems caused by costs that violate this assumption. It is best to
proceed cautiously. Three benefits are desired of the market price. It should:
1. Induce efficient supply in the short run.
2. Induce efficient demand in the short run.
3. Induce efficient investment in the long run.

As is universally acknowledged, marginal cost, $20/MWh in this example, is

the wrong price to induce efficient short-run supply, and a side payment is needed.
This leaves the questions of short-run demand efficiency and long-run investment.
Does setting price equal to SMC induce an efficient demand for power? The
marginal condition seems right. Customers will see the true cost of their next unit
of consumption. But, this analysis tacitly assumes that demand is inelastic, otherwise
demand could not have been known before price was calculated as equal to marginal
cost. If demand is inelastic, then all prices “induce” the efficient level of demand

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-8 The Real-Time Market in Practice 281

Figure 3-8.2
Determining the
maximum market price
before starting a block-
loaded gas turbine.

or at least do not interfere with it. Every price is right for demand efficiency. To
test SMC pricing on the demand side, demand elasticity must be introduced so that
price matters.
Suppose demand is reduced by 1 MW for every $2/MWh increase in price. This
is a modest level of elasticity for a market as large as NYISO’s.6 With elastic
demand, there is a demand-side question of price. If the market price is raised, the
GT will not be needed. This will decrease total surplus by reducing the power
consumed and increase total surplus by avoiding the cost of the GT. When demand
is 20,001 MW at a price of $40, price need only be raised to $42/MWh to reduce
demand to 20,000 MW and balance the system. Because the GT is block loaded,
meeting the extra megawatt of demand instead of raising the price would cost $40/h
for the megawatt itself and 99 × ($40 ! $20)/h for the 99 MW of cheaper power
displaced by block-loading the GT. Clearly the price should be raised and the GT
not dispatched.
As the demand curve shifts to the right, price can be increased to keep the system
balanced, but at some price the demand that is being suppressed becomes so
valuable, that it is better to start the GT and serve the load. This point is reached
when the net consumer surplus of suppressed demand, which is equal to the area
of triangle B in Figure 3-8.2, equals the cost of side payments to the GT, given by
the area of rectangle A. This happens at a price of $109.44 when the demand curve
has shifted to the point where demand would be 20,045 MW at a price of $20/MWh.
From this level of demand until demand reaches 20,100 MW, the market price stays
at $20/MWh. The meaning of these price fluctuations will be discussed after
investment is analyzed, but in brief, SMC pricing often misses the required price
for demand efficiency by a wide margin.
Marginal-Cost Pricing and Investment. The third benefit expected of price
is inducement of the efficient mix of generating technologies. To check this, begin
with one change of assumption. Suppose the GT’s could be flexibly dispatched
with a constant marginal cost of $40/MWh. Suppose there are just enough steam
units to cover the load-duration curve up to a duration of 50%. Suppose the stock

6. If NYISO replaced its zero-elasticity demand for operating reserves with a realistic demand curve, it
might well produce this much elasticity in the demand for power.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

282 PART 3 Market Architecture

of GT’s leaves room for the price spikes required to cover the fixed costs of GT’s.7
In addition, suppose the fixed costs of steam units are $10/MWh greater than the
fixed costs of GT’s. In this case, Result 2-2.2 gives the following market equilibrium
condition.
FCsteam ! FCGT = (VCGT ! VCsteam ) × D*GT
For the given cost values, the market would be in long-run equilibrium. Steam units
would earn short-run profits of $20/MWh half of the time, and this would exactly
cover their extra $10/MWh of fixed cost.
Now imagine that the flexibility of the GT’s is removed and they revert to their
standard all-or-nothing nature. This decreases their value because it increases the
total cost of production relative to what it was when they were flexible. If the market
works properly it will send a signal indicating that slightly less GT capacity and
slightly more steam-unit capacity would be optimal.
What signal does marginal cost pricing send after the change from flexible to
block-loaded GT’s? Whenever demand does not exactly match a whole number
of GT’s, a steam unit must be backed down. And any time a steam unit has been
backed down, it can produce a small additional (marginal) increment of power at
the marginal cost of $20/MWh. Thus, marginal cost will almost always be
$20/MWh. Whenever this occurs, it reduces the short-run profits of steam units
to zero and the short-run profits of GT’s to minus $20/MWh. But the GT’s have
their profits restored through side payments and the steam units do not. The few
peak hours when GT’s and steam units both earn scarcity rents are left undisturbed
as is the capacity market if it exists. Thus GT’s continue to recover their fixed costs
as before while steam units lose nearly all the $10/MWh of scarcity rents they
previously collected when GT’s set the market price. This strongly opposes the
weak positive signal that should have been sent to steam units. In this example,
SMC pricing sends a grossly incorrect investment signal.

The Mystery of System Marginal Cost8

The Zero-Elasticity Case. If demand is totally inelastic then there is no question

of what total supply should be—it should equal demand. In this case, the optimal
dispatch simply minimizes production cost. In the optimal dispatch, there will
always be at least one unconstrained generator (if there is enough generating
capacity) except for the infinitely rare circumstance in which the dispatched
generators can supply only and exactly the demand. When there are several uncon-
strained generators they will all have the same marginal cost in an optimal dispatch.
Thus, when demand is inelastic, system marginal cost is always well defined and
equal to the marginal cost of any unconstrained generators.
In this case, SMC pricing doesn’t hurt demand, but it still sends the wrong signal
to investors. This is ignored in the hope that the long-run damage is smaller than
the short-run advantage of an optimal dispatch. In many cases the same optimal
dispatch could be supported by higher prices and no side payments. This would
7. The coverage of fixed costs may or may not include payments from a capacity market.
8. For a more rigorous treatment of this subject, see Cramton and Wilson (1998, 24).

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-8 The Real-Time Market in Practice 283

cause no distortion of demand or distortion of short-run supply, and it might be

better for investment.
The Elastic Case. With even a small amount of demand elasticity, price can no
longer be set equal to marginal cost without causing demand-side inefficiency. If
the demand curve is imagined as gradually moving from left to right, the price spike
in the above example occurs entirely while the output is stationary and exactly at
the full capacity of the steam units. Without any change in production, price
increases from $20/MWh to $109/MWh thereby holding demand constant in spite
of the moving demand curve. It is reasonable to say that marginal cost is undefined
at this point so none of these values conflict with it. But marginal cost cannot
possibly determine all of these prices or even the maximum price. Instead price
must be set equal to the marginal value of power to customers. In the elastic case,
it is still possible to define SMC pricing and it provides a useful benchmark.

Definition System Marginal Cost (SMC)

SMC is the marginal cost (MC) of all unconstrained units when supply equals
demand, where demand is computed with P = MC. Optimization of the dispatch,
given the output level, is assumed. If there is no output level satisfying these
conditions, SMC is undefined.

Result 3-8.1 SMC Unit Commitment with Elastic Demand Is Inefficient

If production costs are not convex and demand is elastic, SMC prices generally
fail to induce efficient short-run supply, short-run demand, and long-run invest-
ment. Side payments do not prevent any of these inefficiencies.

It is useful to define the pricing procedure illustrated in the block-loading example

as a second benchmark.

Definition System Marginal Value (SMV)

SMV is the marginal value of demand at the optimal output computed by taking
account of demand elasticity. (If marginal cost is defined at the optimal output,
then SMV = SMC.)

Contrary to SMC pricing, SMV pricing is right for demand but wrong for supply.
In the example, when price is set to $109/MWh to curb demand, it is much too high
from a short-run supply perspective. This is to be expected because whenever costs
are nonconvex, there is no assurance of a single efficient price. Also, there has been
no demonstration that SMV costs provide the proper long-run incentives when
coupled with the necessary side payments. There may well be a theoretically correct
pricing scheme and it may require different prices for loads and generators. While
interesting as a theoretical problem, more practical approaches should be pursued
first.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

284 PART 3 Market Architecture

Result 3-8.2 System-Marginal-Value Pricing Provides Efficient Demand Incentives

SMV prices induce efficient short-run demand but generally fails to induce
efficient short-run supply.

How Serious Are the Inefficiencies of SMC Pricing?

In the above example, the maximum short-run inefficiency occurs when the GT
is block-loaded to serve 1 MW of load. This decreases quadratically to zero when
about 45% of the GT’s output is needed. The average over the whole period of
increasing need for this GT, provided load increases linearly, is only in the neigh-
borhood of $300/h for a market that might be selling over $300,000/h worth of
power. That is, 1 part in 1000, and this circumstance probably happens infrequently.
Of course, there are other cost nonconvexities, but it appears that even when
handled poorly, they do not cause much trouble.

Where Does the NYISO’s Non-SMC Pricing Lead?

NYISO’s pricing innovation leads in two directions. First, it leads toward complex-
ity. It introduces an ad hoc rule for pricing in the RT market. Because NYISO
compromised with FERC, adopting FERC’s initial request for SMC pricing under
some circumstances but not others, it adds complex new distinctions. The NYISO-
FERC innovation also introduces different pricing philosophies for the RT and DA
markets without any supporting theory. Together these seem to confirm the standard
concerns over the arbitrariness and impenetrability of real-world power pools.
Viewed more broadly, the pricing innovation may be seen as leading toward
greater simplicity. NYISO saw the benefits to its deviation from SMC pricing as
rooted in a single feature. Its price, as opposed to an SMC price, would pay the
cost of the GT. In its view, this produces several benefits. First, it avoids side
payments and the associated uplift. This prevents one case of market power and
the distortion of transmission pricing. Second, it focuses the payment of the cost
more closely on those who caused the cost.
It also produced a higher price that is somehow more right on average. This
point is intuitively correct but was probably not understood in detail. While SMC
prices send the right signals when the marginal-cost curve is continuous, they miss
the costly block-loading event. They miss it because price cannot be raised to
infinity for a split second thereby tracking marginal cost at the instant the generator
is block-loaded. SMC pricing cannot handle this type of discontinuity in marginal
cost, so it ignores it. NYISO’s cruder pricing scheme does not get tricked by missing
any infinitely high and narrow price spikes.
NYISO has moved away from the complex details of standard SMC pricing
in this one case, and has moved even further from the details of SMV pricing.
Probably, without realizing it, they are giving up on the optimality for which these
approaches strive. Instead they are moving toward a simpler but more robust form

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-8 The Real-Time Market in Practice 285

of pricing, the type of pricing used in every other market. They have moved toward
a price that cannot support the optimal dispatch but can pay for the cost of the
generators it dispatches—they have moved toward a power-exchange approach
to pricing.

3-8.4 THE POWER EXCHANGE APPROACH

Like the power-pool approach, the power-exchange approach is still ill-defined
and problematic. As the power-pool approach needs some power-exchange features,
so a power-exchange approach may need some power-pool features. It may need
two- or three-part bids. Perhaps not, but that is a matter that must be determined
empirically. Only clever design and numeric evaluation followed by experimentation
will yield the answer.
One factor above all others will push markets in this direction. Real-time pricing
of demand will eventually produce a large demand-side response that does not
operate through bids. Then, if the system operator evaluates the bids it has and
determines a market-clearing price and sets it as the market price, the market will
refuse to clear. The new price will provoke a significant demand-side response at
some of the most awkward and delicate times. This is already happening in PJM
but not because of the demand side. That is still incapacitated, but imports and
exports play an active role and do not necessarily bid into the PJM market. (The
NYISO noted the discouragement of “external suppliers from delivering RT im-
ports.”).
When the system operators see load increasing, they look at their bids, which
may be few and high, and choose a bid price that is a few hundred dollars above
the current price. They often find this is followed by a large influx of external power
which causes them to reevaluate and set a new lower price. The imported power
then leaves, or the exports resume, and the system operator is again forced to raise
the price.
System operators at PJM have learned to take account of un-bid supply elasticity,
but the rules still require them to set price at the level of some particular supply
bid. This is a contradictory approach. The system operator may choose not to take
a $400/MWh bid because it believes this will induce too great an increase in net
imports and instead set the price at a $200/MWh bid price. The system operator
can adjust the purely bid-determined price down by $200, but it is not allowed to
adjust it by a mere $100 if there is no supply bid at $300. It cannot set the price
at $300 even if that would more accurately balance the system. If large price
adjustments are made outside of the bidding rules, why can’t smaller adjustments
be utilized?
The un-bid elasticity of the supply and demand curves causes two problems
for the power-pool approach. First it forces the system operator to invent fictitious
bids that represent the un-bid portions of these curves. While this can be done
scientifically and without bias, it cuts against the grain of the underlying philoso-
phy—enter hard data from market participants and compute the optimum. The
second problem is more fundamental. The elasticities are not instantaneous. If price

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

286 PART 3 Market Architecture

Figure 3-8.3
Market clearing at a bid
price and clearing in-
between bid prices.

is based on current SMV, it may cause an unwanted supply or demand response

half an hour later due to the dynamics of the supply and demand elasticities. Just
as supply-side bids account for ramp rate limits, low-load limits, and other lags
and constraints, so the fictitious demand-side bids should account for lags and
discontinuities in demand if optimization is to be successful.
The result of increasing un-bid elasticity will be an order of magnitude increase
in optimization complexity, a new and wide opening for arbitrary decisions on the
part of the system operator.

Improving Prices
The North-East ISOs tend to compute and announce prices every five minutes as
if they meant them, but they never use them to pay for power. Instead they replace
all 12 prices every hour with a single average price used in every five minute
interval. The originally announced prices probably fool only a few. The use of the
five-minute price is to instruct generators as to the desired level of output. In other
Table 3-8.1
words, these prices do not work through the normal economic channel of increasing
Price
and decreasing revenue according to price times quantity. Instead they are regulatory
per
Time MWh instructions which the suppliers translate according to their bid curves. These
instructions are backed by penalties in the form of cancellation of make-whole side
1:00 $24
payments. In PJM not all generators are subject to control; in fact, real-time side
1:05 $24 payments may affect only a minority.
1:10 $24
Use Honest 5-Minute Prices. Five-minute prices matter most when the price
1:15 $24 is changing rapidly. At other times, it might be sufficient to compute price every
1:20 $24 half hour or every hour. Consider an hour with a large price change as shown in
1:25 $24 Table 3-8.1. When the system operator sets a price of $48, the market participants
1:30 $24 will guess that the average for the hour will be about $30 and respond to that
instead. This will mean a 15-minute delay before many participants begin respond-
1:35 $24
ing to these 5-minute prices. If five minutes matters, then the delay undermines
1:40 $24 both control and the precision of the dispatch.
1:45 $48
Use the Best Price. The next improvement is to learn to predict un-bid supply
1:50 $48 and demand responses and use these predictions along with bids to set prices more
1:55 $48 accurately. Stop restricting RT price to bid values, and instead set the price that

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-8 The Real-Time Market in Practice 287

will do the best job of balancing supply and demand. Consider the dynamics of
the supply and demand response.
Use a Downward Sloping Demand for Reserves. A known elasticity makes
control through price adjustments easier. Zero elasticity is inappropriate for operat-
ing reserves. Replacing this assumption with a downward-sloping demand function
for operating reserves will make balancing easier. When reserves are short, the price
automatically rises and this calls forth more bid-in supply and imports and reduces
exports and demand. It also helps to reduce market power by increasing the elastic-
ity of demand. If 10% is the current requirement for all types of reserve, then they
must be worth nearly $1,000/MWh up to the 10% level. Surely another few percent
must be worth something, so the demand curve should go to higher levels of
reserves at low prices as well as to lower levels of reserves at high prices. This
simple change should dampen prices spikes and make the operator’s job easier.
The argument that elasticity will jeopardize reliability is without foundation. It
depends entirely on what demand curve is selected; some would increase reliability
dramatically. Buying more reserves when they are cheaper and fewer when they
are dear will reduce the cost of reliability.
Purchase Balancing Services. The system operator buys balancing services
directly in two ways: by buying regulation and by buying spinning and nonspinning
reserves. Both are limited in the uses to which they may be put and do not cover
many normal circumstances that worry system operators. This concern is expressed
in many ways, such as tying the startup insurance offered by power pools to
following the dispatch. Also, PJM runs a full-scale DA simulation after the DA
market to determine what generators to start. Limits are placed on the flexibility
of external trades. NYISO confiscates the power of generators that overproduce.
While there are many strategies, system operators still find the job difficult, espe-
cially at awkward times such as the morning ramp which occurs just after last-
minute exports are scheduled.
These difficulties exist partly because the balancing service is thought of in terms
of supply and demand instead of in terms of their derivatives (rates of change).
Ramping is a separate service with a separate cost. This cost includes the wear and
tear on machines from changing output as well as startup costs and the costs from
dispatching around low-operating limits. One approach to better control would be
to pay for ramping, that is, for changes in output between successive five-minute
intervals. This price could be either positive or negative.
Sometimes, it is necessary to ask a cheap generator to back down while keeping
a more expensive generator on line. There should be a market for the backing-down
service, separate from the energy market, and perhaps linked with the reserve
markets.
Accepting Bids. For the most part, the previous suggestions ignore bids because
an RT exchange does not utilize bids—it is a Walrasian auction (see Chapter 3-6).
But accepted supply bids discourage market power. If a large generator has sold
all of its output before real time, it will still have the ability to raise the market price,
perhaps significantly, but it cannot profit from it. If it raises the market price, the
increase does not affect its payment on forward contracts. If it withholds power

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

288 PART 3 Market Architecture

it must buy it from the market, so when it withholds it wants a lower, not a higher
price. For this reason the system operator should encourage as much trading as
possible before real time because it is more vulnerable to market power in real time.
This means the RT exchange should be supplemented with an hour-ahead
exchange or even with a forward exchange that is capable of accepting bids at any
point in time. This will minimize the extent of the RT market without the use of
penalties.
Unaccepted forward bids also provide the system operator with information
about the supply curve, although a generator may not follow its own bid if it is
above marginal cost. In spite of this, bids will help set an accurate RT price. In order
to encourage bidding and provide certainty to generators, more complex bids could
be allowed. This does not turn an exchange into a pool; only side payments can
do that.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 Chapter 3-9
The New Unit-Commitment Problem

Let us not go over old ground, let us rather prepare for what is to come.

Marcus Tullius Cicero

(106–43 B.C.)

T HE OLD PROBLEM ASKS WHICH UNITS SHOULD BE COMMITTED; THE

NEW PROBLEM ASKS WHAT MARKET DESIGN WILL BEST SOLVE THE
OLD PROBLEM. The old problem was solved by collecting data on all the genera-
tors and applying the techniques of mathematical programming. The new problem
might be solved by a market designed to induce generators to voluntarily and
accurately provide this same data. The market coordinator could then purchase
power from the generators identified by the old algorithm. This is the power-pool
approach. A power exchange is an alternative approach which pretends the old
problem does not exist.
It seems impossible that ignoring the old problem could be the best way to solve
it, but most market architectures ignore just such complex commitment problems.
When the market coordinator ignores the problem, the suppliers take it up, and they
may do a remarkably good job. Although the programming techniques used to solve
the old problem are astoundingly complex, most generators can get the right answer
on most days simply by looking at the calendar. If you have a baseload plant and
it’s summer, keep your plant committed. If you have a peaker, don’t look at the
calendar, just watch the real (RT) price—day-ahead (DA) forecasts are not needed.
When it really matters, on the hottest days, every supplier knows to commit. But
those who commit units for individual suppliers will do much better. They will
have years of experience and the necessary resources. Moreover, they may have
access to an exchange that uses two-part prices or multiple rounds of bidding. To
beat a good exchange market, a pool must be very good indeed.
Two central concerns have motivated the power-pool approach: efficiency and
reliability. Committing the wrong units costs more, but this problem is limited by
the magnitude of the cost involved and the efficiency of markets without centralized
unit commitment. Reliability is more of a wild card. Perhaps a decentralized market

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

290 PART 3 Market Architecture

would be more prone to occasional undercommitment of plants that are too slow
to start. This could leave the system operator without the necessary resources and
cause a blackout. As with efficiency, this suggestion has not been subject to
quantitative analysis, so the new unit-commitment problem remains unsolved.
The old power pools in the Northeast have adopted power-pool markets, while
the Western markets have adopted or proposed power exchanges. The choice of
design has been historically determined, at least partly because no theoretical or
empirical evidence is available. Considering only one well-defined part of the
problem, efficiency in the context of a workably competitive market, Professor
William Hogan has said:
I am not prepared to argue for or against the need for unit com-
mitment without doing more work. It is neither simple nor obvious
that it is important or not important. If I had to bet, I would go
60/40 in favor of it being important.1
This chapter provides a framework for thinking about the economics of the unit
commitment problem in the context of a competitive market. It suggests that some
problems appear small while others are difficult to assess. It confirms Professor
Hogan’s view that the new unit commitment problem is difficult and our knowledge
is sketchy.
Chapter Summary 3-9 : The cost of committing units is about 1% of retail costs,
and individual generators can come quite close to minimizing this without the help
of a central computation—a power pool. Though generators will sometimes renege
on commitments made in a DA exchange market, they are least likely to do this
when most needed, for that is when the RT price will be highest.
This does not rule out the possibility of improving on the power-exchange
design. But when a first-order approximation is accurate to 1%, a twenthieth-order
approximation may not be required (power pools use bids with at least that many
parts). Alberta allows two-part bidding, the second part being a minimum run-time
limit. While completely artificial, in simple cases, it can substitute perfectly for
bidding a startup cost. The crucial lesson is: Suppliers may not bid the literal truth,
but with sufficient competition and a bit of flexibility, they bid in a way that makes
their operation and the market efficient.
Section 1: How Big Is the Unit-Commitment Problem? Startup costs are
the main costs of commitment. Typically, these amount to less than 1% of retail
costs. Because startup costs are ten times smaller than fixed costs and have first
claim on scarcity rents, marginal-cost prices often send the right signals for unit
commitment.

1. Personal communication, January 5, 2002.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-9 The New Unit-Commitment Problem 291

Section 2: Unit Commitment in a Power Exchange. Four market designs

are considered in the context of a single market structure in which startup costs
are not covered by marginal-cost prices. Case A demonstrates that a power exchange
has no competitive equilibrium in the classic economic sense. Case B demonstrates
that an auction without startup-cost bids or side payments can have an efficient
competitive Nash equilibrium in spite of lacking a classic competitive equilibrium.
Case C reconsiders the simple power exchange design and finds the Nash equilib-
rium which proves to be inefficient but competitive. Case D introduces the possibil-
ity of reneging, that is, failing to generate the power sold in the DA market. This
leads to greater overcommitment in the DA market followed by reneging to the
point of an efficient dispatch.
Section 3: Investment Under a Power Pool. A power pool includes side
payments that reduce risk for generators that are accepted in the DA market and
that produce according to the supply curve they bid. If all generators bid honestly,
the dispatch will be efficient but investment incentives will be distorted. Types
of generators that receive more insurance payments will be encouraged to over-
invest. Like the short-run inefficiency of the power exchange, the long-run ineffi-
ciency of the power pool will be small.

3-9.1 HOW BIG IS THE UNIT-COMMITMENT PROBLEM?

A power exchange does not provide a unit-commitment service, but a power pool
does. Either can set a single price at any given time, a few prices in different zones,
or thousands of prices, one at each electrical bus in the system. This chapter only
considers markets without congestion in order to focus on the unit-commitment
problem. Exchanges and pools were defined in Section 3-3.1, but key parts of those
definitions are repeated here for convenience.

Definitions Power Exchange

A centralized market that trades energy at a single price (per location) at any
given time and location and does not make side payments. An exchange may use
multipart bids.
Power Pool
A centralized market that trades energy and uses side payments that depend on
the suppliers’ bids. Side payments are made only when the pool price is too low
to cover the startup or no-load costs of an accepted bidder.

The purpose of power-pool complexity is to solve the unit-commitment problem,

but the magnitude of that problem has not been assessed in a market context. Prior
to deregulation there was only one unit-commitment problem—the problem of when

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

292 PART 3 Market Architecture

to commit and de-commit a system’s generating units. That problem remains, but
it is not the relevant problem for market design. The new problem is: What market
design will induce the most efficient commitment of generating units. This problem
is different in nature and broader. Sections 3-9.2 and 3-9.3 examine questions of
market design, while this section investigates the magnitude of the old problem
in the market context.

Definition The (New) Unit-Commitment Market Problem

What market design will induce the most efficient commitment of generating
units? “Efficiency” should include the transaction costs of the system operator
and suppliers and the reliability consequences of commitment errors.

The Magnitude of Startup Costs

Startup costs are usually found in the range between $20 and $40/MW.2 Generators
that serve baseload start much less often than once per day, and very few generators
start more often. Hydrogenerators have exceptionally low startup costs. As a rough
estimate, there may be half a megawatt of startup per megawatt-day of load.
If the average startup cost is $30/MW, then the cost of startups is $15/MW for
each MW-day of load. That is $15 per 24 MWh, or about $0.60/MWh. Retail power
costs about $80/MWh in the United States, so startup costs are 3'4 of 1% of retail
costs. (Preliminary estimates for PJM indicate its startup costs may be closer to
half this much.)
A significant amount of this cost is covered by
scarcity rents from marginal-cost pricing.3 If approxi-
Units to Measure Startup Cost
mately 1'3 of startup costs are subject to unit-commit-
Startup costs are measured in dollars per megawatt
ment problems, and these problems add 50% to these
of capacity started. For a plant that is started once
per day, the cost flow is most conveniently stated in startup costs, then the total waste of funds would be
$/MWday. When used in formulas, startup costs are roughly '
1
6 of the cost of startups. This inefficiency
converted to $/MWh by dividing by 24. would raise retail rates about 1'8 of 1%. Because of the
Duration should be expressed as a pure number assumption that the market increases the costs of the
when used in formulas but can be converted to problematic startups by 50%, this should be an upper
hours per day, for convenient interpretation, by bound on the magnitude of the startup problem.
multiplying by 1 in the form (24 hours ' 1 day). Thus Test examples involving two or three generators
a duration of 0.1 equals 2.4 h ' day. usually demonstrate inefficiencies on the order of 1%
if their startup costs are a realistic fraction of total cost.
But startup inefficiencies are inversely proportional to the number of generating
units in the market. Together these give a back-of-the-envelope estimate of ineffi-
ciency on the order of 1 part in 10,000. So far, advocates of power pools have not
produced a contradicting estimate, but the question is still unanswered as no

2. See Hirst (2001). No-load costs are another source of difficulty and are several times larger than startup
costs. They are proportional to the length of time a generator runs and consequently easy for generators
to include in their bids in a power exchange. They may present more of a problem in conjunction with
startup costs.
3. Scarcity rent (see Section 1-6.6) is revenue minus costs that vary with output. After startup and no-load
costs are subtracted, short-run profits remain.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-9 The New Unit-Commitment Problem 293

theoretical proof has been given and no calculations have been made on actual
power systems.4

Result 3-9.1 A Power Exchange’s Unit-Commitment Inefficiency Is Less Than 1%.

Startup costs, the principle source of unit-commitment inefficiency, amount to
roughly 1% of the retail cost of power or less. A significant portion of this cost
is covered by marginal cost pricing, and markets without central unit commitment
can, in principle, solve the unit-commitment problem quite efficiently. Conse-
quently the inefficiency caused by nonconvex production-costs may be less than
'100 of 1%.
1

How Marginal-Cost Bidding Can Cover Startup Costs

One factor mitigating the startup-cost problem is scarcity rents. These must be great
enough to cover fixed costs which are more than ten times greater than startup costs.
Because startup costs have first claim on scarcity rents, they are quite often covered
by them. Example 1 explains how this works and why it sometimes does not.
Example 1A: Startup Costs Covered Completely. Consider a market with
peakers and midload plants that have the costs shown in Table 3-9.1. In long-run
equilibrium, the market will build peakers up to the point where they just cover
their fixed costs. This cost recovery must occur when all peakers are in use, and
the price is driven above $50/MWh by the system operator’s demand for operating
reserves in short supply. The price of energy during this period could be moderate
or extreme, but in either case, it will be just sufficient to cover the peaker’s fixed
costs of $120/MWday (see Chapter 2-2).
Table 3-9.1 Cost Assumptions of Example 1
Variable Cost Fixed Cost Startup Cost
Type of Supplier (per MWh) (per MWday) (per MWday)
Peaker $50 $120 $0
Midload plant $30 $240 $40

During this same period, midload plants will also earn $120/MWday from prices
above $50/MWh, enough to cover half of their fixed costs. The remaining half plus
startup costs total $160/MWday, and this must be recovered from the difference
between the $50/MWh variable cost of peakers and the $30/MWh variable cost
of midload plants (see Chapter 2-2). This $20/MWh revenue stream is available
whenever peakers are running, so peakers must run for 8 hours per day (160/20)
if midload plants are to recover these costs. Consequently midload plants will be
built up to the point, but not beyond the point, where peakers are needed 8 hours
per day on average.
If load had the same pattern every day, high prices set by peaker marginal cost
and system operator demands would allow midload plants to recover their full

4. The author has spent considerable time checking proposed examples, one of which is posted on
www.stoft.com.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

294 PART 3 Market Architecture

$280/MWday of fixed and startup costs by bidding their marginal costs. There
would be no need for complex bids in either a power exchange or a power pool.
Example 1B: Costs Are Still Covered on Some Low-Load Days. Load
fluctuates from day to day, sometimes spending many hours above $50/MWh and
sometimes spending no time at all. On low-load days, this removes $120/MWday
of fixed-cost recovery for both peakers and midload plants. The shorter duration
of peaker use on such days reduces midload cost recovery still further. When
peakers run only half as long as they do on average, midload plants will earn only
$80/MWday of scarcity rent. This will still cover all of their startup cost
($40/MWday) and a bit of their fixed cost without any need for complex bidding.
As long as scarcity rents cover startup cost and just a little more, the unit will be
started, the startup paid for out of scarcity rent, and the remaining rent used to cover
fixed costs. In this sense, startup costs have first claim on scarcity rents.

Result 3-9.2 Marginal Cost Prices Can Solve Some Unit-Commitment Problems
A market in which generators have (nonconvex) startup costs may still have a
normal competitive equilibrium that produces an efficient dispatch at all times.

What Triggers the Startup Inefficiency? Only on days when there is no price
spike and peakers run for less than two hours (1'4 of normal) will the midload plants
be faced with the dilemma caused by startup costs. On these days they must bid
above marginal cost in a power exchange, and they must bid marginal cost and
startup cost separately in a power pool. (Of course they can and will bid startup
costs every day in a power pool, but it does not matter on most days.) Section 3-9.2
considers how well a power exchange performs on such days.

3-9.2 UNIT COMMITMENT IN A POWER EXCHANGE

When California adopted a power exchange, advocates of this design suggested
that suppliers would find it so simple to bid that they could “sit back and read a
book” while the auction proceeded. In spite of subsequent simplification, no optimal
bidding strategy was ever presented. To this day, no one has explained the competi-
tive bidding strategy for such an auction. Example 2C shows the startling complexity
of such a strategy in a trivially simple market.
Fortunately, as explained above, the inefficiencies that are likely to result from
an incorrect solution to the unit-commitment problem are small, and marginal-cost
pricing often solves the problem. When it does not, some rule of thumb will most
likely produce a reasonable outcome, but this has never been demonstrated. A
second problem also causes alarm. Some advocates of power pools see ominous
signs in a power exchange’s lack of a classic competitive equilibrium. They assume
the worst, while advocates of power exchanges assume the best.
This section presents four elementary examples that address the main points
of both positions. These analyze the behavior of a market in which load is low and,
as a consequence, midload plants are unable to cover their startup-costs by bidding

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-9 The New Unit-Commitment Problem 295

only marginal costs. Each example uses the cost structure of Examples 1A and 1B
coupled with a specific low-load condition. The examples differ slightly in auction
design and post-auction trading activity.
Table 3-9.2 Bidding above Marginal Cost in a Power Exchange
Example Auction Bidding Outcome
2A Walrasian — No equilibrium
2B two-part Deterministic Efficient
2C one-part, no reneging Random Inefficient
2D one-part, with reneging Random Efficient

Example 2A demonstrates the absence of a classic competitive equilibrium.

Example 2B shows that allowing bidders to state minimum-run times gives this
market an efficient, though not classic, competitive equilibrium. Examples 2C and
2D utilize the simplest standard power exchange design. Example 2C assumes that
the outcome of the DA market is enforced. In this case the equilibrium is inefficient
due to occasional overcommitment. Example 2D assumes that generators can renege
on their DA contracts, but they must buy back the necessary power in the RT
market. This results in an efficient competitive equilibrium but one with consider-
able complexity and randomness.

Example 2A: Absence of a Classic Competitive Equilibrium

A classic competitive equilibrium is defined by the outcome of a “Walrasian
auction.”5 An auctioneer calls out prices. If a certain price elicits more demand
than supply, he tries a higher price, and vice versa. To define a competitive equilib-
rium, economists consider what the outcome would be if both suppliers and
demanders assumed they could have no influence on the auctioneer’s price. Under
this assumption, if there is a price at which supply equals demand, that is the
competitive equilibrium.
The assumption on trading behavior is referred to as being a “price taker.” If
there is a quantity, Q, and a (competitive) price, P, that satisfy the following two
conditions, supply equals demand. Suppliers could not increase profits by selling
a quantity different from Q at price P. Demanders could not increase their satisfac-
tion (utility) by buying a different quantity Q at price P.

Definition Classic Competitive Equilibrium

A market condition in which supply equals demand and traders are price takers.
More precisely, there is some price P and quantity Q such that suppliers cannot
increase their profit by selling a different Q at the price P, and demanders cannot
increase their utility by buying a different Q at the price P. (A competitive
equilibrium is often neither a Nash equilibrium nor a likely outcome.)

5. The competitive equilibrium of economics will be called “classic” to distinguish it from efficient
competitive Nash equilibria that are not Walrasian equilibria (see Example 2B).

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

296 PART 3 Market Architecture

Figure 3-9.1
Load profile for
Example 2 (A, B,
C, and D).

Economics predicts efficiency for markets that have a classic competitive

equilibrium and price-taking suppliers and warns of possible inefficiency for
markets that do not. The failure of power exchanges to have a classic competitive
equilibrium is sometimes seen as a serious flaw.

Assumptions of Example 2A. There are two 100-MW peakers and two 100-MW
midload plants. Demand, depicted in Figure 3-9.1, is above baseload, requiring
some output from a midload plant for five hours and more than 100 MW above
baseload for only one hour. The lower load-slice has an average duration of three
hours, and the upper load slice has an average duration of one half hour.
Table 3-9.3 Assumptions of Example 2 (A, B, C and D)
Type of No. of Capacity Variable Cost Startup Cost Avg. Duration
Supplier Suppliers (in MW) (per MWh) (per MWday) of Load Slice
Peaker 2 100 $50 $0 0.5 h
Midload 2 100 $30 $40 3.0 h

To show that the market of Example 2A has no equilibrium, it is only necessary

to look for it where it would necessarily be found. Because classic competitive
equilibria are always efficient, if one is to be found it must correspond to the most
efficient dispatch. That consists of dispatching a midload plant for the bottom load
slice and a peaker for the top load slice. The difficulty with this dispatch is that
both plants spend time at fractional output levels.
For the peaker to be in equilibrium at a 50-MW output level, the market price
must be $50/MWh, its variable cost. Otherwise, if the price were higher, it could
increase its profits by producing at full output or, if the price were lower, by
producing at zero output. Similarly the midload plant requires a market price of
$30/MWh to be in equilibrium with a fractional output level. The only possibility
for a classic competitive equilibrium is marginal-cost pricing: a price of $30/MWh
between 2 PM and 7 PM, except for the peak hour when it would be $50/MWh.
Unfortunately, this allows the midload plant to recover only $20/MW for startup
costs (during the peak hour), so it refuses to run. The optimal dispatch is not an
equilibrium for any price path, so there is no competitive equilibrium.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-9 The New Unit-Commitment Problem 297

Result 3-9.3 A Power Exchange Lacks a Classic Competitive Equilibrium

If price is set equal to marginal cost at all times, as required by a competitive
equilibrium, some generators that are part of the optimal dispatch may not cover
their startup costs. Frequently this problem prevents a DA power exchange from
having a classic competitive equilibrium.

How a Power Pool Fixes the “Problem”

Power exchanges frequently lack a classic competitive equilibrium, but, as will
be demonstrated shortly, they do not lack a Nash equilibrium. The classic competi-
tive equilibrium would be perfectly efficient, but the Nash equilibrium might be
extremely efficient.
Using the same example, a power pool would set the market price to variable
cost at all times. This means $30/MWh between 2 PM and 7 PM, except for the peak
hour when it would be $50/MWh. As noted, the midload plant loses $20/MW,
which comes to $2,000. It is offered a side payment of $2,000 on the condition
that it start up and “follow the dispatch,” that is, ramp up and down as directed.
When only part of the midload plant’s output is needed, the market price equals
its variable cost and so it is indifferent as to what output it produces. With this
arrangement, a power pool induces the optimal dispatch in spite of the market’s
lack of a classic competitive equilibrium.
Notice that the power pool solves two problems: (1) how to get the midload
plant to commit, and (2) how to get the midload plant to ramp up and down as
directed. The ramping problem may be more serious than the unit-commitment
problem, but so far neither has been assessed.

Result 3-9.4 A Power Pool with Accurate Bids Induces the Optimal Dispatch
Power pools lack a classic competitive equilibrium. But if bidding is honest,
power-pool side payments, together with marginal-cost prices, will induce the
optimal dispatch. With enough competition between suppliers, bidding should be
honest. (This result ignores demand elasticity and long-run incentives.)

Example 2B: Efficiency with Two-Part Bidding.

Walrasian auctions are found in economic texts but rarely, if ever, in commerce.6
Moreover, the concept of a Walrasian equilibrium used to define a classic
competitive equilibrium is not used in auction theory—the Nash equilibrium
concept is used. The Walrasian equilibrium concept is not intended to predict how
a market will actually function, and it does not. A highly monopolistic market may
have a competitive Walrasian equilibrium, but that does not indicate it will produce
a competitive outcome. The Nash equilibrium is a prediction of a market’s outcome.
This example finds the Nash equilibrium of a power exchange and shows that,
while it has no classic competitive equilibrium, it does have a perfectly efficient
6. An RT power exchange resembles a Walrasian auction but does not wait for a market clearing price
before allowing trade to take place.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

298 PART 3 Market Architecture

Nash equilibrium. In this case, lacking a classic competitive equilibrium is not a

problem.
Alberta uses a simple auction that handles the present example quite nicely.
Suppliers bid an energy price and a minimum run time. Example 2B replaces the
minimum run time with a minimum quantity of energy for convenience. Bids will
only be accepted if at least this much energy is purchased. As with a power pool,
the system operator needs a guarantee of performance. The settlement rule states
that a generator will only be paid for the quantity accepted if it follows the system
operator’s dispatch.
The cost-minimizing set of bids is accepted. When there is a tie, the winner is
chosen by the toss of a coin. The price paid at every point in time is the highest
accepted bid price at that time.
With this auction design, the example market has an equilibrium, which is
defined by midload plants bidding $40 and peakers bidding $50. Midload plants
will specify, as part 2 of their bid, that their minimum output is 300 MWh. Because
of tie bids each generator will be selected half the time, and because they just cover
costs, generators will be indifferent as to whether or not they are accepted. No
generator can do better by changing its bid, so this is a Nash equilibrium. No
generator has exercised market power so it can be termed a competitive equilibrium
although it does not fit the classic economic definition of one. The market price
produced by this auction is a competitive price, though in the classic economic
sense it is not a “market-clearing” price.
Example 2B shows that the lack of a classic competitive equilibrium does not
imply a power pool is needed to achieve a perfectly efficient equilibrium. This
example is important because it achieves the effect of a power pool without the
side-payments of a power pool and without paying any attention to realistic bidding.
Generators bid a false minimum energy and do not bid startup costs.
Note that this two-part-bid auction is more complex than a power exchange
that uses one-part bids. The two-part bids require a more complex evaluation
process, but the result is still transparent. The crucial point is that this two-part-bid
auction is not a power pool because it pays all suppliers the same price at every
point in time and makes no side payments. The single market price is high enough
to cover the costs of all accepted bids.

Result 3-9.5 A Power Exchange Can Be Efficient Without a Competitive Equilibrium

In a market that has no competitive equilibrium, a power exchange may still have
an efficient and competitive Nash equilibrium.

Example 2C: Inefficiency with One-Part Bids

The question remains: How well does a pure power exchange solve the unit-
commitment problem. Example 2C answers this question by computing the Nash
equilibrium for a power exchange operating in the market structure under consider-
ation.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-9 The New Unit-Commitment Problem 299

In a pure power exchange, suppliers can only state their capacity and a price
curve. If suppliers bid the same price in a pure power exchange as in the Alberta
auction, the outcome would be very different. The two midload bids would win.
One of these suppliers would lose the toss of the coin and be accepted for the top
load slice because they cannot specify this as unacceptable.7 This generator would
be paid $40/MWh, which would cover only $5/MW out of its $40/MW startup cost
($10/MWh rent times half a megawatt-hour of output per megawatt of capacity).
Without the peaker to set a high price, the generator in the bottom slice would also
not recover its startup cost.
Generators will need to change strategy to suit the new auction rules. Peakers
will still compete and bid $50, but the midload generators will bid randomly
according to a specific set of probabilities. In game theory, a random strategy is
termed a mixed strategy and a deterministic strategy is termed a pure strategy.8
A mixed strategy specifies a probability for every possible bid that a supplier can
make. The equilibrium strategies have been computed for the present auction and
are presented in Figure 3-9.3. This pair of strategies forms a Nash equilibrium
because neither supplier can increase its expected profits by changing its strategy,
provided the other supplier holds fast.
Even though the two midload suppliers are identical, they use different
strategies—an equilibrium characteristic that will be discussed shortly. Sixty percent
of the time supplier A bids just under $50, while 60% of the time supplier B bids
any amount over $50. How far over does not matter because any amount ensures
that supplier B will not be selected. Supplier A is always selected. When B opts
out (by bidding high), A is quite profitable, but when B bids seriously, B underbids
A 80% of the time. In this case A loses money and B is profitable. All told, the two
suppliers compete away all of their profits, which is not surprising because this
is basically a Bertrand model (pure price competition) and such models lead to
perfect competition.
Although the equilibrium is competitive, it is not efficient. Forty percent of the
time both midload suppliers win. Having a midload supplier serve the top load slice
costs $55/MW, while using a peaker costs only $25/MW.9 Because this inefficiency
occurs only 40% of the time, the average excess cost is $1,200 per day
($30 × 100 × 0.4).
The minimum cost of supplying the total load in this example is $2,500 for the
top load slice and $13,000 for the bottom load slice, so the inefficiency is nearly
8% of wholesale cost or 3% of retail cost. This is unrealistic for several reasons,

7. A one-part-bid auction, though it involves only price and quantity like a Walrasian auction, works
differently. The auctioneer does not call out a price and then allow trade at that price. Instead it sets a price
and accepts certain bid quantities from each supplier. In this way a set of trades can be determined without
a market-clearing price.
8. Mixed strategies are common in the real world, and although difficult to compute, even children learn
to employ them without instruction. Any finite game without a pure strategy Nash equilibrium has a mixed
strategy Nash equilibrium. This game has no satisfactory pure strategy for a midload plant, so it must have
a mixed-strategy Nash equilibrium. All power auctions are finite games because bids are rounded to the
penny and to the megawatt.
9. Cost per megawatt is startup cost plus an average of 0.5 MWh of energy times variable cost for each
megawatt of the peak load slice.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

300 PART 3 Market Architecture

Figure 3-9.2
Randomized bids from
two midload plants in a
power exchange.

two of which are the exaggeration of startup costs and the underestimation of the
number of generators. In an actual market, startup costs are closer to 1% than to
25% and the number of generators is closer to 400 than to 4. Inefficiencies due
to problems with startup cost decline in proportion to the number of suppliers in
the market. Correcting for these two effects could reduce the estimate to less than
'100 of 1%.
1

This auction has two equilibria, one in which generator B opts out by bidding
high 60% of the time, and another in which the roles are reversed. This could lead
to further coordination problems, but it seems likely that each will stumble into
a particular role and realize it does not pay to switch. Still, when such a simple
problem generates such complex behavior, it raises questions about how well a
large market will coordinate.

Result 3-9.6 A Power Exchange Has a Nearly-Efficient Nash Equilibrium

When a power exchange has no classic competitive equilibrium, it still has a Nash
equilibrium. Suppliers with a commitment problem will use randomized strategies
which will cause some inefficiency. In a market with several hundred suppliers
the inefficiency may be near 1'100 of 1% (1 part in 10,000).

Example 2D: Efficiency with One-Part Bids and Reneging

Midload suppliers in Example 2C overcommit 40% of the time. This would tend
to increase reliability, but the reliability of a one-part auction cannot be properly
investigated without allowing a second round in which generators can renege. To
renege means to fail to start up and produce power sold in the DA market. When
a generator reneges, it must buy replacement power in the RT market to cover its
sales in the DA market. Example 2D extends Example 2C by allowing generators
to renege.
When both midload generators are accepted, the one with the higher bid will
be given the top load-slice and will fail to cover its startup cost. In this situation
the generator might renege and leave the system operator with insufficient resources
in real time. To address this concern, a power pool guarantees that suppliers
accepted day ahead will not lose money if they produce the amount accepted.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-9 The New Unit-Commitment Problem 301

Figure 3-9.3
Mixed strategy for
midload generators in
Example 2D.

When a midload generator reneges, it will be required to buy replacement power

at a cost of $50/MWh. (Competition between the peakers will hold the price down
to the competitive level in the RT market.) If the midload generator has sold power
day ahead for $49.99, then it will lose only $0.01/MWh by purchasing instead of
generating the power it sold. This is far less than what it would lose on startup costs,
so it will not commit and will not produce power. Reneging can be profitable, and
some reneging must be expected.
The possibility of reneging and purchasing replacement power changes the
payoffs to both players in Example 2C. Whenever payoffs change, strategies change,
so the auction of Example 2D will have a different equilibrium. This has been
computed, and the results are shown in Figure 3-9.4. Because it has become less
costly to risk commitment, both suppliers bid under the $50/MWh peaker bids,
and both are selected. This produces 100% overcommitment (double the correct
number of midload plants are chosen to startup) with the result that the plant with
the higher bid always reneges to avoid supplying the top load-slice.
As predicted by those critical of a power exchange,
Uncertain Profits in Example 2D the uncertainty produced by a pure power exchange
Although the dispatch is the same as in Example 2B, causes reneging. The net result, however, is an optimal
prices and payments are different. In Example 2B, dispatch. Reneging merely reverses the DA
the price was $40 off peak and $50 on peak, so the overcommitment, so it is difficult to view this behavior
dispatched midload generator exactly covered its as problematic. The resulting dispatch is efficient and
costs. In Example 2D, random bidding causes the suppliers exercise no market power. This outcome
random fluctuations in short-run profits.
occurs in spite of the absence of a competitive equilib-
For example, if one plant bids $42.80 and the
rium and suppliers being unable to submit bids that
other bids $43, the low bidder will lose $1.40/day per
MW of capacity, while the high bidder will lose reflect their true costs.
$3.50/MWday. This model is too simple to tell the whole story,
If one bids $49 and the other $49.50, the low which is still unknown. It demonstrates that if a power
bidder will make a profit of $17.50/MWday and the exchange uses a bid format that is too simple and to
high bidder will lose only $0.25/MWday. far from generation’s actual cost structure, random
(mixed strategy) bidding may well be the result. If it
is, reneging will probably follow. It also shows that reneging does not necessarily
cause a reliability problem. Reneging does not mean turning off a plant that has
been started, and when it results from the random bids of a power exchange it may

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

302 PART 3 Market Architecture

simply cause generators that were incorrectly accepted to correct the error by not
starting up.
Possibly, a more realistic model would show that reneging can be problematic,
but this must be at least an unusual occurrence because the conditions that cause
potential reliability problems are the same conditions that make reneging very
unlikely. If a generator is known to be desperately needed in real time, it is unlikely
to find it unprofitable to startup.

Conclusions
Neither power exchanges nor power pools have a classic competitive equilibrium
given normal production cost structures. Both have Nash equilibria. If bidding is
honest, load forecasts are perfect and demand is inelastic, a power pool’s Nash
equilibrium should be perfectly efficient in the short run, and a typical power
exchange’s Nash equilibrium may well be 99.99% efficient. This difference will
undoubtedly be swamped by other effects.
If a power exchange has a significant disadvantage, it may be the difficulty of
finding the Nash equilibrium. If a one-part-bid power exchange has trouble finding
the equilibrium; with reneging, or with the inefficiency of the equilibrium, this can
be remedied by using two- or three-part bids. This will not turn it into a power pool
because it does not add side-payments. As correctly noted by advocates of power
pools the computation of multipart bids adds only a very small cost.10 Without side
payments, the exchange will retain most of its transparency and the beneficial
aspects of its prices which more accurately reflect costs (see Chapter 3-8 and the
following section).
If a power-pool has a significant disadvantage, it may be its complexity. This
imposes some small transactions costs but more importantly may open the door
for gaming possibilities. Power-pool auctions are far too complex to check with
game theory except with regard to their main features. They are also too complex
to be checked by human intuition.
Other considerations of interest are market power, system operator control over
ramping, and the effects of reneging on the efficiency of the RT market. None of
these areas are currently understood.

3-9.3 INVESTMENT UNDER A POWER POOL

A power pool is designed to optimize the dispatch, provided that bids are truthful
and load forecasts are accurate. If they are, it finds an optimal dispatch. Three
questions have been raised as to the actual efficiency of such markets:
1. How compatible are bid formats with actual generation costs?11
2. How well do these markets behave in the presence of market power?
3. Do prices plus side payments send efficient long-run signals?
10. The main drawback of two- or three-part bids is that the auction becomes less transparent. Still, it will
not develop the opacity of power pools in which it is impossible for bidders to evaluate their bids relative
to prices because the prices involve side payments which are not public information.
11. See Cameron and Camton (1999, Section III.A).

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-9 The New Unit-Commitment Problem 303

Bid Formats. The previous section argued that a power exchange can perform
remarkably well with a bid format far from representative of physical reality; the
generators bid a minimum energy production but had no such physical restriction.
If such an unrealistic bid format can work so well, it seems unlikely that the more
realistic bid formats of power pools could cause any problems. They might, if
generators bid honestly in a literal sense. But generators will certainly compensate
for small deviations from the realism of power-pool bidding formats just as they
compensate for large deviations from realism in power-exchange bidding.
Market Power. The question of market power remains open. Neither auction
design has any obvious claim to an ability to reduce market power. The question
is particularly vexing because optimal bidding in the presence of market power
often requires a random bidding strategy. Such strategies are notoriously difficult
to analyze. In simulations, the author has found cases in which a power pool
controlled market power better than did a power exchange but in which the power
pool produced a less efficient result. But the opposite situation was also found.
Most likely, the choice between a power pool and a power exchange has little
impact on the exercise of problematic levels of market power. What may make
more difference is whether bids are hourly or daily (see Section 4-4.4). Pools are
more likely to do this, but exchanges can if they use two- or three-part bids.

Long-Run Inefficiency Due to Startup Insurance

Insurance is notorious for dampening economic signals, and side-payments, a form
of startup insurance, proves no exception. Example 3 demonstrates that by insuring
generators against startup costs, the power pool can encourage overinvestment in
generators with high startup costs. Startup insurance prevents the generator from
“feeling” the startup cost on days when the insurance is called into play.

Result 3-9.7 Side Payments in Power Pools Distort Investment in Generation

Power pools pay the difference between marginal-cost prices and the actual costs,
including startup costs, of generators that are required for an optimal dispatch.
This encourages too much investment in the types of generation that receive the
greatest side payments per megawatt of capacity.

Assumptions of Example 3. There are two daily load profiles, with equally
high peaks as shown in Figure 3-9.4, but the second load peak is one third as wide
as the first. While this profile is unrealistic, it appears to illustrate a general principle
and simplifies the example. Peak-load and midload generators have the cost
structures shown in Table 3-9.4. The unrealistically low fixed costs are required
by this simple example to ensure that side payments are made on the low load days.
The first step is to find the minimum-cost equilibrium. Peakers will be built
up to the point where their fixed costs are covered, but the exact level at which
that happens is not important. What matters is that the high prices which allow
peakers to recover their fixed costs allow midload plants to recover an equal amount

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

304 PART 3 Market Architecture

Figure 3-9.4
Alternating daily load
profiles for Example 3.

of their fixed costs. This amounts to $48/MW over the two-day load profile, 3'4
earned on day 1 and 1'4 on day 2. So midload plants will recover $36/MW on day 1,
and $12/MW on day 2 from prices above $40/MWh. This is two-days’ worth of
price-spike revenue.
Table 3-9.4 Cost Assumptions of Example 3
Variable Cost Startup Cost Fixed Cost
Type of Supply (per MWh) (per MWday) (per MWh)*
Peaker $40 $0 $1
Midload $30 $40 $1
* The fixed-cost units, of $/MWh, are explained in Chapter 1-3.

Next the optimal level of investment in midload plants must be determined.

This investment level will be characterized by the duration of peaker use, Dpeaker.
The more midload capacity, the shorter the duration of peaker use. Given the load
profile, Dpeaker determines a particular level of investment. At this duration the
average cost of owning and using capacity, ACK, of a peaker should equal the ACK
of a midload plant. The definition of ACK given in Equation 1-3.1 must be modified
to include startup cost as follows:
Average cost per MW of capacity: ACK = FC + D × VC + SC
Equating ACK of peakers and midload plants and then substituting the example
values with SC converted to $/MWh, and cancelling out the units gives:
FCpeak + Dpeaker × VCpeak + SCpeak = FCmid + Dpeaker × VCmid + SCmid
1 + Dpeaker × 40 + 0 = 1 + Dpeaker × 30 + 40 ' 24
Solve for Dpeaker and then convert this duration from a pure number to the convenient
units of hours per day by multiplying by 1 in the form of (24 hours ' 1 day).
*
Dpeaker = 4 ' 24 × (24 h/day)
*
Dpeaker = 4 h/day.
This is the optimal duration because it minimizes the total cost of production
including variable, fixed, and startup costs. This average duration can be achieved

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-9 The New Unit-Commitment Problem 305

only by a duration of six hours on day 1 and 2 hours on day 2 because of the relative
widths of the load profiles.
If the optimal mix of generation has been built and if the power pool sends the
right long-run signals, then the economic profit of a midload plant should be zero
(short-run profits cover fixed costs), so that there will be no incentive to increase
or decrease its share of the mix. Actual profit can be calculated as follows using
*
the values for Dpeaker just computed.
Table 3-9.5 Optimal Daily Profit of Midload Plants for Example 3
Scarcity Rent Non-Energy Cost
Day *
Dpeaker P > $40 P = $40 SC FC Profit
1 6h $36 6 x $10 $40 $24 $32
2 2h $12 2 x $10 $40 $24 !$32
Dollar figures indicate $/MWh.

Scarcity rent for midload plants comes from two sources: price-spike revenue
(computed above), and prices equal to the peaker’s variable cost, $40/MWh. This
*
price only lasts for the duration Dpeaker and then the price falls to $30/MWh. This
contribution to scarcity rent is Dpeaker × ($40 ! VCmid)/MWh. Notice that economic
*
profit averages zero over the two days, just as it should. Notice also that on day
2, net energy revenue is only $32 so it does not cover startup costs. Consequently,
midload generators receive a side payment of $8/MW on day 2. This makes midload
plants profitable and encourages additional investment.
The additional investment in midload capacity will move the system away from
the optimal mix of generation and increase the total cost of production. In an actual
power market this effect should be small because the side payments will be small,
but it may be large enough to cancel any positive effect of side payments. Without
quantitative analysis little can be said about the efficiency of pool prices relative
to the efficiency of power exchange prices.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 Chapter 3-10
The Market for Operating Reserves
A Telegrapher’s Valentine, by James Clerk Maxwell, 1860.

The tendrils of my soul are twined O tell me, when along the line
With thine, though many a mile apart. From my full heart the message flows,
And thine in close coiled circuits wind What currents are induced in thine?
Around the needle of my heart. One click from thee will end my woes.
Constant as Daniel, strong as Grove. Through many a volt the weber flew,
Ebullient throughout its depths like Smee, And clicked this answer back to me;
My heart puts forth its tide of love, I am thy farad staunch and true,
And all its circuits close in thee. Charged to a volt with love for thee.

A MARKET FOR OPERATING RESERVES PAYS GENERATORS TO BE-

HAVE DIFFERENTLY FROM HOW THE ENERGY MARKET SAYS THEY
SHOULD. If generators are cheap and will produce at full output, the market might
tell them to produce less. If they are too expensive to produce at all, it may tell them
to start “spinning,” and this may require them to produce at a substantial level. Its
purpose is to increase reliability and moderate price spikes.
Chapter Summary 3-10 : Not maximizing profits has an opportunity cost, and
generators must be paid for this to secure their cooperation. There are two philoso-
phies: (1) have the system operator calculate this value from the real-time price
and pay them accordingly; (2) have the generators guess this value and include
it in their bids. The first approach may be quite susceptible to gaming while the
second is optimal in theory but risky for generators in practice and may increase
the randomness of the outcome.
Section 1: Types of Operating Reserve. Operating reserves come in several
qualities classified by how quickly the generator can respond. “Regulation” keeps
the system in balance minute by minute. Ten-minute spinning reserve can start
responding almost instantly and deliver its full response within ten minutes. This
type of reserve will serve as a model for considering market designs. The problem
of linking the different reserve markets is not considered.
Section 2: Scoring by Expected Cost. One approach to conducting a market
for spin is to have suppliers submit two-part bids, a capacity price, CCbid, and an
energy price, VCbid. An obvious way to evaluate such bids is to score them by their

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-10 The Market for Operating Reserves 307

expected cost CCbid + ^h × VCbid, where ^h is the fraction of the time that energy is
expected to be required from a supplier of spin. Unfortunately there is no single
correct h, because the amount of energy used depends on the supplier’s energy
price. Evaluating the bids by using the wrong ^h leads to gaming, which, depending
on the structure of the auction, can be either extreme or moderate. Although a
reasonably efficient auction based on expected-cost scoring seems possible, this
has not yet been demonstrated.
Section 3: Scoring Based on Capacity Price Only. An alternative scoring
approach calls for the same two-part bids (CCbid, VCbid), but evaluates them by
picking those with the lowest capacity price, CCbid.1 Remarkably, this is sufficient,
provided the bidders are exceptionally well informed.
Information requirements are problematic under capacity-bid scoring. To bid
efficiently, bidders must know how the probability of being called on to provide
energy will depend on their energy-cost bid, but this probability function depends
on who wins the auction and what the winners have bid.
Section 4: Opportunity-Cost Pricing. Auctions in which suppliers offer both
a capacity and an energy bid require that they guess what their opportunity cost
will be in the real-time market. For instance, if they believe that the spot market
price will be $50/MWh, and their marginal cost is $49/MWh, they may offer spin
capacity for $1/MWh. If the market price turns out to be $80 and they are not called
on to provide energy, they will have missed a significant opportunity. Of course
with a low spot price they would have won their gamble. The problem is not with
the averages but with the randomness such guesswork will introduce into the
bidding process. As a remedy for this problem, suppliers can be paid their opportu-
nity cost, whatever that turns out to be.

3-10.1 TYPES OF OPERATING RESERVE

Spinning reserve (spin) is the most expensive type of reserve because a generator
must be operating (spinning) to provide it. Spin is typically defined as the increase
in output that a generator can provide in ten minutes.2 Steam units can typically
ramp up (increase output) at a rate of 1% per minute which allows them to provide
spin equal to 10% of their capacity.3 Spin can also be provided by load that can
reliably back down by a certain amount in ten minutes.

1. This approach was developed by Robert Wilson for the California ISO and is explained along with the
problems of expected-cost bidding in Chao and Wilson (1999b).
2. Australia often defines spin as the five-minute increase in output.
3. This value can be improved by the generator owner, and markets may lead to such improvements. Some
reports indicate this value may be more an economic than a physical limit.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

308 PART 3 Market Architecture

The next lower qualities of operating reserves are ten-minute nonspinning

reserves, typically provided by gas (combustion) turbines, and 30-minute
nonspinning reserves. The highest quality of operating reserve is called "regulation"
and is used continuously, not just for contingencies. It is bought separately and
is often not included when operating reserves are discussed. This chapter will
consider only spinning reserve because it is the most critical and illustrates many
important design problems.
Spin can range in price from free to expensive. “Incidental spin” is provided
by marginal generators whose maximum output is not required but more often by
those in the process of ramping down. Sometimes generators are given credit for
spin when they are ramping up at full speed to keep up with the “morning ramp.”
While these may meet the letter of the definition, they do not meet its spirit because
they could not help to meet a contingency such as another generator dropping off
line. Typically the spinning reserve requirement of a system is roughly equal to
the largest loss of power that could occur due to a single line or generator failure,
a “single contingency.”
Providing spin from generators that would not otherwise run is costly for several
reasons. Most importantly, generators usually have a minimum generation limit
below which they cannot operate and remain stable. If this limit requires a generator
to produce at least 60 MW, and its marginal cost is $10/MWh above the market
price, it will lose $600/h if it provides any spin. If it can provide 30 MW of spin
when operating at 60 MW, the spin will cost $20/MWh. In addition, there will be
a no-load cost due to power usage by the generator that is unrelated to its output.
Startup costs should also be included.
Providing spin from inframarginal generators, ones with marginal costs below
the market price, is also expensive. If a cheap generator has been backed down
slightly from full output, its marginal cost may be only $20/MWh while the competi-
tive price is $30/MWh. In this case, backing it down 1 MW will save $20 of
production cost but will require that an extra megawatt be produced at $30/MWh.
The provided MW of spin costs $10/MWh. Sometimes it is necessary to provide
spin in this manner because too little is available from marginal and extra-marginal
generators. This is typically the case when the market price is above the variable
cost of most generating capacity.
The three operating reserve markets are tightly linked to each other and to the
energy market. California demonstrated the folly of pretending differently and
managed to pay $9,999/MWh for a class of reserves lower than 30-minute nonspin
at times when the highest quality reserves were selling for under $50/MWh.4 This
chapter will not consider the problem of how the markets should be linked, although
the most straightforward suggestion is to clear all three simultaneously using a single
set of bids that can be applied to any of the markets.

4. The root of this problem was a “market separation” ideology, although several peculiar rules played a
role as did FERC.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-10 The Market for Operating Reserves 309

3-10.2 SCORING BY EXPECTED COST

The expected cost of spinning reserve depends on the cost (CC) of the reserve
capacity, the cost of the energy provided (VC) when reserves are called on, and
the chance they will be called (h). The cost of providing capacity is always associ-
ated with a megawatt value of the capacity and the length of time for which it is
provided, so CC is measured in dollars per megawatt-hour as is VC. (See Chapter
1-3 for a discussion of units.)
Because the expected cost of spin is CC + h × VC, it seems natural to allow two-
part bidding and to score the bids using this expected-cost formula. If h has a single
value known to the system operator, this is a reasonable scoring procedure. But
if the system operator is mistaken about h and the bidders know h, the procedure
is susceptible to a classic form of gaming.5 Say h is the correct probability but the
system operator believes it is ^h , and the bidder bids CCbid for capacity cost and
VCbid for energy cost. The bid’s score, S, the payment to a winning bid, and the
net profit of a winner are shown in Table 3-10.1.
Table 3-10.1 A Bidder’s Accounting in an Expected-Cost Auction
True expected costs: CC + h × VC
Bids: CCbid, VCbid
Score: S = CCbid + ^h × VCbid
Expected auction payoff: CCbid + h × VCbid
Expected profit: CCbid + h × VCbid ! (CC + h × VC)

The lowest score wins. Say the bidder wants to achieve a score of S. It must
choose CCbid = S ! ^h × VCbid, where it is free to choose any energy bid, VCbid. With
this chosen and CCbid determined, expected profit will be
Expected profit = CCbid + h × VCbid ! (CC + h × VC )
= (S ! ^h × VC ) + h × VC ! (CC + h × VC ),
bid bid

= [S ! (CC + h × VC )] + [(h ! ^h ) × VCbid]

S ! (CC + h × VC ) is unaffected by the bidder’s choice of VCbid, so profit is con-
trolled by the term (h ! ^h ) × VCbid. The bidder can first choose any score, S, and
then achieve any profit level by choosing the correct VCbid! The choice of VCbid
will determine the bidder’s choice for CCbid, as described, but together VCbid and
CCbid will produce any desired level of profit and any desired score. This depends
on the bidder knowing h, and the system operator choosing ^h =/ h. If ^h < h, then
the bidder should bid an extremely high cost for energy and a low cost for capacity.
If ^h > h, the bidder should bid an extremely low (negative) VCbid. (California's 1993
series of Biennial Resource Planning Update auction contained this flaw, and the
winning bids proposed huge up-front capacity payments and negative energy

5. The discussion of this point is based on Chao and Wilson (1999b).

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

310 PART 3 Market Architecture

payments.6 ) As long as the system operator gets ^h wrong and the bidder knows
it, the bidder can achieve any score and any level of profit simultaneously.
While this would seem to make the expected-cost auction entirely useless, a
closer look is required. The probability, h, of selling energy from the accepted
reserve capacity depends on the energy price of the accepted bid. Suppose the
system operator uses an ^h that is intermediate. Those who believe their h will be
lower, will enter very low energy bids. This will make their probability of use, h,
higher than ^h , in contradiction of their assumption, and their strategy will prove
ineffective. This self-limiting effect does not eliminate the problems of gaming
but provides some hope for expected-cost auction design.

3-10.3 SCORING BASED ON THE CAPACITY BID ONLY

The best theoretical work on spinning-reserve auctions (Chao and Wilson 1999b)
suggests the energy-price part of the bid should be ignored when accepting bids,
and winning bids should be those with the lowest capacity-cost values regardless
of their energy-price bids. This surprising result is based on the observation that
suppliers of spin with a low cost of energy will profit from the spin energy market
when more expensive spinning reserve is called. This gives them a strong incentive
to be selected in the spin auction and they will submit a capacity bid, CCbid, lower
than their capacity cost, CC, in order to win the auction. In this way, the relatively
low cost of their energy bid, VCbid, is indirectly accounted for when the auctioneer
evaluates only their capacity bids.
An example will make clear that precise mathematics lies behind this qualitative
argument. Consider two types of suppliers, the second of which has the higher
energy price. Assume that all suppliers bid competitively, and that enough spin
is needed so that some spin must be bought from each type.7 If capacity scoring
works, the cheaper type will win and all of them will be selected before any of the
more expensive types is taken. If capacity scoring fails, it will do the reverse.
In this example, when spin is called on to supply energy, all available spin is
called into service, and the more expensive energy supplier will set the market price
for energy. The costs and bids of these two types are indicated in Table 3-10.2.
Table 3-10.2 Capacity Scoring Example 1, Costs and Bids
Supplier Type Expected Cost Capacity Bid, CCbid Energy Bid, VCbid
1 CC1 + h × VC1 CC1 ! h × (VC2 ! VC1) VC1
2 CC2 + h × VC2 CC2 VC2 ( > VC1 )

A key advantage of capacity scoring is that it makes the energy part of the bid
irrelevant for winning acceptance as spinning reserve. Competitive bidders will
consider only the spin energy market when they set their energy bids and so will

6. This was anticipated by Bushnell and Oren (1994) and described by Gribik (1995).
7. Although cheaper suppliers could bid above the competitive level, this chapter is only concerned with
the efficiency of spinning reserve auctions under competitive assumptions.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-10 The Market for Operating Reserves 311

bid their true variable costs (in a competitive market). This is reflected in Table
3-10.2.
Since competitive bids are always break-even, both bidders subtract from their
capacity bids any profit they expect to make in the spin energy market. To accom-
plish this, they must know what energy bids will set the spin energy market price.
In this case the problem is simple. Some bidders of each type will win, so VC2 will
set the market price. Type 1 bidders will earn h × (VC2 ! VC1) from this relatively
high price so they subtract this from their capacity cost when bidding. Type 2
bidders make no money in the spin energy market and so bid their true capacity
costs.
If this auction is efficient, the type of supplier that can provide power most
cheaply must win the auction. The winner is selected purely on the basis of its
capacity bid, so type 1 will win if and only if
CC1 ! h × (VC2 ! VC1) < CC2
Type 1 can provide spin more cheaply if and only if its cost is less:
CC1 + h × VC1 < CC2 + h × VC2.
These conditions are algebraically equivalent, so the cheaper supplier will always
win the auction (see Chao and Wilson (1999b) for a rigorous treatment). The bidders
have adjusted their capacity bids to take account of their difference in marginal
cost, so the system operator need only evaluate the capacity part of the bid.

Result 3-10.1 Capacity-Bid Scoring for Spinning Reserves Is Optimal

With perfect information, capacity-bid scoring is efficient. Suppliers should bid
a capacity cost and an energy cost. Bidders with the lowest capacity bids should
be selected to supply spin, and from these, the ones with the lowest energy bids
should be selected to run when energy is needed. Accepted bids are paid the
market-clearing capacity price and energy price.

Inframarginal Reserve Capacity

A second example will help demonstrate the robustness of the Chao-Wilson
efficiency result. Spinning reserves can be provided by generators that back down
from their maximum output even though their cost of production is less than the
average price of spot energy, P. By producing at only 90% of capacity, a steam
unit can convert the last 10% of its capacity into spinning reserves. When extra-
marginal sources of spin are too expensive, such inframarginal sources are preferred.
When reserves are in short supply, extra-marginal sources prove insufficient, and
inframarginal sources are required.
Redefine the type 1 supplier in the previous example to be an inframarginal
steam unit. This changes both the cost of using this generator and the generator’s
bidding strategy. First consider its strategy. Instead of comparing its profit in the
spin market to zero profit, the inframarginal generator compares it with its profit
in the spot energy market. If accepted in the spin auction, it would be paid its
capacity bid, CCbid-1, plus (VC2 ! VC1) with probability h. In the spot market it would

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

312 PART 3 Market Architecture

earn its normal scarcity rent at all times. These profits are shown in Table 3-10.3.
Competitive type-1 generators will bid to break even in the two markets thereby
ensuring they will at least cover their opportunity cost. To find their bid, CCbid-1,
equate the profits in the two markets and solve for CCbid-1.
Table 3-10.3 Equating Profits Determines the Capacity Bid Price
Type Spin-Market Profit Spot-Market Profit
1 CCbid-1 + h × (VC2 ! VC1) P ! VC1

The expected cost of a type-1 generator providing spin also must be found in
order to determine the efficiency of this auction. This time there is no physical
capacity cost, CC1, but there is a cost to replacing the energy that would have been
provided by this inframarginal generator had it not withdrawn some of its output
to provide spin. The cost of providing a small amount of additional energy is the
spot price, P, so the net replacement cost is (P ! VC1). This is the actual increase
in production cost in a competitive market, so it is the real cost, not the opportunity
cost, of providing spinning reserve with an inframarginal generator. In addition,
there is the possibility, h, that spin energy will be needed; this increases the produc-
tion cost of type-1 spin by h × VC1. This total as well as the value of CCbid-1 found
by equating the two profit levels in Table 3-10.3 are shown in Table 3-10.4 for the
type-1 generators. The entries for the type-2 generators remain unchanged from
the previous example.
Table 3-10.4 Expected Costs and Capacity Bids
Type Expected Cost Capacity Bid, CCbid-1
1 (P ! VC1) + h × VC1 (P ! VC1 ) ! h × (VC2 ! VC1 )
2 CC2 + h × VC2 CC2

This auction is efficient if the type of supplier capable of providing spin most
cheaply wins the auction. The winner is selected on the basis of its capacity bid,
CCbid, so type 1 will win if and only if:
(P ! VC1 ) ! h × (VC2 ! VC1 ) < CC2
Type 1 can provide spin more cheaply if and only if
(p ! VC1) + h × VC1 < CC2 + h × VC2

These are algebraically equivalent, so the auction is efficient.

The Information Burden of Capacity-Bid Scoring

In the standard theory of competitive markets, it is optimal for suppliers to bid their
marginal cost curves. This information is readily available. Knowing one’s own
costs is merely a matter of good accounting and requires no theory of how the
market works and no data on other suppliers and market demand conditions.
Optimal capacity-only bids cannot be based simply on information about one’s
own costs. They require knowledge of h(VCbid), the probability of being called
to provide energy as a function of the energy bid, VCbid. The bidder must use this

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

 CHAPTER 3-10 The Market for Operating Reserves 313

function to compute how much energy it will be called on to provide and the
distribution of prices when it is called. This crucial function is determined by the
outcome of the auction. If many low bids for VCbid are accepted, h(VCbid) will be
low, the profit from supplying energy will be low, and the bidder will need to submit
a higher capacity bid.
In a very stable market, learning h(VCbid) would be relatively easy. Bidders
would observe yesterday’s h(VCbid), base their bids on it, and that would determine
a new h(VCbid) the next day. This would be observed, and, as the process repeated,
h(VCbid ) would most likely converge to some stable function which would be known
to all. But the conditions in the spinning reserve market vary with weather, time
of year, and the state of repair of generators. The market is also unpredictable from
year to year due to load growth, rainfall, and new investment in generation. Conse-
quently, the function h(VCbid) is likely to be difficult to predict. If not correctly
predicted, suppliers will bid incorrectly and the dispatch will be inefficient. Because
the bidding problem is so complex, larger suppliers will probably have an advantage
unless a market develops to supply estimates of h(VCbid). In this case, part of the
inefficiency will become visible in the form of revenues to the suppliers of informa-
tion, but presumably, total inefficiency would decrease.
Many proposals for “simple” auctions prove their efficiency by assuming that
bidders have perfect information and unlimited computational abilities. Sometimes
this assumption is a reasonable approximation of reality. As a first step toward
checking this approximation, those who propose an auction mechanism should
be required to provide an explicit method by which bidders could bid optimally,
or at least very efficiently, in a competitive market. If the auction designer cannot
demonstrate how a bidder could compute its bid under the simplifying assumption
of perfect competition, then market participants may find the information and
computational requirements of the design overly burdensome. This may lead to
inaccurate and conservative bidding.

3-10.4 OPPORTUNITY-COST PRICING

One complaint against paying the clearing bid price for spinning reserve capacity
is that those selected to provide reserves often find themselves losing money and
tempted to cheat. For instance, if the expected spot price is $50, a generator with
a marginal cost of $49 might offer to provide reserves for $1/MWh and be accepted.
If the real-time price then turns out to be $80/MWh, and no reserves are called on,
the generator will perceive that it is losing $30/MWh relative to what it could be
making if it were not providing reserves.
Although this situation will lead generators to wish they could cheat, it is not
a difficult matter to check on them. The capacity of a generator, which is easily
verified, together with its ramp rate change little from day to day. As long as a 200-
MW generator is paid for no more than 180 MW of output while it is supposedly
providing 20 MW of spin capacity, it cannot cheat.
In spite of this, there may be good reasons for helping generators to avoid such
uncertainty. If a generator can base its bids on information about its own costs

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.

314 PART 3 Market Architecture

instead of on estimates of future market outcomes (in a very volatile market), it

seems reasonable to assume that bidding will be more accurate, at least in a competi-
tive market. This is the philosophy behind the approach to spinning reserve markets
used in NYISO.

Steps Toward an Opportunity-Cost Pricing Design

In a spin market based on opportunity costs, generators who are required to spin
are paid the short-run profit they would make if they were producing energy. This
is calculated based on their marginal-cost bid and the market price. For example,
suppose inframarginal generators have variable costs ranging from $30 to
$40/MWh, and the expected spot price is $40/MWh. In a competitive market, these
would bid their true variable costs and a CCbid of zero. Suppose several extra-
marginal generators bid their true capacity cost of $8/MWh. The system operators
uses the expected spot price to impute opportunity costs in the range of zero to
$10/MWh to the inframarginal generators. The imputed opportunity costs are then
added to their capacity bids.
If some, but not all, of the extra-marginal generators are needed, they set the
clearing price for capacity at $8/MWh. All inframarginal generators with variable
costs above $32/MWh are accepted, and all are paid $8/MWh provided the spot
price is $40/MWh. If the market price proves to be $50/MWh, then they are all
paid a capacity price of $18/MWh. Because they are paid a market clearing price
and treated as if they had been the most expensive accepted bid, they have no reason
to regret their bids and they will bid honestly.8
This market, based on opportunity costs, uses a capacity-bid design. This is
unusual. Opportunity-cost markets tend to be unit-commitment markets with
multi-part bids and optimization programs. In this setting it is not standard to
“score” the bids by a simple rule, but in effect an optimization program is just a
complex scoring rule. As a result, there are winners and losers. To decide if a bid
with a high marginal cost is nonetheless a winner in the spin auction, the program
assigns some probability to the chance of the supplier being called on to provide
energy—it assigns an estimate of h. Such markets are closely related to the two-part-
bid-evaluation, expected-cost designs described in Section 3-10.2.
Unfortunately there is little if any literature on the gaming possibilities in this
kind of market. Section 3-10.2 suggested they may be serious, but paying opportu-
nity costs gives generators with market power in the ancillary service market a
reason to lower their marginal-cost bids. Since their marginal-cost bids are used
as well in the energy auctions, this reason to bid low should partly cancel a tendency
to exercise market power in the energy market.

8. This point was suggested by David Mead in a personal communication, September 18, 2001.

© February2002. Steven Stoft, Power System Economics (IEEE/Wiley) ISBN 0-471-15040-1.