Directory UMM :Data Elmu:jurnal:J-a:Journal Of Economic Dynamics And Control:Vol25.Issue3-4.March2001:

Journal of Economic Dynamics & Control
25 (2001) 561}592

Experimental analysis of the e$ciency
of uniform-price versus discriminatory
auctions in the England and Wales
electricity marketq
John Bower, Derek Bunn*
London Business School, Regent's Park, Sussex Place, London NW1 4SA, UK

Abstract
The question of whether the uniform price or discriminatory auction format is the
better multi-unit auction mechanism is addressed in the context of the 1999 debate on
reforming the England & Wales electricity market. Each generator is modelled as an
autonomous adaptive agent capable of endogenously developing its own bidding strategies using a naive reinforcement learning algorithm. The discriminatory auction results
in higher market prices than the uniform-price auction. This is because market prices
are not publicly available and agents with a large market share gain a signi"cant
informational advantage in a discriminatory auction, thereby facing less competitive
pressure. ( 2001 Elsevier Science B.V. All rights reserved.
JEL classixcation: C63; C7; D43; D44; L94
Keywords: Agent-based computational economics; Auction design; Electricity markets;

Market power

q
This research was partly funded by ESRC Postgraduate Training Award No. R00429724400 and
a London Business School Financial Award.

* Corresponding author. Tel.: #44(0)20-7262-5050; fax: #44(0)20-7724-7875.
E-mail address: dbunn@london.edu (D. Bunn).

0165-1889/01/$ - see front matter ( 2001 Elsevier Science B.V. All rights reserved.
PII: S 0 1 6 5 - 1 8 8 9 ( 0 0 ) 0 0 0 3 6 - 1

562

J. Bower, D. Bunn / Journal of Economic Dynamics & Control 25 (2001) 561}592

1. Introduction
Auction mechanisms have been attracting increasing attention in recent years,
motivated in part by their increasing use in selling network access rights, and
creating competitive markets, in telecommunications, gas and electric utilities,

as well as in periodic auctions of various government securities. Research into
alternative designs for auctions is mainly concerned with understanding how the
di!erent rules which govern the auction process, the number of items on o!er,
and the timing and quantity of information available to bidders, a!ect the
market price. However, most of this is theoretical research and assumes that
bidders will be making competitive bids, that these bidders are symmetrical in
size, that they bid once only, for a unique good, and that they are risk neutral. In
reality, however, many auctions involve an oligopoly of asymmetric bidders who
meet repeatedly, and frequently, to bid for the same commodity. Under these
circumstances, bidding is unlikely to be competitive, but strategic in nature, with
bidders seeking out opportunities to exercise market power. As a result,
oligopolistic bidders will have a powerful incentive to exploit, or &game', any
auction mechanism and, as a result, cause price ine$ciency. Despite the frequent
occurrence of multi-period, multi-unit, auctions, it is therefore still an open
question as to which mechanism is most e$cient, especially when the auction
involves an oligopoly of bidders.
In this paper, we attempt address this question by using a novel agent-based
computational economic (ACE) approach to simulate the behaviour of an
oligopoly of bidders in a range of multi-unit, multi-period, auction settings. The
application of ACE to the analysis of imperfect competition is a particularly

promising area of research, because of the obvious potential for emergent
learning and collusive behaviour as "rms interact through time. In this
case, we are interested in understanding what behavioural e!ects alternative
bidding rules may have on bidders of di!erent sizes, how that impacts
their ability to exercise market power, and what happens to market price
as a result. In terms of a practical application, we have been particularly
motivated by the 1999 proposal to transform the 10 yr old England &
Wales wholesale electricity market, which had been run as a uniformprice auction, into a commodity market operating as a discriminatory
auction. We have therefore developed a detailed model of electricity trading in
England and Wales, and used this to experiment with di!erent auction market
mechanisms.
In Section 2, we brie#y review the theoretical and empirical background,
highlighting the di$culty of applying single-unit auction theory to multi-unit,
multi-period, auctions, and survey the con#icting evidence from empirical
studies. Section 3 introduces the Pool and the current debate on changing the
auction mechanism. Section 4 describes the agent-based simulation model we
have developed to analyse the likely impact of such a change market, with the

J. Bower, D. Bunn / Journal of Economic Dynamics & Control 25 (2001) 561}592

563

results in summarised in Section 5. Finally, the wider theoretical and economic
implications of our "ndings are discussed in Section 6.

2. Background
Game theoretic analysis of various auction types (e.g. &First Price', &Second
Price', &English', &Dutch') generally assume a benchmark model where bidders:
f are risk neutral;
f have their own private valuations of the good (the &independent-privatevalues' assumption);
f are symmetric; and
f make or receive payments as a function of bids alone.
As a result, the general results are that the "nal price achieved is invariant to
the auction mechanism (see Eatwell et al., 1998; Rothkopf and Harstad, 1994;
McAfee and McMillan, 1987). However, despite being a cornerstone of modern
auction theory, this Revenue Equivalence Theorem only applies where a single
unit of an indivisible good is being auctioned, in a single-period setting. In
contrast, theory is much less well-developed where these conditions are relaxed,
and particularly where multi-unit, multi-period bidding occurs between an
oligopoly of asymmetric bidders.

The multi-unit auction analogue of the "rst-price auction is the discriminatory
auction where bidders make sealed bids indicating the quantity of goods they
are willing to buy at a range of prices. The auctioneer allocates the goods to the
highest bid "rst, according to the quantity demanded, and so on down the
sequence of received bids until all the goods have been allocated. In some cases,
the seller may reserve the right to increase the total number of goods allocated,
for example to take advantage of unexpectedly high demand, or to scale back
the sale if bids fall below some reserve price. Bidders pay for the goods at the
price equal to their individual bids (&Pay Bid'). Likewise, the uniform-price
auction is the multi-unit analogue of the second-price auction which is run in the
same way as a discriminatory auction except that successful bidders all pay the
same price regardless of the bids they actually made. This price is equal to the
highest (marginal) bid price accepted (&Pay Marginal').
Some researchers have argued that the uniform-price auction has a lower
winner's curse and results in greater revenue to the seller than would a discriminatory auction (see Milgrom, 1989; Bikhchandani and Huang, 1993). However,
this work was based on single-unit auction theory and Back and Zender (1993)
were able to show the result was critically dependent on the assumption that
the good was indivisible. They concluded that uniform-price auctions are no
longer universally superior to discriminatory auctions. Wang and Zender (1994)
go on to show that the two auction types cannot be ranked because, at least

564

J. Bower, D. Bunn / Journal of Economic Dynamics & Control 25 (2001) 561}592

theoretically, there are always discriminatory auction equilibria that dominate
uniform-price equilibria, and vice versa. However, as Das and Sundaram (1997)
point out even this conclusion is not clear cut. Not all equilibria are equally
plausible; auction markets do not work in isolation and often interact with
forward markets, and other secondary markets; and the impact of non-competitive bidders creates noise that a!ects the actions of competitive bidders.
Nevertheless, largely as a result of the early theoretical research described
above, a number of governments began experimenting with uniform-price
auctions. However, analysis of their results presents a mixed picture. For
example, when the US Treasury switched to uniform-price auctions, for 2 and
5 yr treasury note sales in 1992, the results were inconclusive. A study by Simon
(1994) argues that the uniform-price auction cost the US Treasury money while
Nyborg and Sundaresan (1996), compared actual bids and market price data,
before and after the switch, and found no signi"cant di!erence in prices under
the two di!erent sets of trading arrangements. Their explanation for the apparent revenue equivalence related to the fact that both a &when issued' forward
market and a secondary market, operated in parallel with the regular auctions.

This signi"cantly contributed to the dissemination of information between
potential bidders and helped relieve information asymmetries that would otherwise create market ine$ciencies.
Similar analysis in other markets include a Nyborg et al. (1997) study of bid
price data during 1990}1994 for the Swedish Treasury market. This suggested
that the government might be better switching to a uniform-price auction from
the current discriminatory auction, although the results were not conclusive.
Umlauf (1993) and Tenorio (1993) studied Mexican treasury and Zambian
Foreign exchange auctions, respectively. Both studies report higher seller revenues under a uniform-price auction format. A subject that has recently become
topical is the sale of gold reserves by European central banks in preparation for
the Euro, and by the IMF to fund debt relief to third world nations. When, in
1959, the IMF sold some of its reserves in a series of 35 discriminatory auctions
and 10 uniform-price auctions signi"cantly higher revenue was achieved from
the discriminatory auctions. Interestingly, when the UK Treasury began selling
its gold reserves in 1999 it chose to use a uniform-price format (Bank of England,
1999). Though it is too early to tell what e!ect the choice of auction mechanism
has had on revenues, the Treasury has reserved the right to change this format in
later auctions. This clearly suggests there is still some uncertainty on its e!ectiveness. Indeed, the UK authorities decided against introducing uniform-price
auctions for treasury securities, a few years earlier, because of the inconclusive
results from the US treasury note auctions described above (Bank of England,
1995).

None of the studies cited above, explicitly address the question of what impact
the numbers, and size, of bidders have on the market price. This is surprising
given that the switch to uniform-price auctions by the US Treasury was

J. Bower, D. Bunn / Journal of Economic Dynamics & Control 25 (2001) 561}592

565

prompted, at least in part, by concerns raised by the 1991 Salomon bond trading
scandal. Here, one large broker-dealer, Salomon Brothers, made unauthorised
use of client accounts to secretly acquire substantial quantities of bonds, in 2 yr
treasury note auctions, which exceeded explicit US Treasury limits that prohibit
any one bidder acquiring more than 35% of an issue (see Jegadeesh, 1993 for
further details). Nyborg et al. (1997) also note there are only 16 dealers in
Swedish treasuries and that one dealer will frequently succeed in bidding for the
entire auction. The possibility of implicit, or explicit, collusive behaviour developing among a relatively small number of bidders who compete regularly at
these auctions, or even one bidder dominating the entire auction, is an important factor not so far considered. Indeed, one explanation for the inconclusive
results of the empirical studies on auctions in "nancial markets may be in part
due to the concentration of large dealers that dominate these markets. Smith
(1990) addresses the possibility of non-competitive behaviour in auctions, from

a sociological rather than economic point of view, and suggests that bidders
should be thought of as social groups with norms, values, and behavioural rules.
He provides empirical evidence from New York jewellery auctions, where
regularly repeated auctions are attended by an essentially closed group of
bidders who have come to clear understandings about how and when they will
compete for particular lots on o!er. These resulted in bids, and prices, which
"tted the group's notion of &worth', in contrast to the typically assumed reservation price. The importance of behavioural learning, through repetition, in
auctions is further illustrated in experiments carried out by Cox et al. (1984).
These show that subjects often do not follow the dominant bidding strategy, to
begin with, but gradually learn to adopt the strategy over time as their experience increases through repetition. Laboratory experiments by Isaac and Walker
(1985), where subjects were allowed to discuss and coordinate bidding strategies,
over many di!erent auction types including multi-unit auctions, also resulted in
a signi"cant level of collusion in many cases.
Overall, the literature therefore presents a confusing picture. Studies of
treasury and other "nancial market auctions are rather inconclusive and this is
often taken to imply that there may be no signi"cant di!erence between
uniform-price and discriminatory auctions. However, the existence of forward
&when issued' markets before treasury auctions and secondary markets afterwards suggests that the results may be confounded by the interrelationships
between prices in linked markets. This result may also explain why out of 42
countries, surveyed in 1994, all but two were still using discriminatory auctions

to market government debt (Bartolini and Cottarelli, 1997). What is also
unclear, is the potential e!ect that imperfect competition between bidders might
be having on these results. A number of experimental and empirical studies
suggest that where bidders participate in repeated rounds of auctions they can
quickly learn to cooperate to prevent prices reaching competitive levels. As all of
these examples are of regularly repeated auctions, often with the same relatively

566

J. Bower, D. Bunn / Journal of Economic Dynamics & Control 25 (2001) 561}592

small group of participants bidding, it is possible that implicit collusion may
be masking underlying di!erences between auction mechanisms that would
otherwise become apparent in a more competitive environment. In fact,
the only really clear cut conclusion that we can draw from the literature is
that single-unit auction theory cannot successfully be applied to multi-unit
auctions.
A new set of auction markets, that may help resolve these issues, has become
available in the last few years due to the deregulation of the global electricity
industry. Electricity market auctions are of particular interest, because they

involve multi-unit bidding, often by relatively concentrated groups of bidders,
some of whom have very signi"cant market shares. Not only that, the rules
which govern these markets usually require the same bidders to compete
regularly, at least daily, in order to o!er their output through a centralised
auction process which o!ers signi"cant potential for learning and implicit
collusion.
Rothkopf (1999) highlights the central importance of repetition in electricity
auctions, a fact that has often been ignored in traditional auction theory, and
stresses the fact that tacit collusion is a much greater problem when bidders
meet repeatedly, whichever auction method is chosen. To combat these implicit
and explicit collusion problems, he recommends delaying the release of information about bids and auction outcomes, particularly from sealed-bid auctions,
and stresses the need for e!ectively enforced prohibitions on active conspiracy
between bidders. However, delaying information is at odds with the common
belief that extra transparency and information availability will improve market
e$ciency. The conclusion he draws about electricity trading arrangements in
California is that a sealed-bid auction is better than a progressive auction but to
avoid collusion, care must be taken in deciding which information to release.
For the remainder of this paper we will focus on the wholesale electricity market
in England and Wales and the proposal made by the O$ce of Electricity
Regulation (OFFER)1 to transform the day-ahead market, which has been
a uniform-price auction (the &Pool'), into a discriminatory auction (&the bilateral
model').

3. Electricity trading in England and Wales
When the UK government introduced competition into the England & Wales
electricity market, in 1990, it created a pool-based market through which all

1 OFFER amalgamated with the O$ce of Gas Regulation (OFGAS) in 1999, creating the O$ce of
Gas and Electricity Market Regulation (OFGEM) but for clarity we shall continue to refer to
OFFER throughout.

J. Bower, D. Bunn / Journal of Economic Dynamics & Control 25 (2001) 561}592

567

Fig. 1. Monthly time-weighted average Pool prices.

physical supplies of bulk electricity are traded, and which many other countries
have since copied. However, in 1998, OFFER launched an industry-wide review
of the trading arrangements because of the widely held belief that the market
mechanism might be creating, or at least contributing to, generator market
power in the Pool and as a result allowing them to keep market prices well
above their marginal production costs.
OFFER had noted that, on a time-weighted basis, the average winter Pool
price in 1997/98 was 12% higher, in real terms, than in the previous year, and
35% higher than in winter 1990/91 (see Fig. 1) when the Pool began trading
(OFFER, 1998). The regulator even resorted to making an explicit statement
that &Pool prices must come down' (OFGEM, 1999a) after estimating that
gaming in the Pool had cost consumers over C90m more, in December 1998 than
in previous years (OFFER, 1999). The fact that OFFER had to intervene so
frequently throughout the Pool's history, most recently in July 1999 (OFGEM,
1999a), put reform of the Pool trading arrangements high on the UK government's agenda, and pool reform become viewed as an essential prerequisite for
reducing market prices. Indeed, OFFER estimated that introducing the bilateral
model would cause prices to fall by at least 13%, a saving C1.5 billion per year
(Taylor, 1999). However, in the past, OFFER had been less convinced. When the
whole issue of Pool trading arrangements was previously reviewed (OFFER,
1994) the regulator found there was insu$cient evidence to instigate further
reform and concluded that the uniform-price auction format o!ered some
economic advantages over a mechanism that paid generators in a discriminatory fashion with their own bid prices.

568

J. Bower, D. Bunn / Journal of Economic Dynamics & Control 25 (2001) 561}592

Empirical analysis of the strategic bidding behaviour of generators in the
Pool, Wolfram (1998), shows that strategic bid increases are indeed occurring
and that this is particularly prevalent for bidders with large portfolios of plant.
They tend to bid higher than other bidders, all else equal, because they receive
a larger payo! on all their lower price units, have the necessary spare capacity,
and biggest incentive to bid in such a way as to raise the marginal price setting
bid. This is also consistent with the "ndings of Bunn et al. (1998) that, even in the
absence of collusion, generators could still exercise market power in the current
Pool. The importance of generators' ability to in#uence the marginal bid price
was also investigated in the early days of the Pool by von der Fehr (1991) who
believed that a small changes to the auction mechanism could make a signi"cant
di!erence to the "nal market price. However, despite the conviction with which
new trading arrangements were proposed by OFFER and the UK government
in 1999, there has been no substantial theoretical, or empirical, evidence that
changing the Pool's trading mechanism from a uniform-price to a discriminatory auction would reduce market prices.
In the remainder of the section we give a brief overview of the Pool and
bilateral models. Extensive documentation covering current Pool trading arrangements, and OFFER's bilateral model proposal can be found at the web
sites of the Electricity Pool of England & Wales, OFFER, and the RETA
Group2 (see particularly the comprehensive summary in OFGEM, 1999b, c).
Here, we will limit ourselves to identifying those elements of the Pool that the
bilateral model would replace.
3.1. The Pool
The focal point of the Pool is the day-ahead market that sets a price3 for
electricity in each of the 48 half-hourly periods of the next day. This is closely
tied to a forward market which generators and consumers use to hedge their
exposure to volatile prices in the day-ahead market.
The day-ahead market begins with each generator submitting price, and
quantity, bids to the National Grid Company (NGC, 1998), for up to three
incremental levels of output from each of their generating units (&gensets'). NGC
is both the Independent System Operator (ISO), as owner of the transmission
system, and also the Independent Market Operator (IMO) which runs the
market's clearing, settlement and payment systems. The day-ahead market is
mandatory and all generators, who wish to have their plant despatched, must
2 All of this documentation, including more detailed descriptions of the Pool mechanism can be
downloaded from http://elecpool.com, http://www.open.gov.uk/o!er, and http://www.reta.org.uk,
respectively.
3 Bids and market prices are quoted in C/Megawatt hour (C/MWh) and quantities in Megawatts
(MW).

J. Bower, D. Bunn / Journal of Economic Dynamics & Control 25 (2001) 561}592

569

submit their bids by 10.00 a.m. each day. Once submitted, these bids cannot be
changed and apply to the whole 24 h period, starting at 5.00 a.m. on the next
day. However, generators are allowed to withdraw their bids at any time by
re-declaring their gensets unavailable right up to the moment of despatch.
On receipt of the price and quantity bids from the generators, NGC constructs a supply curve by stacking the bids in price merit order, and identi"es the
optimal, lowest cost, combination of gensets that will meet its forecast of
demand in each half-hour of the next day. This process creates the Unconstrained Schedule that NGC uses to begin planning the operational despatch of
plants and to calculate the half-hourly system marginal price (SMP). This is
e!ectively the bid price of the most expensive generating set that NGC has
forecast will run in each half-hour of the next day. At around 4.00 p.m., on the
day-ahead, NGC publishes4 SMP and also the Pool Purchase Price (PPP),
which is SMP plus a Capacity Element (CE). This covers capacity payments
made to all gensets included in the Unconstrained Schedule, and is meant to
encourage investment in new plant designed to meet peak loads. CE rises
exponentially when forecast demand begins to approach total system capacity
and can make up a signi"cant proportion of a generator's "nal revenue.
Consumers also pay a uniform price, the Pool Selling Price (PSP) which, in
low-demand periods (the so-called &Table B') is equal to PPP. In high-demand
periods (the so-called &Table A') PSP includes an additional element, called
Uplift, which is made up of Energy Uplift and Unscheduled Availability Payments. Energy Uplift covers the cost of calling additional plants to run, not
originally included in the Unconstrained Schedule, to meet deviations from
forecast demand, and replace plant that has become unavailable through outages or generators' availability redeclarations. Unscheduled Availability Payments are made to gensets which were not included in the Unconstrained
Schedule, but which were bid into the day-ahead market, to encourage investment in peaking plant in the same way as capacity payments operate. Consumers have no direct involvement in the price setting mechanism except for
a few very large power users, who are permitted to make demand reduction bids,
but the quantities involved typically amount to less than 2% of winter peak
demand.
Despite the central importance of the day-ahead market, an equally vital part
of the Pool is the forward market where consumers and generators meet
voluntarily to trade bilateral contracts. In practice, this is where the price of
around 90% of electricity, consumed in England and Wales, is actually "xed.
This operates like any other commodity forward market except that no physical
deliveries of power occur as a result of the contracting process. Instead, the

4 NGC noti"es all relevant Pool members electronically at 4.00 p.m. and prices are also made
publicly available in the Financial Times, published on the following day.

570

J. Bower, D. Bunn / Journal of Economic Dynamics & Control 25 (2001) 561}592

mandatory nature of the day-ahead market means that both of the main
contracts traded, Contracts for Di!erence (CFDs), and Electricity Forward
Agreements (EFAs), must be "nancially settled against prices set in the dayahead market, usually PPP.5 The CFD market covers long-term forward
contracts, with durations stretching from six months up to 15 yr and beyond,
while the EFA market is for much shorter-term contracts, usually one week to
six months in duration. In contrast to the day-ahead market, EFA and CFD
market prices are not publicly quoted and, despite the large volumes of power
being contracted for, both markets are also highly illiquid.
3.2. The bilateral model
The proposed bilateral model has the following key components:
(i) voluntary forward market as required by consumers and generators;
(ii) voluntary screen-based, half-hourly balancing market operating from 4 h
before despatch right through the particular half-hour in question; and
(iii) mandatory settlement process for imbalances.
Under this arrangement generators and consumers who have contracted for
physical deliveries of power in any given half-hour, through the forward or
short-term market, would be responsible for self-despatching those contracts.
Their only other responsibility would be to provide their &Final Physical
Noti"cation' (FPN) to the ISO before the &gate closure' point has been reached,
some four hours prior to each half-hourly despatch period. The ISO will assume
full responsibility for any further contracting necessary to maintain system
security until despatch is completed. This will be achieved partly by buying in
ancillary services, probably under long term contracts similar to those which
NGC now has for frequency response and black-start capabilities, and partly by
buying and selling in the balancing market to cover any imbalances between
noti"ed positions and actual demand and supply.
At least initially, the operation of the forward market under the bilateral
model is likely to be almost identical to the way that current EFA and CFD
markets operate, except that generator and consumers will be able to contract
for physical forward delivery in addition to "nancial settlement. New trading
instruments may eventually be developed, and there is still some uncertainty
about which market will set the benchmark price for "nancially settled contracts, but it is envisaged that the forward market will continue to be the
dominant mechanism for setting price for the majority of electricity consumed.
The balancing market will also be optional and is similar in concept to that
operating in the Scandinavian electricity markets, NordPool and EL-EX, and

5 See Hoare (1995) for a detailed description of CFD and EFA forward markets.

J. Bower, D. Bunn / Journal of Economic Dynamics & Control 25 (2001) 561}592

571

also the UK gas market. Here generators and consumers will be able to make
location-speci"c bids and o!ers to ISO, presumably also on a screen-based
system, for increments6 and decrements7 to their FPN positions. The balancing
market will therefore create an explicit market mechanism for allocating the cost
of maintaining an energy balance on the transmission system, and replace the
complex and opaque process that NGC currently uses to calculate and allocate
Energy Uplift. By applying a market price to imbalances, it is also hoped that
this will make supply and demand much more responsive to short-term changes
in the operational status of the system.

4. Modelling the Pool and bilateral mechanisms
We have developed an agent-based simulation (ABS) model of the wholesale
market for electricity market in England and Wales which allows us to compare
market prices, and the bidding strategies of individual generators under the
di!erent trading arrangements. The key feature of this simulation approach is
that it uses a micro-level, bottom-up, representation of the market with each
generating "rm represented, at the level of its individual power plants, by
a separate computer-generated autonomous adaptive agent (&agent'). The agents
are capable of developing their own bidding strategies, to explore and exploit
the capacity and technical constraints of plant, market demand, and di!erent
market clearing and settlement arrangements.
4.1. Modelling objective
The crucial focus of comparison is the Pool's day-ahead market versus the
bilateral short-term market. We model both as a daily repeated auction and
compare the market clearing prices set under the four di!erent combinations of
trading, and settlement, arrangements set out in Table 1.
Clearly these are stylised models, ignoring much of the complexity of the real
Pool day-ahead and bilateral short-term markets but they allow us to isolate the
following key issues:
(i)
(ii)
(iii)
(iv)

Pay SMP versus Pay Bid settlement;
daily bids versus hourly bids;
learning through repetition; and
impact of information availability.

6 The price that generators wish to be paid for an increase in output or consumers are willing to
pay for an increase in demand.
7 The price that generators are willing to pay for a decrease in output or consumers wish to be paid
for a decrease in demand.

572

J. Bower, D. Bunn / Journal of Economic Dynamics & Control 25 (2001) 561}592

Table 1
Alternative auction models tested
Market

Bidding

Settlement

1.
2.
3.
4.

Daily
Daily
Hourly
Hourly

Uniform price
Discriminatory
Uniform price
Discriminatory

Pool Pay SMP
Pool Pay Bid
Bilateral Pay SMP
Bilateral Pay Bid

4.2. Modelling architecture
The ABS model we have developed has a trading environment, a set of agents,
and an economic environment. The trading environment is a daily repeated
auction upon which di!erent combinations of bidding, clearing and settlement
arrangements, described above, are exogenously imposed along with any regulatory controls on allowed agent bidding behaviour. Each agent represents one of
the generating "rms operating in the Pool during 1998, and which is endowed
with a portfolio of plants characterised by capacity, fuel type, e$ciency, availability, etc. The economic environment de"nes the demand pattern for electricity, and the input costs, both of which are imposed exogenously as static
variables.
4.2.1. The trading environment
At the start of each simulated trading day, each agent is allowed to submit one
bid for each plant in its portfolio and we assume that all the expected available
capacity makes a &"rm' bid for the whole day. Each bid is therefore linked to
a speci"c plant capacity, which means that agents are, in e!ect, submitting "rm
bid supply functions. This is consistent with the work of Klemperer and Meyer
(1989), later used by Green and Newbery (1992) and Green (1996a) in their
analysis of the bidding behaviour of generators in the Pool.
Bidding is allowed in one of two ways, either a single price for each plant for
a whole day (daily bids) as in the current Pool's day-ahead market, or 24
separate hourly prices for each plant (hourly bids) as in the bilateral model. The
market is cleared by stacking the plant bids, low to high, and allocating demand
to plants, in strict merit order, until demand is exhausted for each hourly period.
There are 24 separate hourly settlement prices in each trading day, and any
plant that has bid above the bid price of the marginal plant in any given hour
has a zero utilisation rate. The auction results are simultaneously calculated, for
all 24 h, at the end of the trading day. Revenues are calculated on the basis of
demand allocated multiplied by the price bid by the marginal plant in each hour
(Pay SMP settlement), or on the basis of a plants own bid (Pay Bid settlement).
All agents simultaneously receive the results of their bids at the end of the

J. Bower, D. Bunn / Journal of Economic Dynamics & Control 25 (2001) 561}592

573

trading day, and even where separate hourly bids are submitted there is no
opportunity to observe the outcome of these until trading is completed.
4.2.2. The agents
The agents, summarised in Appendix A, represent the di!erent generating
"rms, and centrally despatched plant capacities, marginal production costs, and
expected plant availabilities, during 1998. These were synthesised from a range
of public, and private, sources as well as our own estimates.
The ABS approach allows us to avoid making the usual restrictive assumptions that are required by traditional economic analysis of imperfect competition. Instead, the agents use simple internal decision rules, summarised in
Table 2, that allow them to &discover' and &learn' strategic solutions which satisfy
their pro"t and market share objectives over time. Taken together, these rules
constitute what is essentially a naive reinforcement learning algorithm8 that
seeks out and exploits successful bidding strategies while discarding unsuccessful bidding strategies. As a result, the behaviour of the simulated market is
almost entirely emergent as it is created endogenously by the aggregate interaction between agents and their environment.
The agents' bidding strategies are therefore not speci"ed exogenously by the
modeller but are developed by the agents themselves. The model also has the
advantage of allowing bidding strategies to be observed for asymmetric bidders,
right down to the individual plant level. ABS is therefore a distinctly bottom-up
approach focussing on individual strategic decision-making behaviour, rather
than top-down aggregate market behaviour.
Strategic learning is driven by each agent attempting to jointly satisfy the two
objectives of:
(i) continuously increasing its own overall pro"tability, from one day to the
next; and
(ii) reaching a target utilisation rate on its plant portfolio every day.
To reach these objectives, agents may follow either a &price-raising' strategy,
by adding a random percentage to the bid(s) they submitted in the previous
trading day or a &price-lowering' strategy, by subtracting a random percentage.9
The agents may raise or lower bid prices to any level, between 0 and C1000.00,
but plants with high marginal costs of production must always bid higher prices
than plants, in the same portfolio, with lower costs of production. To replicate
the impact of forward contract cover, we assume that forward contracting
8 See Sutton and Barto (1998) for a de"nition and fuller discussion of the many di!erent forms of
reinforcement learning which have been developed.
9 In all the simulations discussed here, agents draw their random percentage values from a uniform distribution with a range $10% and a mean of 0% though we have tested other distributions
with little e!ect o the relative level of prices among the auction mechanisms tested.

574

J. Bower, D. Bunn / Journal of Economic Dynamics & Control 25 (2001) 561}592

Table 2
Summary of agent bidding rules and objectives
Rule 1. Self awareness
Agents receive feedback data from their own trading activities for the previous two trading days:
(i) Plant avoidable costs of production
(ii) Plant bid prices
(iii) Plant sales prices
(iv) Plant and total portfolio expected available capacity
(v) Plant and total portfolio sales volume
(vi) Plant and portfolio rate of utilisation
(vii) Plant and portfolio pro"t
(viii) Portfolio target utilisation
(ix) Portfolio target pro"t
Rule 2: Information restrictions
Agents do not know the past, current, or future, actions of other agents or the state of the market.
Rule 3: Objective functions
Agents have common objectives for each new trading day which are to achieve:
(i) At least their target rate of utilisation for their whole plant portfolio
(ii) A higher pro"t on their own plant portfolio, than for the previous trading day
Rule 4: Strategy selection
Agents submit bid price(s) for each plant in their portfolio, at the beginning of the current trading
day, using decision criteria in the following order of precedence:
(i) If the target rate of utilisation was not reached across the portfolio, on the previous trading
day, then randomly subtract a percentage from the previous day's bid price for each plant in
the portfolio
(ii) If any plant sold output for a lower price than other plants across the portfolio, on previous
trading day, then raise the bid price of that plant to the next highest bid price submitted
(iii) If total pro"t did not increase across the portfolio, on the previous trading day, then
randomly add or subtract a percentage from the previous day's bid price for each plant in the
portfolio
(iv) If pro"t and utilisation objectives were achieved across the portfolio, on the previous trading
day, then repeat the previous trading day's decision
Rule 5. Strategy restrictions
Agents can follow any strategy on condition that the bid prices in their plant portfolio are always:
(i) No less than C0.00
(ii) No more than C1000.00
(iii) Rounded to two decimal places
(iv) Higher for high marginal production cost plant than for low marginal production cost plant
in the portfolio

re#ects each generator's desire to guarantee itself a minimum level of market
share, or output, in a given period. For each agent we have therefore estimated
a minimum target rate of utilisation for its plants, expressed as a percentage of
expected total available MWh of capacity across its whole portfolio. From the

J. Bower, D. Bunn / Journal of Economic Dynamics & Control 25 (2001) 561}592

575

point of view of the simulation, if an agent failed to reach its target utilisation
rate on the previous trading day, then it lowers the bid price(s) on all of its plants
for the current trading day. Though this response disregards the potential
impact on pro"tability, and the success of previous strategies, the target utilisation rate is attached to an agent's portfolio, not to particular plant(s), so they
are still free to explore a wide range of bidding strategies which will satisfy both
pro"t and utilisation objectives. Finally, an agent can transfer a successful
bidding strategy, from one of its plants, to all other plants in its portfolio. This
favours agents with large plant portfolios as they naturally have more opportunities to experiment with, identify and adopt successful bidding strategies than
a single plant operator does. This is achieved by allowing agents to automatically raise the bid price on any plant, to the level of the next highest bid price
submitted, if it sold its output for less the marginal sales price achieved in the
portfolio on the previous trading day.
In practice each agent is continuously updating its pro"t objective, as the
simulation progresses, always using the previous trading day's pro"t as a
benchmark against which it compares the current day's pro"ts. By continuously
updating their pro"t objective, at the end of each trading period, agents
are forced to continuously compete against each other. As in the real world,
not all the agents can increase their pro"ts inde"nitely and, at some point,
a pro"t increase by one agent will cause a pro"t decrease for another agent.
When an agent su!ers a pro"t decrease it is prompted to abandon its
current bidding strategy and randomly look for a more successful one.
When it eventually "nds a better strategy, which might mean taking pro"t from
another agent, this would trigger a new strategy search by the a!ected agent,
and so on.
A criticism of the proposed bilateral model is that less information will be
available to participants than is currently available in the Pool. We have
eliminated this potential informational di!erence in our model so that we can
focus purely on the impact of alternative bidding and settlement arrangements.
We do this by assuming that agents know everything about their own portfolio
of plants, bids, output levels, and pro"ts, but nothing about other agents or the
state of the market. Their ability to capture and retain data is very limited, they
have no powers of strategic reasoning, and hence they exhibit a high degree of
bounded rationality.
4.2.3. The economic environment
In contrast to the supply side of the market, we assume that all the agents on
the demand side are price takers with no ability to in#uence the market through
strategic behaviour. For simplicity, we therefore model them as an aggregate
demand curve. To simplify the analysis, we have created a standardised daily
load pro"le, corresponding to the demand pattern corresponding to the 1997/98
NGC Winter peak Demand (1 December 1997).

576

J. Bower, D. Bunn / Journal of Economic Dynamics & Control 25 (2001) 561}592

We know that demand response to increasing prices is very low in the Pool
day-ahead market, which is one of the major criticisms of it, because of the
limited amount of demand-side bidding that occurs. Typically, about 750 MW
of demand-side bids are usually submitted between C80/MWh and 250/MWh.
An uncertain amount of private load management also occurs, which we assume
accounts for another 4000 MW of demand response, which is also likely to be
from industrial customers. For the purposes of our simulations, we have therefore assumed a linear load-shedding response of 25 MW for every C1 MWh that
SMP rises above C75/MWh. Therefore, at a SMP of C175/MWh we are assuming that a total of 2500 MW of demand-side response occurs. Our estimate of
short-term demand-side elasticity is much lower than that used by Green and
Newbery (1992), who assumed a load drop of 500 MW per C1/MWh. However,
it is close to the empirical estimates of Patrick and Wolak (1997) who calculate
demand elasticities between !0.1 and !0.3 for large industrial customers.
Given that they contribute about 15,000 MW of demand in the Pool this would
be equivalent to a load drop of 15}45 MW per C1/MWh on-peak on typical
winter day. Appendix B shows the daily load pro"le, as it would be at prices
below C75/MWh.

5. Simulation results
Using the agents, demand curve and hedging pro"le above, we simulated 750
working days of trading under the four di!erent sets of trading arrangements,
with summary statistics for each of the 24 h settlement periods calculated from
the "nal 250 working days of data. For the Pool model this means that over
50,000 separate daily bids decisions were simulated, while the bilateral model
simulates over 1.2 million hourly bids.10
Though data on the quantity, and distribution, of forward contract cover in
the Pool is commercially sensitive, we know that, in general, all independent
power producers (IPPs) with Combined Cycle Gas Turbine (CCGT) plant,
nuclear generators, and interconnector trade is almost fully contracted and we
assume a target utilisation rate of 100% for these companies. For Eastern,
National Power, and PowerGen we assume an average target utilisation rate of
60% across their plant portfolios. In our simulation of the Pool day-ahead
market, this accounts for 97.1% of total industry output, a value consistent with
levels of contracting reported anecdotally, and in Government inquiries (MMC,
1996a, b).
10 Agents are modelled as data arrays in Excel 97 and manipulated with Visual Basic. This allows
run speeds of approximately 8 min for 1000 working days (Pool), and 24 min per 1000 working days
(bilateral model), on a standard desktop PC equipped with a 400 MHz Pentium processor and
128 MB of RAM.

J. Bower, D. Bunn / Journal of Economic Dynamics & Control 25 (2001) 561}592

577

Simulated market-clearing prices, for each of the four trading arrangements, are summarised in Fig. 2. These show that the Pool with Pay SMP
settlement and single daily bids, as per the current trading arrangements,
produced the lowest peak clearing price at C99.22. In contrast, Pay Bid settlement with and hourly bids, as per the bilateral model produced the highest peak
clearing prices of C279.34. The two alternative sets of trading arrangements
produced intermediate peak clearing prices of C147.05 for Pay SMP settlement with hourly bids of C147.05 and for Pay Bid settlement with daily bids
of C114.42.
Fig. 3 shows the aggregate supply functions bid by agents under the four
di!erent sets of trading arrangements. Both of the Pay SMP settlement
simulations show most low cost plant is bidding at close to zero, a strategy seen
in the real Pool day-ahead market, which is truly emergent from the model
because this behaviour is not explicitly speci"ed in the agents' decision rules. In
contrast, using exactly the same decision rules, agents quickly learn to bid
a much #atter supply function, well above zero, when Pay Bid settlement
applies. This gives us con"dence that the model is capable of successfully
replicating both the actual, and potential, micro-level strategic behaviour in this
market.
We know from observing real bidding and from analysis done by Helm and
Powell (1992), Powell (1993), and Green (1996b), that the optimal bidding
strategy for generators with hedged plant, is to bid at short-run avoidable cost.

Fig. 2. Simulated market-clearing prices for alternative trading arrangements.

578

J. Bower, D. Bunn / Journal of Economic Dynamics & Control 25 (2001) 561}592

Fig. 3. Simulated supply functions for alternative trading arrangements.

Therefore, prices in the Pool day-ahead market do not re#ect purely
competitive bidding, because forward contract cover limits generator's ability
to exercise market power. In general, therefore, the higher the level of contracting, the lower the expected level of prices. We have been able to replicate
these well-known "ndings, summarised in Fig. 4, by changing the target
rate of utilisation for Eastern, PowerGen, and National Power. For both,
the Pool day-ahead simulation, and the bilateral model simulation, prices
fall as the target rate of utilisation (i.e. generators' desire for market share)
rises.
The results, described above, seem to contradict OFFER's expectation that
prices will fall as a result of the introduction of the bilateral model. Closer
inspection of the bidding strategies of individual plants in our simulations
reveals that this is due to two separate, but complementary, phenomena that
result in:
(i) Pay Bid settlement increasing the risk of overbidding by baseload generators, especially IPPs with small plant portfolios, which reduces competitive
pressure on generators with mid-merit plant; and
(ii) hourly bidding allowing generators to e!ectively segment demand into
on-peak and o!-peak hours, thereby extracting a greater proportion of the
consumer surplus than under daily bidding.

J. Bower, D. Bunn / Journal of Economic Dynamics & Control 25 (2001) 561}592

579

Fig. 4. Impact of target utilisation rate on simulated supply functions for Pool Pay SMP.

5.1. Analysis of Pay Bid versus Pay SMP ewect on price
In the Pool day-ahead simulations, the Pay Bid price is generally higher than
Pay SMP. The same result occurs in the bilateral model simulations with Pay
Bid producing on-peak and o!-peak prices higher than Pay SMP. In Table 3,
the sales-weighted average bid price for each generator has been calculated. This
con"rms that, under Pay SMP settlement simulation, the interconnectors, IPP
CCGT, and nuclear baseload operators bid all their plant at prices close to zero.
Table 3
Sales weighted average o!-peak and on-peak bid prices under alternative trading arrangements
Generator

Eastern
EDF
IPP/Other
Magnox Electric
Mission Energy
National Power
Nuclear Electric
Power Gen
Scot' & Southern
Scottish Power

Current pool
All hours

C23.30
C0.05
C1.76
C0.62
C38.67
C13.89
C1.19
C12.59
C1.65
C13.94

Pool Pay Bid
All hours

C104.13
C98.53
C98.46
C99.07
C120.42
C101.69
C83.89
C104.34
C102.29
C97.78

Bilateral Pay SMP

Bilateral Model

O!-Peak

On-Peak

O!-Peak

On-Peak

C16.25
C0.05
C4.96
C1.02
N/A
C10.13
C3.12
C5.97
C0.37
C25.30

C76.53
C0.05
C0.68
C3.00
C96.45
C65.47
C19.81
C70.15
C9.53
C16.13

C54.34
C42.75
C45.81
C35.66
N/A
C83.14
C30.09
C51.21
C43.25
C47.00

C183.49
C52.68
C71.78
C35.93
C68.60
C164.83
C28.41
C208.36
C54.20
C120.51

580

J. Bower, D. Bunn / Journal of Economic Dynamics & Control 25 (2001) 561}592

Conversely, under Pay Bid settlement simulation, the sales weighted average
bid price of baseload generators rises to a level similar to the mid-merit
generators because they are being forced to bid closer to the market clearing
price in order maximise their pro"ts. This change in behaviour is exactly what
OFFER wants to encourage, rather than allowing baseload plant to just bid
zero and leave the market price setting to mid-merit plant.
However, Table 4 shows that Eastern, National Power, and PowerGen
produce a greater percentage of total output under Pay Bid than under Pay
SMP settlement. It seems that moving to Pay Bid settlement does not increase
competitive pressure on the marginal plant, as is hoped, rather it diminishes it,
because the probability of baseload plant being underbid by mid-merit plant is
increased. Baseload generators bid zero in the current Pool day-ahead market in
order to eliminate this risk, and to guarantee that their plants keep running at all
times. This reduction in market sh