Information Technology Reference
In-Depth Information
2.2
Auction Dynamics
The market functioned according to an adapted continuous double auction mechanism.
The standard continuous double auction allows buyers and sellers to submit bids to the
rest of the market for consideration at any time. First, in order to simplify the implemen-
tation of a continuous double auction on a network, we adopted the system presented
by Cliff and Bruten [4]. In their simulations, the auction mechanism acts in discrete
time and has no order topic. Each time period, one active agent (one who is still able to
trade) is selected at random to make an offer or a bid. The other agents in the market
are then polled in random order for responses to the shout. If the response and the shout
cross then a trade is executed at the first shouted price, if not the next agent is polled.
If no trader accepts the shout then the shout is removed. Second, we limit an agent's
ability to trade such that they are only able to make offers to, or accept bids from, their
network neighbours. Each market was simulated for a fixed number of time steps.
2.3
Trading Agents
Here, the ZIP trading algorithm is used to govern agent behaviour. ZIP, or Zero Intel-
ligence Plus, agents were created by Cliff and Bruten [4] in response to work by Gode
and Sunder [3], who created the “Zero Intelligence” trading algorithm in some of the
first agent-based market simulations. The Zero Intelligence algorithm was designed to
be the simplest possible algorithm that would allow trade to occur in a market. Two
types of Zero Intelligence trader were introduced. The first, unconstrained traders (ZI-
U), choose shout prices at random from a uniform distribution across the whole range
of possible prices permitted, disregarding any limit prices. It was found that markets
populated by these traders exhibited none of the normal properties associated with mar-
kets, such as convergence to the equilibrium price. The second type of zero intelligence
traders (ZI-C) were
constrained
in the range of prices that could be shouted. Shout
prices were again drawn at random from a uniform distribution. However, this distri-
bution was now constrained by an agent's limit price. In the case of sellers, shouts
were constrained to be greater than the limit price, while in the case of buyers, shouts
had to be less than the specified limit price. Importantly, markets populated by traders
using this algorithm were shown to behave analogously to real markets in that they
converged to the theoretical equilibrium price [3]. This was interpreted as indicating
that the market mechanism itself was the most significant factor in market behaviour,
and that the design of the trading algorithm was not as important. Cliff and Bruten [4],
however, showed this to be incorrect, demonstrating that the convergence observed dur-
ing each trading period was an artifact of the supply and demand schedules used by
Gode and Sunder. They demonstrated that, for a certain type of supply and demand
schedule that was close to symmetric, the probability distribution of likely ZI-C bids
and offers would result in convergence to the mean price. They then performed sim-
ulations to verify these results with a broader range of supply and demand schedules.
For non-symmetric schedules, markets populated by ZI-C traders failed to converge, or
converged to a non-market-equilibrium value.
The ZIP agent differs from the ZI-C agent in that it learns from the market. Each ZIP
trader has a profit margin associated with its limit price. In the case of buyers, the profit
