AGENT-HUMAN INTERACTIONS IN THE CONTINUOUS
DOUBLE AUCTION, REDUX
Using the OpEx Lab-in-a-Box to explore ZIP and GDX
Marco De Luca and Dave Cliff
Department of Computer Science, University of Bristol, Merchant Venturers Building
Woodland Road, Bristol BS8 1UB, U.K.
Keywords: Algorithmic trading, Continuous double auction, Experimental economics, Trading agents.
Abstract: In 2001, a team of researchers at IBM published a paper in IJCAI which reported on the first experiments
that systematically studied the interactions of human traders and software-agent traders in electronic
marketplaces running the continuous double auction (CDA) mechanism. IBM found that two software-agent
strategies, known as GD and ZIP, consistently outperformed human traders. IBM's results received
international press coverage, probably because the CDA is the mechanism that is used in the main electronic
trading systems that make up the global financial markets. In 2002, Tesauro & Bredin published details of
an extension to GD, which they named GDX, for which they wrote: "We suggest that this algorithm may
offer the best performance of any published CDA bidding strategy". To the best of our knowledge, GDX
has never been tested against human traders under experimental conditions. In this paper, we report on the
first such test: we present detailed analysis of the results from our own replications of IBM's human vs. ZIP
experiments and from our world-first experiments that test humans vs. GDX. Our overall findings are that,
both when competing against ZIP in pure agent vs. agent experiments and when competing against human
traders, GDX's performance is significantly better than the performance of ZIP.
1 INTRODUCTION
At the 2001 International Joint Conference on
Artificial Intelligence (IJCAI-01), a team of IBM
researchers presented a paper (Das, Hanson, Kephart
& Tesauro, 2001) that generated press coverage
around the world (e.g. Graham-Rowe, 2001). Das et
al.’s paper was the first to apply the laboratory
methods of experimental economics (e.g. Kagel &
Roth, 1997) to the systematic comparative
evaluation of adaptive autonomous software-agent
“robot” trader strategies, in controlled experiments
that pitted the robot traders against human traders in
a continuous double auction (CDA) mechanism. The
IBM team explored their own robot strategy, a
modified form of the Gjerstad-Dickhaut algorithm
(Gjerstad & Dickhaut, 1998) which we will refer to
as EGD (Extended GD), and a version of the Zero-
Intelligence Plus (ZIP) algorithm developed by Cliff
at Hewlett-Packard Labs (Cliff & Bruten, 1997). Das
et al. reported on results from six experiments
involving a number of human subjects being pitted
against a similar number of a particular type of
trading-agent: EGD in four experiments, and ZIP in
the remaining two. The results from all six of these
experiments were conclusive: the average efficiency
of the robot traders, i.e. their ability to enact
profitable transactions, was consistently higher than
that of the human traders, and this was true for both
the trading strategies. The IBM paper concluded
with the following memorable passage:
“[…] the successful demonstration of machine
superiority in the CDA and other common
auctions could have a much more direct and
powerful financial impact—one that might be
measured in billions of dollars annually”
Somewhat curiously, in the decade since that
paper was first published, as far as we can determine
no-one has yet reported on a replication of those
results. We speculate here that this is because, back
in 2001, to set up an experimental economics
laboratory such as that used by the IBM team
required a considerable investment. However, as the
real cost of personal computers (PCs) and data-
networking hardware has fallen dramatically in the
351
De Luca M. and Cliff D..
AGENT-HUMAN INTERACTIONS IN THE CONTINUOUS DOUBLE AUCTION, REDUX - Using the OpEx Lab-in-a-Box to explore ZIP and GDX.
DOI: 10.5220/0003293903510358
In Proceedings of the 3rd International Conference on Agents and Artificial Intelligence (ICAART-2011), pages 351-358
ISBN: 978-989-8425-41-6
Copyright
c
 2011 SCITEPRESS (Science and Technology Publications, Lda.)

past ten years, we observed that it is now possible to
re-create the necessary laboratory apparatus using
low-cost “netbook” PCs for a total cost of only a few
thousand dollars. With that motivation, we have
designed and implemented an experimental
economics laboratory network trading system, where
“trader terminal” netbooks communicate with a
central “exchange” server, with the potential for
multiple instruments to be traded simultaneously in
varying quantities, and with every quote in the
marketplace, and details of all transactions, written
to a database as a single “consolidated tape” record
of the trading events (to sub-second timestamp
accuracy) over the course of a trading experiment.
This trading system, which is called “OpEx” (from
Open Exchange) will be open-sourced under a
creative commons license in the near future (De
Luca, forthcoming 2011). In this paper, we report on
the use of OpEx to replicate IBM’s IJCAI-01 results
from testing human traders against ZIP and the most
recent evolution in the “GD” class of algorithmic
traders: GDX (Tesauro & Bredin, 2002). To the best
of our knowledge, these are the first results from
testing GDX against humans. We find that our
results agree with IBM in the respect that the GDX
and ZIP robot traders consistently out-perform the
human traders, but our results differ from IBM’s in
that we find that GDX outperforms ZIP, while in the
IBM study ZIP slightly outperforms EGD on
average. Our results are also in line with those
achieved by Tesauro & Bredin: in pure robot vs.
robot competitions, GDX outperforms ZIP and
proves to be a major improvement of the original
GD algorithm.
2 BACKGROUND
Today, the vast majority of financial products are
traded electronically: following exact rules, buyers
and sellers, collectively known as traders, interact in
a common virtual “marketplace” to trade those
products. The numerous organisations that are in
place to allow electronic trading of financial
securities are known as exchanges, or sometimes
markets. The set of rules that define the exchange
process between traders on a market forms its
market mechanism, of which the continuous double
auction (CDA) is the most used due to its high
efficiency:
“Markets organised under double-auction trading
rules appear to generate competitive outcomes
more quickly and reliably than markets organised
under any alternative set of trading rules.” (Davis
& Holt, 1993)
In a CDA, traders can make bids and accept
offers asynchronously at any time during the trading
day (that is, the fixed-duration trading period during
which trading is allowed). All the offers are usually
publicly visible by all market participants, and a
trade is made whenever the outstanding bid is
greater than or equal to the outstanding ask.
Although it is made up of simple rules, the
nonlinearities of the CDA are too complex to be
analysed by traditional mathematical methods such
as game theory: as a results, researchers have turned
to empirical approaches.
In his Nobel-prize-winning work, Vernon Smith
(1962) ran several experiments with human traders,
and demonstrated that markets governed by the
CDA can reach close-to-optimal efficiency. Also, he
proved that transaction prices converge to the
market’s theoretical competitive equilibrium price,
where the supply and demand curves intersect.
Furthermore, he found that if the supply and demand
of markets suddenly changed, the transaction prices
would rapidly converge to the new equilibrium
price. In many of his experiments, Smith studied the
dynamics of CDA-based markets by assigning one
unit to sell(buy) at no less(more) than a specific
price to each of the traders. The price of the unit,
known as limit price, represents the maximum
amount of money l a buyer can spend to buy the
unit, or the minimum value c for which a seller can
sell the unit. As a consequence, buyers make a profit
l-p if they buy at a price p that is less than their limit
price, whereas sellers make a profit p-c if they sell
for a price p higher than their limit price. The limit
prices are private, each trader knowing only her
limit. The traders interact by quoting the price at
which they are willing to trade their units. In Smith’s
early experiments this happened by speaking the
number out loud, thus the public quotes in a CDA
are often referred to as shouts. A random player is
selected every turn to make a shout, and the game
finishes after a fixed number of turns. Following the
rules of the CDA, a trade occurs when the
outstanding bid is greater than or equal to the
outstanding ask. Smith measured the performance of
a trader in terms of allocative efficiency, which is the
total profit earned by the trader divided by the
maximum theoretical profit of that trader, expressed
as a percentage. The maximum theoretical profit of a
trader is the profit that trader could have made if all
the market participants would have traded their units
at the theoretical competitive market equilibrium
price. A further measure of the performance of a
ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence
352

market is the profit dispersion: this is defined as the
cross-sectional root mean squared difference
between the actual profits and the maximum
theoretical profits of individual traders. Formally, if
a
i
is the actual profit earned by trader i, and p
i
is the
theoretical equilibrium profit for that trader, then for
a group of n traders the profit dispersion is given by:

1


(


−

)



(1)
3 OPEN EXCHANGE
We ran our experiments on Open Exchange (OpEx),
an experimental algorithmic trading platform
developed by De Luca (forthcoming 2011). OpEx
was designed to resemble closely the structure and
the behaviour of modern commercial financial-
market electronic trading systems, and to be generic
enough to support experimental economics
simulations of arbitrary complexity. Figure 1
illustrates the interaction between the core
components in a simple configuration. The
connections between the components on the left
Figure 1: An instance of Open Exchange. The solid lines
and the dotted lines represent the flow of order data,
respectively the requests and the replies. The sparsely
dotted lines indicate the market data flow, from the
Exchange back to the order generators through the Market
Data Bus.
hand side show the flow of order data. Orders
represent the agents' instructions to buy or sell a
specific quantity of a given product at a particular
price condition. Human traders enter their orders in
the Trading GUI, a graphical application that allows
users to view the market order book (i.e. the
descending-ordered list of currently outstanding
bids, and the ascending-ordered list of currently
outstanding offers), their “blotter” (personal history
of orders and trades), and their assignments. Agents,
on the other hand, produce orders automatically,
without the need of human intervention, on the basis
of the market conditions that they observe. Once
generated, orders are sent to the Order Manager,
which routes them to the appropriate order processor
(in this example, the single Exchange) depending on
the destination specified by the sender. Once
received by the Exchange, orders are processed
according to the specific order matching logic
implemented (the order matching logic that we will
cover in detail here is the price-time priority
matching logic, which constitutes the foundation of
the CDA) and order completion data is passed back
to the Order Manager, which in turn dispatches it to
the appropriate sender. It is worth noting that order
data are private, as only the originator of an order
receives the order completion data relative to that
specific order, which will let him/her know its
progress. Conversely, market data are published on
the Market Data Bus and can be seen by every
market participant.
4 AGENTS
In the Open Exchange framework, automated
trading agents are implemented as individual plug-
ins running on an instance of the Agent Host. In line
with the distributed architecture of OpEx, there can
be multiple instances of the Agent Host, each one
running a particular set of Agents. Every Agent
implements one specific algorithm and has its own
configuration settings, loaded at start-up. One
instance of the Agent Host is capable of running
multiple instances of the same Agent, so that more
than one automated trader following a specific
strategy can participate in the market
simultaneously. The behaviour of an OpEx Agent
consists of cyclically listening to stimuli and reacting
to them sequentially by performing one or more
actions. Agents are idle as they wait for the next
stimulus, whereas they perform calculations and
they can send a signal to the market when they are
active. Each stimulus represents a precise event (e.g.
“the activation timer has expired”, “an order has
been sent”, or “there has been a trade”) and it is
produced by a specific source asynchronously.
Unprocessed stimuli are convoyed to a common
collector, and then the resulting queue, sorted
chronologically, is processed sequentially. Our
choice of timing mechanism is consistent with the
previous IBM work (Das et al., 2001), where similar
timing rules were used to regulate the activity of the
Agents. However, while the results presented in
AGENT-HUMAN INTERACTIONS IN THE CONTINUOUS DOUBLE AUCTION, REDUX - Using the OpEx
Lab-in-a-Box to explore ZIP and GDX
353

(Das et al., 2001) are from experiments run using
two different timer periods (“fast”, 1 second; and
“slow”, 5 second) for the different algorithms, in our
work reported here we used the same timing across
all the experiments in order to simplify the
comparison of the performances of the different
trading agents. In particular, our Agents primary
timer period is set to 1 second, equivalent to the
“Fast” configuration used in (Das et al., 2001). On
the other hand, OpEx schedules the activity of the
Agents in a much more basic way when running in
“Discrete Event Simulator” (DES) mode. DES
simulations are turn-based (300 turns in one trading
day), and at each turn only one Agent is chosen at
random among the active Agents, each of which has
the same probability of being selected.
4.1 ZIP
In 1996, Cliff invented the Zero-Intelligence Plus
(ZIP) algorithm to investigate the minimum level of
intelligence required to achieve convergence to
market equilibrium price (Cliff & Bruten, 1997). ZIP
has been used in several subsequent studies, e.g.
(Tesauro & Das, 2001) and (Das et al., 2001), as a
benchmark for evaluation of strategy efficiency, and
it was subsequently extended to ZIP60 by Cliff
(2009). Each ZIP trader agent maintains a real-
valued profit margin and employs simple heuristic
mechanisms to adjust their margin using market
data. In this context, the profit margin represents the
difference between the agent’s limit price and the
shout price, which is the price that the agent sends to
the market to buy or sell the commodity. By
observing market events, ZIP buyers (sellers)
increase their profit margin, and therefore make
cheaper bids (more expensive offers), when a trade
at a lower (higher) price than their current shout
price occurs. Conversely, ZIP buyers that observe an
accepted offer (bid) at a price higher (lower) than the
one they have put on the market move towards that
price by lowering their profit margin, that is bidding
(offering) a higher (lower) price. The same applies
to buyers (sellers) that witness a rejected bid (offer)
at a higher price than the one they are advertising.
The profit-margin adaptation rule used in the ZIP
algorithm to dynamically respond to the market
conditions is based on the Widrow-Hoff “delta rule”
with an additional noise-smoothing “momentum”
term. The profit margin of the ZIP traders is adjusted
by a small random quantity, proportional to the
learning rate of the individual agent. Consistently
with (Preist & Van Tol, 1998) and (Das et al., 2001),
we altered the original ZIP implementation to fit in
the OpEx infrastructure by introducing an “inactivity
timer”. The timer triggers a procedure that adjusts
the shout price of the agents moving it towards the
best price on the opposite side of the order book. As
a result, the piece of information “nothing is
happening in the market” is used by the agents as an
additional pricing heuristic.
4.2 GD/GDX
In 1998 Gjerstad & Dickhaut introduced a bidding
algorithm, now widely referred to as GD, centred on
“belief” functions that agents form on the basis of
observed market data. GD agents collect the orders
(rejected shouts) and trades (accepted shouts)
occurred during the last M trades, and store them in a
history H. When a GD agent prices an order, from
the history H it builds the belief function f(p), which
represents the probability that an order at price p
will result in a trade. For example, the belief
function for a GD buyer is:

(

)
=

(

)
+

()

(

)
+


(

)
+
(

)
(2)
Here, TBL(p) represents the number of accepted
bids found in H at price ≤ p, AL(p) is the number of
asks in H with price ≤ p, and RBG(p) is the number
of rejected bids in H at price ≥ p. Note that f(p)
depends on H, and therefore it can potentially
change every time a market participant sends an
order to the market. Because f(p) is defined only for
some values of p, the function is interpolated to
provide values over the domain of all the valid
prices. Finally, the price p that maximises the
product of the interpolated f(p) and the profit
function of the agent (equal to l - p for buyers and p -
l for sellers) is chosen as the order price. The
original GD algorithm was modified by Tesauro &
Bredin (2002), who christened their version “GDX”.
GDX substantially differs from GD in that it makes
use of Dynamic Programming (DP) to price orders.
The pricing function takes into account both the
effect of trading the current unit immediately, and
the effect of trading it in the future, discounting the
latter by a parameter γ. As a result, GDX agents do
not just maximise the immediate profit, but instead
optimise the pricing process in order to achieve
overall higher returns over the entire trading period.
5 EXPERIMENTAL SETUP
All of our human vs. robot experiments involved 6
human traders and 6 robot traders, both equally split
ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence
354

Figure 2: Trade price time series for a humans-vs.-GDX experiment. The vertical lines represent the start of a new round.
The 10 rounds of 3 minutes each were divided into 5 phases, each of which with its own set of limit prices. The theoretical
equilibrium price for each phase is indicated by the horizontal dashed lines. Trades between two humans are marked with
open squares, between two agents with open circles, and between an agent and a human with solid circles. Mean efficiency
per phase (vertical bars) and per rounds are shown for Agent Buyers (AB), Agent Sellers (AS), Human Buyers (HB) and
Human Sellers (HS).
into 3 buyers and 3 sellers, a structure used in the
original IBM experiments. Before each experiment,
the human subjects were briefed about the rules, and
were given some time to familiarise with the Sales
Trading GUI (briefing and tutorial typically took
less than 30 minutes). The subjects had no previous
professional experience in electronic trading, and
they were only allowed to participate in one
experiment. We motivated all 6 participants by
giving each of them a token worth £20, plus a bonus
of £40 and £20 to the first and the second best
trader, respectively. An experiment consisted of 10
consecutive “rounds” 3 minutes long. At the
beginning of a round, each of the 12 players
received a fresh supply of 13 units to buy or to sell
during that round, according to their role. At the end
of the round the unused units were discarded,
without any profit or loss for the traders. Players had
to trade their units sequentially, and the sequence of
their limit prices was arranged in an arithmetic
progression. Only 3 “generator” sequences were
actually used to produce the prices for all the
players: a human and his/her robot counterparty had
the same limit prices; and buyers and sellers share
the same values except for the order, that is
increasing for sellers and decreasing for buyers. The
progressions had the same slope, and they were
chosen so that each player had approximately the
same maximum theoretical surplus in a given round.
In line with (Das et al., 2001), we introduced market
shocks by periodically altering the limit prices
adding or subtracting a constant to them every 2
rounds. Thus, a 30 minutes simulation was
constituted by 5 consecutive trading periods at
different equilibrium prices.
6 EXPERIMENTAL RESULTS
6.1 Agents vs. Humans
The results of the four agent-human experiments,
summarised in Table 1, present several significant
findings, all of which are in line with (Das et al.,
AGENT-HUMAN INTERACTIONS IN THE CONTINUOUS DOUBLE AUCTION, REDUX - Using the OpEx
Lab-in-a-Box to explore ZIP and GDX
355

Table 1: Summary of the four agent-human experiments. For each experiment, the table displays: the strategy employed by
all six agents; the percentage of trades made between two Agents, an Agent and a Human, and two Humans; the average
efficiency of Agents and Humans; the percentage difference between Agents surplus and Humans surplus; the market
efficiency and the profit dispersion. The mean maximum theoretical profit per trader per simulation is 2107. Lower profit
dispersion and higher mean efficiency values are better.
Experiment Trades Performance Market
ID Strategy A-A A-H H-H Eff(A) Eff(H) Δ Profit (A-H) Eff Profit Disp
UoB01 ZIP 35% 35% 30% 1.010 0.965 5% 0.987 536
UoB04 ZIP 39% 30% 32% 1.037 0.931 11% 0.984 468
UoB05 GDX 36% 40% 24% 1.055 0.789 36% 0.923 707
UoB06 GDX 33% 44% 22% 1.074 0.809 35% 0.943 704
2001).
First, the agents as a group consistently
outperformed the humans in all four experiments:
the total surplus extracted from the market by the
agents was on average ~22% more than the total
surplus extracted by the human counterpart. Also,
the efficiency achieved by the agents is constantly
above 100%, which evidently implies that the agents
managed to exploit human flaws.
Second, there was a substantial interaction
between agents and humans: on average, ~37% of
the trades happened between an agent and a human,
which confirms that the humans as a group were
well integrated in the mixed humans-agents market.
Third, we found that for each experiment, either
all the buyers (but one) did better than all the sellers,
or vice versa. Because this pattern was found neither
in the numerous robot vs. robot experiments we ran
under identical conditions, nor in the many human
vs. human trials documented in (Smith, 1962), we
speculate that this asymmetry is due to the
heterogeneous nature of our market.
Finally, our analysis shows that although GDX
agents as a group achieve higher values of allocative
efficiency than ZIP agents when competing against
humans, both the overall market efficiency and the
profit dispersion values are better for ZIP.
6.1.1 GDX Agents vs. Humans
The trade price time series of the human vs. GDX
experiment UoB06 is shown in Figure 2. We will
refer to this specific experiment, although the
observations we made on UoB05 are very similar.
The dashed vertical lines separate the 10 trading
periods, whereas the dashed horizontal lines mark
the theoretical equilibrium price p
*
. The time series
exhibits a recurring pattern of convergence towards
a price that is often somewhat lower than p
*
. Most of
the trades were made at lower prices than p
*
, since
buyers closed deals at reasonably lower prices than
their limit prices, and therefore kept a higher profit
margin than their sellers counterparty. This is
confirmed by the fact that the five best traders in
terms of mean allocative efficiency are buyers, for
both the human vs. GDX experiments.
A more detailed analysis of the efficiency per
trading period reveals that the discrepancy between
buyers and sellers is accentuated by the raising of
the equilibrium price (e.g. between trading periods 6
and 7), and attenuated by the drop (e.g. between
trading periods 2 and 3, and 8 and 9). We explained
this by looking at the first few trades made in the
trading period following the market shock: their
prices tend to remain close to the previous value of
p
*
, resulting in better opportunities for buyers or for
sellers, if there was a raise or a drop of p
*
respectively. This confirms that the GDX strategy
requires a few samples before it can adapt to the new
market condition.
6.1.2 ZIP Agents vs. Humans
Figure 3 illustrates the first four trading periods of
experiment UoB04, which are quite representative
for the two human vs. ZIP experiments we ran. By
visual inspection, it can be verified that human-ZIP
markets display better capabilities of tracking the
equilibrium price, as convergence to p
*
is more
pronounced than in human-GDX markets. It is clear
that the patterns displayed by this time series are
quite different from those in Figure 2. It can be
noted that, qualitatively, the shape of the time series
is reasonably consistent across the trading periods,
and that the curve presents a higher price excursion
in a shorter time than GDX before converging to p
*
.
We ran a detailed quantitative analysis of the time
series to confirm this, and found that the mean trade
time relative to the trading period is ~45 seconds for
ZIP-humans and ~69 seconds for GDX-humans
markets. Moreover, by isolating the trades between
two agents (A-A), between two humans (H-H), and
between a human and an agent (A-H), we found that
the mean trade time of the three types of trades is
ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence
356

Figure 3: The first four trading periods of experiment UoB04.
Table 2: Summary of three sets of robot vs. robot experiments between GDX & ZIP agents. For each set of experiments, the
table presents: the type of experiment, the value of the discount parameter γ, the number of experiments won by the two
agents, and the mean number of rounds per experiments won by GDX (±1 s.d.).
Type γ ZIP GDX GDX rounds won
DES 0.0 46 1011 8.567 (± 1.817)
DES 0.9 14 985 9.094 (± 1.273)
RT 0.9 316 654 5.736 (± 1.518)
consistently higher in GDX than in ZIP. Also, the
mean trade time of A-A trades is the smallest and
that of H-H trades is the largest consistently across
trading periods in the experiments involving ZIP,
while this relationship does not hold for some
trading periods of experiments UoB05 and UoB06.
6.2 Robots vs. Robots
In order to further benchmark ZIP and GDX, we ran
three sets of experiments between the two agents, in
a pure robot vs. robot market. The results are
outlined in Table 2.
Qualitatively in line with (Tesauro & Bredin,
2002), GDX clearly outperforms ZIP in discrete
event simulations, both when run in optimal mode (γ
= 0.9) and when degenerated to GD (γ = 0); in
particular, the performance of GDX improves
slightly for γ = 0.9. However, the win-lose ratio
changes radically when the experiment is run in
Real-Time (RT) mode, that is using the same set-up
described for human vs. robot markets. This is also
confirmed by the values of the mean number of
rounds won by GDX.
We speculate that the difference between the
DES and the RT results is mostly due to the very
nature of the two simulators: DES simulations are
essentially single-threaded, and the agent selected
for the current move has a virtually unlimited time to
perform its calculation before ending the move.
Conversely, each agent is represented by (at least)
one thread in a RT simulation: agents are woken up
asynchronously, therefore two or more of them may
happen to operate “simultaneously” (compatibly
with the software and hardware scheduling policies
in force on the system running the simulation). This
discrepancy is particularly relevant when comparing
GDX and ZIP because the calculations performed by
the latter are much more light-weighted than those
performed by the former: while the GDX strategy
may fare overwhelmingly better than ZIP if it is
given all the required time to execute the pricing
calculations, the difference between the performance
of the two is dramatically reduced when time is
critical, and the fastest agent to hit a price makes
more profit.
7 DISCUSSION & CONCLUSIONS
We were pleased to employ our low-cost, portable
experimental economics laboratory to, for the first
time ever, pit humans against what is known to be
the most evolved version of the “GD” class of
algorithms. The results we obtained are, at the best
AGENT-HUMAN INTERACTIONS IN THE CONTINUOUS DOUBLE AUCTION, REDUX - Using the OpEx
Lab-in-a-Box to explore ZIP and GDX
357

of our knowledge, unique, and they present several
noteworthy characteristics.
The application of Dynamic Programming
techniques indeed proves its validity in terms of
overall efficiency achieved by the agents as a group,
against both human and automated rivals. The
advantage of GDX over its predecessor is also
confirmed by comparing our results to those realised
in the IBM study, which present consistently lower
values of mean efficiency of the trading agents.
On the other hand, human-ZIP markets certainly
display better overall performance, in terms of
market efficiency and profit dispersion. This
suggests that ZIP agents would be better companions
for human traders in a CDA-regulated market where
the objective is to maximise the whole profit
extracted, whereas GDX would be a better choice in
a scenario where humans and agents are pitted
against each other as two separate teams, each one
trying to exploit their rivals’ weaknesses to
maximise their own profit.
Moreover, we note here that several features of
the market dynamics observed in our experiments
deserve further investigation: the curved price
trajectories and their convergence to the theoretical
equilibrium price; the distinct separation between
buyers and sellers in terms of overall performance;
and the effect of timing constraints on the
algorithmic traders.
Ultimately, it would be interesting to test our
algorithmic traders in two additional scenarios,
compatibly with the time and money issues related
to running the experiments. One where the period of
the agency interventions is forced to be comparable
to the estimated reaction time of the human traders:
this would reveal in what measure the superiority of
the agents is bound to their speed. And a second
scenario where professional traders are used instead
of amateurs, which would explain whether solid
trading skills in the global financial markets make
any difference in a competition against automated
traders.
ACKNOWLEDGEMENTS
The authors are very thankful to the UK Engineering
and Physical Sciences Research Council (EPSRC:
http://www.epsrc.ac.uk) for funding the equipment
used for this research, and to the many students and
staff at University of Bristol who took part in the
experiments. Both authors are members of, and
receive financial support from, the UK Large-Scale
Complex IT Systems (LSCITS) Initiative. Further
details at http://www.lscits.org.
REFERENCES
Cliff, D., 2009. ZIP60: Further explorations in the
evolutionary design of online auction market
mechanisms. IEEE Transactions on Evolutionary
Computation, 13(1):3-18.
Cliff, D., Bruten, J., 1997. Minimal Intelligence Agents
for Bargaining Behaviours in Market-Based
Environments. Technical report, HPL-97-91, Hewlett
Packard Labs.
Das, R., Hanson, J. E., Kephart, J. O., Tesauro, G.,
2001.Agent-Human Interactions in the Continuous
Double Auction. In: Proc. 17th International Joint
Conference on Artificial Intelligence, pp. 1169—1176.
Davis, D. D., Holt, C. A., 1993.Experimental Economics.
Princeton University Press.
De Luca, M. Forthcoming 2011. OpEx: An Open
Exchange Experimental Economics Laboratory in a
Box. Technical Report in preparation.
Gjerstad, S., Dickhaut, J., 1998.Price Formation in Double
Auctions. In: Games and Economic Behaviour, vol.
22, pp. 1—29.
Graham-Rowe, D., 2001. Bots knock the socks off city
slickers. New Scientist, 11 August 2001.
Kagel, J., Roth, A., 1997.The Handbook of Experimental
Economics. Princeton University Press.
Preist, C., Van Tol, M., 1998. Adaptive agents in a
persistent shout double auction. In: Proc. of the First
International Conference on Information and
Computation Economies, pp 11—18, ACM Press.
Tesauro, G., Das, R., 2001. High-performance bidding
agents for the continuous double auction. In: Proc.
Third ACM Conference on Electronic Commerce, pp.
206—209.
Tesauro, G., Bredin, J. L., 2002. Strategic Sequential
Bidding in Auctions using Dynamic Programming. In:
Proc. 1st International Joint Conf. on Autonomous
Agents and Multiagent Systems, pp. 591—598.
Smith, V., 1962. An Experimental Study of Comparative
Market Behaviour. In: Journal of Political Economy,
vol. 70, pp. 111—137.
ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence
358