Using Time as a Strategic Element in

Continuous Double Auctions

Marcel Neumann, Karl Tuyls, and Michael Kaisers

Maastricht University, P.O.Box 616, 6200 MD Maastricht

Abstract. Numerous pricing strategies have been proposed and incor-

porated into software trading agents. Early simulation experiments have
shown that markets are efficient even with only a few traders placing ran-
domly priced offers at each time step using the Zero-Intelligence strat-
egy. This article investigates the strategic effect of timing on the per-
formance of trader profits and market efficiency. The trading strategies
Zero-Intelligence and Zero-Intelligence-Plus are enhanced by new timing
strategies, replacing their heuristics with random and strategic behav-
ior. As expected, agents make less profit with random timing against
the heuristic. However, market efficiency remains the same, confirming
that continuous double auctions are highly efficient mechanisms even
with traders placing profitable offers with random prices and timing.
Furthermore, Zero-Intelligence-Plus agents also achieve high efficiency,
but strategic timing reduces the risk of being exploited by trading far
from the equilibrium price. Thus, there is a clear individual incentive to
exploit timing strategically.

Keywords: Continuous Double Auctions, Zero-Intelligence, Timing

1 Introduction

Brokers of the New York Stock Exchange (NYSE) exchange about 1.6 billion
shares worth $45 billion per day. The dominant market mechanism for the ex-
change of commodity goods is the continuous double auction. Each trader is
assumed to have private valuations for the goods to be traded that are only
known to itself. Traders place offers at self-elected times to indicate that they
are willing to buy (bid) or sell (ask) a good at a certain price. Market clearing
occurs each time a new offer arrives. The highest open bid is then tried to be
matched with the lowest open ask.
Classic economic theory predicts that the price for one unit of a commod-
ity settles at an equilibrium price where the quantity demanded is equal to
the quantity supplied. Smith [8] showed that the transaction price time series
has a tendency to converge towards the market equilibrium price even with a
relatively small number of human traders. Gode & Sunder [4] introduced two
types of “Zero-Intelligence” computer trading strategies that follow a random
pricing scheme. Market efficiency was shown to reach values up to 99%. Cliff
[1] followed with a more sophisticated trading strategy called “Zero-Intelligence-
Plus”. Traders are able to observe market activity and adjust their offer prices
accordingly. Both trading strategies apply the same implicit heuristic about the
timing of offers. Offers are placed as soon as their price is being determined.
This article extends both trading strategies Zero-Intelligence and Zero-Intelligence-
Plus with alternative timing strategies and compares their performance through
a series of simulation experiments with their original counterparts. The complex-
ity of the strategies used in this article is summarized in Table 1. The concept of
random pricing by Zero-Intelligence agents is transferred and similarly applied
to time in order to investigate the performance of agents that do not use time
strategically. The complex form of pricing by Zero-Intelligence-Plus agents is
enhanced with a sophisticated timing strategy to investigate the performance of
agents that use time as a strategic element. This allows to evaluate the perfor-
mance of traders and markets with varying strategic reasoning about time.
The outline of this paper is as follows. Section 2 introduces the trading strate-
gies to be investigated together with their enhancements and performance mea-
sures. Section 3 describes the experiments. It gives a description of the simulation
environment and the specific experimental setups together with their results. Fi-
nally, the last section draws the conclusions.

2 Background and Methodology

Trading Strategies
Gode & Sunder [4] introduced two “Zero-Intelligence” (ZI) trading strategies.
A trader of the first version submits random offers independently, identically,
and uniformly distributed over the range of feasible trading prices from 1 to the
maximum trading price enforced by the market mechanism and is referred to as
“ZI unconstrained” (ZIU). The second version is subject to a budget constraint
that prevents it from engaging in loss making transactions. Thus, buyers submit
the uniform random bids between 1 and the valuation of the unit in question.
Sellers submit the uniform random asks between the cost of the unit in question
and the maximum trading price of the market. It is referred to as “ZI with
constraint” (ZIC). Both ZIU and ZIC submit offers as soon as they are computed.
ZI traders are considered to behave randomly. However, the originally pro-
posed “Zero-Intelligence” traders submit offers at every round and thus implicitly
use the heuristic always submit. “Zero-Intelligence-Time” (ZIT) traders combine
the random pricing of ZI traders with random timing, picking a time t∗ for the

Table 1. This table lists the sophistication of the strategies’ reasoning about the price
and time. Each of the strategies is explained in Section 2.

Price\Time Random Heuristic Strategic

Random ZIUT ZIU
Rational ZICT ZIC
Heuristic ZIP ZIPT
submission of the intended offer, such that t∗ is distributed uniformly over the
anticipated remaining time of the auction.
The trading strategy “Zero-Intelligence-Plus” (ZIP) proposed by Cliff [1]
allows agents to observe market activity and adjust future offer prices accord-
ingly. Events in the market that suggest higher profit for a trader result in more
profitable offer prices. Similarly, events that show the competitiveness of other
traders results in less profitable offer prices. At last, offer prices converge towards
unmatched offers in order to stay on a competitive basis. The reader is referred
to [1] for a full specification.
“Zero-Intelligence-Plus Time” (ZIPT) enhances the trading strategy of ZIP
with strategic timing behavior. The bid-ask-spread is defined by the difference
of the price of the lowest open ask and the highest open bid. The size of this
spread is used as an indicator of whether prices are going to change in the near
future. Offers by ZIPT agents are only submitted when the size of the bid-ask-
spread has fallen below a threshold value. The use of this threshold is inspired by
the Kaplan strategy that has won the 1990 Double Auction Tournament despite
being among the simplest strategies [7].
Performance Measures
The profit of a trader i considers all his successful transactions and measures
them against the private values λi,j of all units j. It is given by
X
pri = k(λi,j − pj ),
j∈traded

k = 1 if i ∈ Buyers, k = −1 if i ∈ Sellers, where pj denotes the transaction

price of unit j.
The coefficient of convergence α introduced by Smith [8] measures how close
a group of traders trades to the theoretical equilibrium. It is given in equation
1 and it is defined as the standard deviation of the actual trade prices pi from
the equilibrium price p0 as a percentage of it.
v
u n
t1
100 u X
α= (pi − p0 )2 (1)
p0 n i=1

The allocative efficiency e is a measure of performance of an entire market. It is

given by equation 2 and defined as the ratio of total actual profit and theoretical
profit. Total actual profit is the sum of profits made by each trader while the
theoretical profit is the sum of buyers’ sb and sellers’ ss surplus.
P
pri
i∈traders
e= (2)
ss + sb
These three measurements give insight into the individual profit, the collective
efficiency, and the convergence of the market to equilibrium prices. They provide
the basis for the experiments presented in the following section.
3 Experiments
The experiments are conducted using a discrete-event simulation. A simulation
process runs for six periods. During each period, traders are allowed to submit
offers at every time step, and a period is closed after 20 time units without new
offers. Traders receive a list of private values at the beginning of each period.
The distributions of private values are given in Figures 1, 2, 3 and 4. Offers
are for a single unit and subject to the NYSE spread improvement rule. Only
one open offer is allowed per trader. A transaction occurs when the price of the
highest open bid and the lowest open ask cross each other.

3.1 ZI vs. ZIT

A first set of experiments involves homogeneous populations of either ZIU, ZIC,
ZIUT and ZICT traders. The population consists of six buyers and sellers re-

Supply and demand functions of market 1 Supply and demand functions of market 2
200 Supply 200 Supply
Demand Demand
Equilibrium Equilibrium
price=80 price=70

150 150
Price

Price

100 100

50 50

0 0
0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70
Quantity Quantity

Fig. 1. Supply and demand functions intro- Fig. 2. Supply and demand functions intro-
duced by Gode & Sunder [4]. Sellers’ the- duced by Gode & Sunder [4]. Sellers’ the-
oretical profit = 1050, Buyer’s theoretical oretical profit = 500, Buyers’ theoretical
profit = 500. profit = 1000.

Supply and demand functions of market A

400 Supply
Demand Supply and demand functions of market B
Equilibrium
350 price=200 200 Supply
Demand
Equilibrium
price=100
300

250 150
Price

200
Price

150 100

100

50 50

0
0 20 40 60 80 100 120
Quantity
0
0 10 20 30 40 50 60 70
Quantity

Fig. 3. Supply and demand functions intro-

duced by Cliff [1]. Sellers’ theoretical profit Fig. 4. Seller’s theoretical profit = 900,
= 3750, Buyers’ theoretical profit = 3750. Buyer’s theoretical profit = 900.
spectively. The simplification by Gode & Sunder [4] of dropping all open offers
after a transaction is adopted for the sake of comparison.
Gode & Sunder [4] report that ZIU traders are able to achieve an allocative
efficiency of 90% for market 1 and 2. ZIC traders are able to raise this value to
99.9% for market 1 and 99.2% for market 2. It is expected that the two types of
traders will do equally good in this set-up of experiments. Parsons [5] noted that
the efficiency depends on the distribution of private values only. Since ZIUT
traders should be able to trade all their units, they should achieve the same
efficiency value as ZIU agents. On the other ZICT traders should achieve lower
values of allocative efficiency because of a higher risk of extra-marginal units to
be traded.
A second set of experiments is concerned with heterogeneous populations of
ZI and ZIT traders. ZIU traders are confronted with ZIUT traders while ZIC
traders are confronted with ZICT traders. The population consists of twelve
buyers and twelve sellers, equally divided into the respective type of traders.

Results Figure 5 shows the transaction price time series in market 1 with a
population of ZIU traders. The shape of the graph is identical to the one created
by Gode & Sunder [4]. The pattern can be stipulated as being random. This is
also the case for market 2 and B.
The transaction price time series in market 1 with a population of ZIC traders
is shown in Figure 6. Again, the shape of the graph is identical to the one created
by Gode & Sunder [4]. The price series is less volatile than the price series of
the ZIU traders. Also, the price series converges towards the equilibrium price
within each period. The observed pattern is similar for market 2 and B.
Transaction price time series in market 1 with a population of ZIUT traders
is shown in Figure 8. The shape of the graph is similar to Figure 5. A random
pattern can be observed while periods take longer to finish because traders do

Transaction price time series ZIU Transaction price time series ZICT
200 Trades 200 Trades
Period range Period range
Equilibrium Equilibrium
price=80 price=80

150 150
Price

Price

100 100

50 50

0 0
0 50 100 150 200 250 0 100 200 300 400 500 600
Time Time

Fig. 5. Transaction price time series for Fig. 6. Transaction price time series for
market 1 with 6 buyers and 6 sellers of ZIU market 1 with 6 buyers and 6 sellers of ZIC
traders (heuristic timing). traders (heuristic timing).
 Traders Market 1 Market 2 Market B
x̄ (s) x̄ (s) x̄ (s)
ZIU 90.32 (± 0.00) 90.00 (± 0.00) 0.00 (± 0.00)
ZIC 99.17 (± 0.23) 98.96 (± 0.27) 97.21 (± 0.61)
ZIU-T 90.32 (± 0.00) 90.00 (± 0.00) 0.00 (± 0.00)
ZIC-T 99.12 (± 0.24) 98.91 (± 0.28) 96.95 (± 0.67)

Fig. 7. Sample mean and sample standard deviation of the efficiency of ZI traders,
sample size = 1000.

not submit offers as soon as possible anymore but randomly distributed over the
least remaining time of the period. The pattern is similar for market 2 and B.
Figure 9 shows the transaction price time series in market 1 with a population
of ZICT traders. The price series looks similar than the one in Figure 6. It still
converges to the equilibrium price but not as gradually as it was the case for
ZIC traders.
Figure 7 shows the sample mean and the sample standard deviation of the
allocative efficiency for the experiments above over 1000 runs. The designs of
market 1 and 2 imply efficiencies of at least 90 % which occurs when all available
units are being traded as it is the case for both types of unconstrained traders.
The intra-marginal and the extra-marginal side of the symmetric market B are of
equal size. Hence the allocative efficiency of the unconstrained traders in it is 0.
Thus, Gode & Sunder [4] could have used truly ZIU agents for their experiments
already. Introducing a timing heuristic does not increase the efficiency value. In
the constrained case however, the introduction of a timing heuristic does lead to
a slightly higher efficiency than it is the case with random timing.
The mean actual profits achieved by ZIC and ZICT traders per period in
market B when placed in competition is shown in Figure 10. ZIC traders are
able to achieve a mean actual profit of above 920 price units while ZICT traders

Transaction price time series ZIUT Transaction price time series ZICT
200 Trades 200 Trades
Period range Period range
Equilibrium Equilibrium
price=80 price=80

150 150
Price

Price

100 100

50 50

0 0
0 200 400 600 800 1000 1200 1400 1600 0 2000 4000 6000 8000 10000 12000 14000 16000 18000
Time Time

Fig. 8. Transaction price time series for Fig. 9. Transaction price time series for
market 1 with 6 buyers and 6 sellers of market 1 with 6 buyers and 6 sellers of
ZIUT traders (random timing). ZICT traders (random timing).
are only able to achieve a mean actual profit of around 800 price units. Market
theory predicts a theoretical profit of 900 units for this market. This shows that
ZIC traders are more competitive due to their heuristic of time. ZICT traders
are not able to achieve high profit values. They are even exploited by ZIC traders
which explains their profit being above the predicted theoretical profit.

3.2 ZIP vs ZIPT

ZIP traders require the following set of parameters. During all experiments R
is uniformly distributed over [1.0, 1.05] for price increases and over [0.95, 1.0] for
price decreases. These values are suggested by Cliff [1]. The values for A are
changed due to a larger scale for the price units. A is uniformly distributed over
[0.0, 5.0] for price increases and over [−5.0, 0.0] for price decreases. The learning
rate β is set to 0.3 and the momentum coefficient γ is set to 0.05 for all traders
as suggested by Preist et al. [6]. At last, buyers start with profit margins that
give a price of 0 as their first offer. Sellers on the other hand start with profit
margins that give the limit price of the market as their first offer.
ZIPT traders start submitting offers when the size of the bid-ask-spreak has
fallen below 5.0 price units. If the bid-ask-spread has not changed for 5 time
units ZIPT traders start testing the market themselves before considering the
size of the bid-ask-spread again. Also, ZIP traders adjust their profit margins
in the direction of the best best open respective offer when there has not been
a trade for 5 time units. This policy was introduced by Das et al. [2] and is
necessary for ZIP traders to be applied in continuous double auctions.
Unlike the set-up in section 3.1 the list of offers persists after the occurrence
of a transaction to provide a more realistic market place.
A first set of experiments is concerned with homogeneous populations of
ZIP and ZIPT traders. Cliff [1] reports that transaction prices of ZIP traders
converge to the equilibrium price during early periods and stays stable during
later periods. The performance of ZIP traders in this set-up should be equally
good. Transaction prices of ZIPT traders should also be able to converge to the
equilibrium price since they employ the same pricing strategy as their original
complement. However, clues for the profit margin adjustments of the traders are
given less frequently which might lead to longer period lengths. In a second set
of experiments ZIP traders are confronted with ZIPT traders. The performance
of either trader is measured by their actual profit earned during any one period.

Results Figure 11 shows the transaction price time series in market A with a
population of ZIP traders. As can be seen the transaction prices scallop around
the equilibrium price of 200 price units during the first four periods. However,
the price series converges towards the equilibrium price where it stays stable for
the last two periods.
Figure 13 shows the transaction price time series in market A with a homo-
geneous population of ZIPT traders. The shape of the graph looks similar to
figure 11. The transaction price time series scallops around the equilibrium price
for the first four periods and stabilizes in period five.
 Transaction price time series ZIP
Actual profits ZIC vs. ZICT
400 Trades
940 ZIC Buyer Period range
ZIC Seller Equilibrium
ZICT Buyer price=200
ZICT Seller 350
920

300
900

250
880
Actual profit

Price
200
860

150
840

100
820

50
800

0
780 0 50 100 150 200 250 300 350 400
0 1 2 3 4 5
Time
Period

Fig. 10. Actual profits earned in market B Fig. 11. Transaction price time series for
by a heterogeneous population of ZIC and market A with 6 buyers and 6 sellers of ZIP
ZICT traders. traders (heuristic timing).

Traders Market 1 Market 2 Market A

x̄ (s) x̄ (s) x̄ (s)
ZIP 99.79 (± 0.24) 99.86 (± 0.19) 99.86 (± 0.30)
ZIPT 99.81 (± 0.22) 99.83 (± 0.22) 98.84 (± 0.31)

Fig. 12. Sample mean and sample standard deviation of the efficiency of ZIP and ZIPT
traders, sample size = 1000.

The mean allocative efficiency of markets with homogeneous populations of

either ZIP or ZIPT traders are around 99% and summarized in figure 12. The
efficiency value for ZIP traders conforms with the findings by Cliff [1].
The second set of experiments was concerned with heterogeneous populations
of ZIP and ZIPT traders. Figure 15 shows the transaction price time series while
Figure 16 shows the sample mean of the actual profits for each period. It can be

Coefficient of convergence ZIP vs. ZIPT

25
ZIP Buyer
Transacion price time series ZIPT ZIP Seller
400 ZIPT Buyer
Trades ZIPT Seller
Period range
Equilibrium 20
350 price=200

300
15
Alpha

250
Price

200 10

150

5
100

50
0
0 1 2 3 4 5
Period
0
0 200 400 600 800 1000 1200 1400 1600 1800
Time

Fig. 14. Course of the sample mean of the

Fig. 13. Transaction price time series for coefficient of convergence for heterogeneous
market A with 6 buyers and 6 sellers of populations of ZIP and ZIPT traders, sam-
ZIPT traders (strategic timing). ple size = 1000.
 Actual profits ZIP vs. ZIPT
6500 ZIP Buyer
ZIP Seller
Transaction price time series ZIP vs. ZIPT 6000 ZIPT Buyer
400 ZIPT Seller
Trades
Period range 5500
Equilibrium
350 price=200
5000

300 4500

Actual profit
4000
250
3500
Price

200 3000

2500
150
2000
100
1500

50 1000
0 1 2 3 4 5
Period
0
0 100 200 300 400 500 600 700 800
Time

Fig. 16. Course of the sample mean of ac-

Fig. 15. Transaction price time series for tual profits for a heterogeneous population
market A with a heterogeneous population of ZIP and ZIPT traders in market A, sam-
of ZIP and ZIPT traders. ple size = 1000.

seen that ZIPT traders trade closer to the equilibrium price and are therefore
less vulnerable to being exploited. It is underlined by Figure 14 which shows the
sample mean of the coefficients of convergence for each period.

4 Conclusions

This article expanded the horizon of timing strategies employed by trading

agents. The implicit heuristic “sooner offers yield higher profits” was made ex-
plicit and has been investigated in two directions. First, the Zero-Intelligence
trading strategy was enhanced with random timing. The efficiencies of these ac-
tual Zero-Intelligence agents does only differ by a very small margin in the case
of constrained pricing and it does not differ at all in the case of unconstrained
pricing in contrast to their original counterparts. However, when put in compe-
tition results of Figure 10 show that using the heuristic yields higher profit than
placing random bids in time. This finding confirms that continuous double auc-
tions are highly efficient mechanisms even with only a small number of the most
random traders. This efficiency can be expected to increase with the number of
traders, but additional experiments are necessary to quantify the scalability and
efficiency gains of larger markets.
Second, the Zero-Intelligence-Plus strategy was enhanced with a more so-
phisticated timing strategy. Again, the results show that the efficiency of the
enhancement does not differ from its original complement. In competition to
heuristic timing, using time as a strategic element prevented the agents from
being exploited by trading far from the equilibrium price, as indicated in Fig-
ure 16.
The outcome of this paper indicate that markets are efficient even with the
most random traders, but individual traders do have an incentive to exploit
timing as an essential part of their strategies. Intelligent timing strategies are
indispensable and it can not be made standard practice to only consider timing
strategies that place offers as soon as their prices are determined. We suggest to
extend this strategic timing to more complex trading strategies such as Gjerstad
& Dickhaut [3], and to discuss the effect explicitly.

References
1. Dave Cliff. Minimal-intelligence agents for bargaining behaviors in market-based
environments. Technical report, School of Cognitive and Computing Sciences, Uni-
versity of Sussex, 1997.
2. Rajarshi Das, James E. Hanson, Jeffrey O. Kephart, and Gerald Tesauro. Agent-
human interactions in the continuous double auction. In Proceedings of the 17th
international joint conference on Artificial intelligence - Volume 2, pages 1169–1176,
San Francisco, CA, USA, 2001. Morgan Kaufmann Publishers Inc.
3. Steven Gjerstad and John Dickhaut. Price formation in double auctions. GAMES
AND ECONOMIC BEHAVIOR, 22:1–29, 1998.
4. Dhananjay K Gode and Shyam Sunder. Allocative efficiency of markets with zero-
intelligence traders: Market as a partial substitute for individual rationality. Journal
of Political Economy, 101(1):119–37, 1993.
5. Simon Parsons, Marek Marcinkiewicz, Jinzhong Niu, and Steve Phelps. Everything
you wanted to know about double auctions but were afraid to (bid or) ask. Technical
report, in preparation.
6. Chris Preist and Maarten van Tol. Adaptive agents in a persistent shout double
auction. In Proceedings of the first international conference on Information and
computation economies, ICE ’98, pages 11–18, New York, NY, USA, 1998. ACM.
7. John Rust, Richard Palmer, and John H. Miller. Behavior of Trading Automata in
a Computerized Double Auction Market. In D. Friedman, J. Geanakoplos, D. Lane,
and J. Rust, editors, The Double Auction Market: Institutions, Theories and Evi-
dence. Addison Wesley, 1992.
8. Vernon Smith. An experimental study of competitive market behavior. Journal of
Political Economy, 70, 1962.