home
***
CD-ROM
|
disk
|
FTP
|
other
***
search
/
Trading on the Edge
/
Trading On The Edge - CD-ROM Toolkit (Wayzata Technology)(2031)(1994).bin
/
pc
/
shared
/
azte_pro
/
robots.txt
< prev
Wrap
Text File
|
1993-06-11
|
34KB
|
756 lines
-------------------------------------------------------------------------
ANDERSON
David Anderson
University of Minnesota
This strategy attempts to buy at prices better than the average
sale price in a period. If the average of the current bid and
the current offer is above the average sale price my program
bids aggresively (in my case bidding aggressively is bidding
2*cbid/3+coffer/3.) I also bid aggressively if all other buyers
have sold one more unit than me. If neither of these conditions
hold I bid nonaggressively (in my case this means bidding
10*cbid/11+coffer/11.) Since this is what I expect to happen in
most cases my stategy amounts to attempting to colude with the
other buyers to bring prices down. I only bid aggresively when
a favorable counteroffer has been made or if others are doing all
the trading.
Finally, if I'm running out of time or if I haven't bought in a
long time I bid coffer. My stategy for selling is symetric.
CLASSIFICATION: simple, nonadaptive, nonstochastic, nonpredictive,
nonoptimizing
Program rank in 1990 SFI Double Auction Tournament: 8th out of 29
-------------------------------------------------------------------------
ANON1
Strategy: delay as long as you dare before moving towards the opponent's
price. This works GREAT if all your teammates do it also. If they don't,
it's still nice for them, but not so nice for me. How noble.
In order not to miss trades by running overtime, I define my waiting
period as a fraction of the remaining ticks.
CLASSIFICATION: simple nonadaptive nonstochastic nonpredictive nonoptimizing
Rank in 1990 SFI DA tournament: 9th out of 29
-------------------------------------------------------------------------
ANON2
This program does the following:
- keeps an average transaction price
- goes quickly to somewhat below the average
- goes slowly otherwise
- speeds up toward the end of the period
- preserves the average price across periods of the same round
CLASSIFICATION: simple nonadaptive nonstochastic nonpredictive nonoptimizing
Rank in 1990 SFI DA tournament: 4th out of 29
-------------------------------------------------------------------------
BOLCER
Author(s) name: Gregory Alan Bolcer
Phone: (714) 856-4047
Institution: University of California at Irvine
Mailing address: Department of Information and Computer Science
University of California, Irvine
Irvine, CA 92717
e-mail address: gbolcer@ics.uci.edu
This program attempts to predict specific trading prices which can
then be used to evaluate bid/offer values by utilizing concepts from
an acceleration method in economics which assumes that as prices near
equilibrium the velocity with which they approach this point decreases,
and a method of encoding feature vectors into a scalar in such a way that
is meaningful, (incompletely) similar to the method used for
sensor-fusion in robotics.
CLASSIFICATION: simple nonadaptive nonstochastic nonpredictive nonoptimizing
Rank in 1990 SFI DA tournament: 29th out of 29.
-------------------------------------------------------------------------
BRETON
Marc Breton
University of Minnesota
This program attempts to allow for flexible initial bidding /offering
(at time 1, period 1) and a fairly aggresive strategy for achieving the low
offer/high bid. In contrast, the strategy for closing in on a deal is less
aggresive in that it requires reciprocation on the part of by the agents in
the other half of the market. It responds to percentage changes as well as
interval changes in the bid/offer from the other side of the market. The
acceptance criteria for offers is simply enough to guarentee a reasonable
return (i.e., ratio of bid to offer) and probably will be very susceptible
to other competitors "stealing" the deal. C'est la vie!
Rank in 1990 SFI DA tournament: 7th out of 29
-------------------------------------------------------------------------
BROMILEY
Philip Bromiley, Mgmt. Dept., U of Minnesota
271 - 19th Av S #835
Minneapolis, MN 55455
ASP sets an aspiration level and adjusts it on the basis of
last time's activity. It does a simple forecast of next time's
offers and bids and then decides to enter the market if trading
is expected to reach the aspiration level.
By initially being quiet, ASP hopes to avoid counter-productive
competition with traders on its own side.
ASP takes care of a couple of odd contingencies: time running
out and no messages from its own side, but leaves other, rather
obvious contingencies uncovered (e.g. monopoly for or against,
or only one unit to trade).
ASP requires little computational ability and little memory. It
is an outgrowth of work here to try to simulate some of the
experimental market (human) results with minimal "expense" in
terms of computational ability and memory.
Rank in 1990 SFI DA tournament: 18th out of 29
-------------------------------------------------------------------------
BURCHARD
Paul Burchard
Institute for Advanced Study, ECP-8
Princeton NJ 08540
This program uses an adaptive curve fitter based on "cell division"
to predict the trading behavior of its fellow players. (For more
information on the curve fitter, see "cells.h".) Four curve fitters
are used in the current version of the program. They predict: [1] the
highest bid to be made by others on the next time step (assuming no trade
takes place this time), [2] the same (but assuming a trade does take
place this time), [3] the lowest offer to be made by others on the next
time step (assuming no trade takes place this time), the same (but
assuming a trade does take place this time). Given its predictions,
it bids just high enough or offers just low enough that it should win
on the next timestep.
In addition to obviously acceptable deals, during a buy-sell step it will
also accept deals (not objectionable in any obvious way) which come at
a better than average price. At the very end of a period it will accept
anything with positive profit.
Rank in 1990 SFI DA tournament: 10th out of 29
-------------------------------------------------------------------------
EXP
Roland Michelitsch, Shawn LaMaster, Mark VanBoening
Economic Science Laboratory, University of Arizona, Tucson, Arizona
Exp computes an expected equilibrium price based not only on the
information we gain by gametype, but also taking into consideration
our own token values. Expected eq is updated if we are a seller if
a trade occurs above the expected eq and if we are buyer if a trade
occurs below the expected eq. At the end of every period (in the
last quarter) Exp goes below this expected eq as far as a markup
above/below the current token value. This markup depends on how much
time is left.
Rank in 1990 SFI DA tournament: 27th out of 29
-------------------------------------------------------------------------
ESL Free
Designed by Dr. Corinne Bronfman, Finance Department, University of Arizona
Programmed by Shawn LaMaster, Economic Science Lab, University of Arizona
Economic Science Laboratory, University of Arizona, Tucson, Arizona
Free uses the information in the price process to update the estimate
of the equilibrium price (free rides on other peoples updating). As the
number of trades increase, the weights on the current price information
increase.
Rank in 1990 SFI DA tournament: 23rd out of 29.
-------------------------------------------------------------------------
GAMER
Shawn LaMaster, Mark Van Boening, Kevin McCabe
Economic Science Laboratory, University of Arizona, Tucson, Arizona
ESL Gamer simply bids/offers a 5% markup of its value to win
the bid/offer step, and then accepts any bid/offer that provides non
negative profits. Because of the discreet steps involved, the bid/offer
and buy/sell steps, and the manner in which ties for both sides accepting
the winning bid or offer are broken, we believe this simple strategy will
work. At worst, ESL Gamer would receive 5% of its value, if a winner of
the bid/offer step. At best, ESL Gamer would receive the difference between
its value and the offer that won the bid/offer step. This strategy would
not work in an auction market where any trader can accept the standing
bid/offer.
Rank in 1990 SFI DA tournament: 24th out of 29
-------------------------------------------------------------------------
JACOBSON
Guy Jacobson, Carnegie Mellon University
A simple speed-up strategy that maintains an estimate of the equilibrium
price for this round based on a weighted average of previous trade prices,
and moves to that estimate at rate propotional to its confidence in the
estimate's accuracy. (As more trades are made, this confidence
increases.) Once there, it is pretty stubborn about getting its price,
and only capitulates when time is running out relative to the rate at
which the gap between cbid and coffer is closing in the current period,
and when the profit is a relatively large fraction of the profit + gap.
Rank in 1990 SFI DA tournament: 17th out of 29
-------------------------------------------------------------------------
KAPLAN
Todd R. Kaplan
University of Minnesota
Mpls, Mn. 55406
This routine involves several strategies.
1. it stays out of the bidding war hoping that by doing so the market
becomes more favorable to buyers
2. It waits until the bids and offers are within a certain range, then
it jumps in and buys before someone else does. It also makes sure
that before jumping in that the transaction is somewhat profitable
3. If its been a while since the last transaction, it decides to buy
preventing the period from ending with incomplete transactions.
4. When it is stuck with a one on one, (No transactions can take place
without it participating) it just uses (3) to wait until the seller
saturates and offers its lowest possible price.
5. It tries to improve in future periods and not get 'taken' twice.
The main idea is to make the market favorable to buyers and get its fair
share of the profits. This is a semi-collusive strategy.. If the other
buyers/sellers keep there prices favorable, then they all can win big.
In practice, it needs the majority of one side of the market to be non
aggressive. Problems: the range it decides to jump in might be too big
or too small so it might be to ambitious and buy too expensive, or wait
too long and get shut out of the market.
It also relies heavily on the other buyers being smart if they are
non collusive. An aggressive side member can also hurt it. It also assumes
that its opponents are computers and won't recognize its obvious patterns.
This strategy takes no account of past history or performance. It also
doesn't allow for tokens being outside the minprice-maxprice range.
Rank in 1990 SFI DA tournament: first place out of 29.
-------------------------------------------------------------------------
KINDRED
Darrell Kindred
Duke University
Basis of strategy: Similar to the simple one in the skeleton, but
we adjust our bids/offers and propensity to sell based on previous
sales, potential profits, and time left in the period.
Rank in 1990 SFI DA tournament: 22nd out of 29
-------------------------------------------------------------------------
LEE
Bernard P. Lee
University of British Columbia
The strategy employed is a passive "observe your competition"
strategy. An assumption is made that the competition has very
similar token values to what was given to the program. As each
competitor (buyer or seller) makes a bid or offer, their approximate
`margin' is calculated by dividing the amount of the bid or offer by
the approximated value of their token.
For instance, of buyer #2 has made an offer of 500, and had
already bought one in the same period, this program approximates that
buyer #2's cost will be the same as the value of the second token given.
All the calculated margins of the most recent offers for each
competitor with each token is saved, and the program's next bid and
offer for the nth token will be based on the average margin on the nth
token of all other competitors, instead of the current bid or offer.
This strategy isolates the fact that different competitors with
different token values will be bidding or offering at similar amounts,
each with different levels of profit. The strategy also allows for
setting of a fixed goal that whatever the bid or offer will be, it
must me equal to or greater that the average profit gained by all
other competitors.
To prevent the strategy from being stubborn, data on how much the
program is `ahead of' or `behind' the competition is also collected.
If the program is `ahead of' the competition, the `requirement' on
the profit will be greater. If the program is `behind', the `requirement'
on the profit will drop in order to meet the furious competition.
Buying and selling are also very simple strategies. A ratio between
the profits if the current bit becomes the price and if the current offer
becomes the price is made. Accept the bid/offer if the ratio falls to
a certain limit. If the ratio is too high, do not accept, and hope that
the other side accepts.
If the held current bid or offer is bettered, let the bid/offer
strategy beat it again. If the other side refuse to accept for a fixed
number of times, assume the other side has reached its lower limit on
profit and accept the offer/bid. If it is about the last time in the
period, accept it because minimal profit is better than none.
Both the ratio and the maxinum number of times to reject is adjusted
by the position relative to the competition, i.e. how much ahead of or
behind the competition.
No assumption on consistant behaviour on other players is made,
and parameters are cleared and recalculated each period.
Rank in 1990 SFI DA tournament: 13th out of 29
-------------------------------------------------------------------------
LEINWEBER
David Leinweber
MJT Advisers
800 West 6th Street
LA, CA 90066
This is a simple strategy based on the fact that it is
better to take a smaller profit at the end of a period
than to be stuck with token that "expire" worthless.
The basic strategy is the same as the skeleton provided by
SFI, modified as follows:
The maximum number of trades possible is calculated at
the start of each period. When more than half of this number
have actually occurred, the program becomes less "greedy" in
its buy/sell decisions.
Rank in 1990 SFI DA tournament: 14th out of 29
-------------------------------------------------------------------------
LIN
Prof. Kuan-Pin Lin, Department of Economics
Portland State University, Portland, OR 97207
My strategy in playing DA game is very simple. First the "bid" is a weighted
average of the standing bid (if any) and a reservation (target) price drawn
from a normal distribution with mean and standard error of historical prices
in the market. The weight used in computing this bid depends on number of
sellers and buyers in the market, the number of total and current time units
in each period, and the number of total and current trades. Next the "buy"
decision is based on the profitability rule that current offer must be lower
than the mean market price plus one standard error. The seller's "offer" and
"sell" strategies are constructed similarly.
Rank in 1990 SFI DA tournament: 26th out of 29
-------------------------------------------------------------------------
MAX
Vernon Smith, Shawn LaMaster, Mark Van Boening, Roland Michelitsch,
Robert Dorsey, Steve Rassenti, Kevin McCabe, Corinne Bronfman, Eric Lu
Economic Science Laboratory, University of Arizona, Tucson, Arizona
Max is an expected utility maximizing buyer/seller that uses
the cumulative distribution function of offers/bids to calculate
the probability that a seller/buyer will accept our bid/offer.
Until the number of observations of offers+accepted bids/
bids+accepted offers are equal to (2*ntimes*sellers)/(2*ntimes*buyers),
or 20, whichever is smaller, Max uses an exponential decay based on the
expected equilibrium price computed from the gametype to determine a
bid/offer.
Risk in this model is held at 1 for a risk nuetral player.
Max uses a linear decay rule for minimum acceptable profit when
deciding to accept a offer/bid once he wins the bid/offer step. This
decay starts with the maximum amount of profit and then decays to zero.
Rank in 1990 SFI DA tournament: 19th out of 29
-------------------------------------------------------------------------
MAX-R
Vernon Smith, Shawn LaMaster, Mark Van Boening, Roland Michelitsch
Robert Dorsey, Steve Rassenti, Kevin McCabe, Corinne Bronfman, Eric Lu
Economic Science Laboratory, University of Arizona, Tucson, Arizona
Max is an expected utility maximizing buyer/seller that uses
the cumulative distribution function of offers/bids to calculate
the probability that a seller/buyer will accept our bid/offer.
Until the number of observations of offers+accepted bids/
bids+accepted offers are equal to (2*ntimes*sellers)/(2*ntimes*buyers),
or 20, whichever is smaller, Max uses an exponential decay based on the
expected equilibrium price computed from the gametype to determine a
bid/offer.
Risk in this model is held at 1.5 for a risk loving player.
Max uses a linear decay rule for minimum acceptable profit when
deciding to accept a offer/bid once he wins the bid/offer step. This
decay starts with the maximum amount of profit and then decays to zero.
Rank in 1990 SFI DA tournament: 20th out of 29
-------------------------------------------------------------------------
GA-NN-1
Richard Dallaway Inman Harvey
richardd@cogs.susx.ac.uk inmanh@cogs.susx.ac.uk
School of Cognitive & Computing Sciences
University of Sussex
Falmer BRIGHTON UK
Genetic algorithm has been used to evolve recurrent neural network.
Program is described in more detail in Rust, Miller, Palmer
"Characterizing Effective Trading Strategies: Insights from
a Computerized Doulbe Auction Tournament" Journal of Economic
Dynamics and Control (forthcoming) also available by anonymous
ftp from fido.econ.arizona.edu in azte/da2.ps.
This program reads the result of the best player (garnet.dna) we had
at the time we stopped evolution. Garnet.dna is a specification
of a fixed (i.e., no learning) recurrent neural network.
Disqualified for excessive losses in 1990 SFI DA tournament
Program modified to yield average performance in subsequent
"scientific tournaments" held at SFI.
-------------------------------------------------------------------------
PERRY
Stan Perry
Portland State University
This trader uses a statistical bidding strategy based on the average
price and the price variance. In addition, a measure of expected
efficiency is included and is used to adjust a tuneable parameter
at the end of each trading period. The Statistical Player uses a fairly
conservative non-statistical strategy until the first 3 trades are
completed in the first period.
In what follows the market strategy is described from the buyer's
viewpoint, the sellers viewpoint is generally just the inverse. The
Statistical Player computes the average price and standard deviation
from the beginning of the round, and for the most recent period within
the round. The player weights the statistics from the previous period
and combines this with the round statistics. These simple statistics
are combined with information on the number of players, the seller/buyer
ratio, the apparent efficiency and the time remaining within the round
to develop a bid.
This player opens bidding at about 2 standard deviations below the
average price at the beginning of a period. (The precise amount the
bid is below average price depends on the tuneable parameter which
includes some random noise.) As the period progresses, the player
moves the bid closer to the average price, and at the end of the period
it may bid up to 0.2 standard deviations above the average price.
The buying strategy is very simple, the Statistical Player sets an
acceptance region at the beginning of the period at approximately
1 standard deviation below the average price. As the period evolves,
the player moves the acceptance region closer to the average price.
By the end of the period, the player will accept an offer that is
0.2 standard deviations above the average price.
At the end of each period an evaluation function is called to assess
the efficiency of the player. The possible profit is computed by
comparing the number of trades that were feasible at the average price
with the players token values. The apparent efficiency is calculated
from actual profit and the possible profit at the average price.
If the efficiency is below an acceptable level a tuneable parameter
is adjusted in proportion to the size of the error.
Rank in 1990 SFI DA tournament: 6th out of 29
-------------------------------------------------------------------------
PRICETAKER
John Rust
University of Wisconsin
Madison, WI 53706
Pricetaker emulates "naive price-taking" behavior, treating
market prices as good estimates of CE prices. If PTAKER is a buyer, he will
accept any outstanding offer that yields a profit if time is running out:
i.e. if there are fewer than STPLFT steps remaining, where STPLFT is a
random fraction of the total number of available time steps (see precise
efinition below). If there are more than STPLFT steps remaining, Pricetaker
won't buy until it sees a "good deal". A good deal is defined as an offer
which is no more than a certain percentage MRKUP over a running average of
the transaction price in the current and previous periods. However even if
it detects a good deal, Pricetaker will still wait to buy if offer prices are
coming down and there are more the STPLFT steps remaining in the period.
Pricetaker makes bids according to a simple rule: if there is a current bid
outstanding and this bid is not higher than a certain percentage MRKUP
over a running average of the transaction prices in the current and
previous periods, then Prcietaker overbids the current bid by a random
percentage PREM of the profits earned on the purchase of its token at the
current bid price, or passes if profits at the current bid are negative.
If the current bid is "too high", i.e. more than MRKUP percent over a
running average of transaction prices in the current and previous periods,
then Pricetaker waits for bid prices to come down before trying to bid. If
there is no current bid then Pricetaker overbids a random percentage PREM
of the running average transaction price, provided it is positive. However
if the running average transaction price is zero (as it is in the start of
period 1 in each round), then Pricetaker starts by making an "outrageous
bid", i.e. a random fraction of his lowest redemption value. Such an
outrageous bid guarantees that Pricetaker will earn a hansom profit if
accepted by a "sucker" seller at the beginning of the game, reducing the
risk of being committed to buy at too high a price while Pricetaker is
getting a sense of what kinds of bids and offers the other traders will
be making.
The strategy Pricetaker follows as a seller is simply a mirror image of
the strategy it follows as a buyer.
Program did not participate in 1990 SFI DA tournament, but
ranks as an average player in subsequent "scientific tournaments".
-------------------------------------------------------------------------
RINGUETTE
Marc Ringuette, Carnegie Mellon Univeristy, Comp Sci Dept.
This is a simple-minded strategy which does surprisingly well in local
run-offs. I'll describe it from the buyer's point of view.
It has two phases. For a long time, it just watches, making no
significant bids; but if a buyer and seller converge within a trigger
margin, it jump-bids and makes the transaction. In other words, it waits
for fellow buyers to do the negotiating, then "steals" the transaction.
The surprising thing is that it doesn't check what the selling price is
as long as it's not a loss -- it trusts its fellow buyers to do the
negotiating.
For robustness, it has a second phase for when no one else is bidding.
In that case, I chose to have the program bid as slowly as possible.
It can get away with slow bidding because (a) when phase 2 is invoked,
there is usually no competition from other buyers, and (b) because it
speeds up as the deadline approaches, so as not to lose transactions.
Given this, slower is better, to pull the selling price down.
General comments: I conjecture that, since it is fairly difficult to
write a computer program that recognizes the behavior of other programs,
most programs will be almost oblivious to the behavior of the other
players. Given this, there are two possible tactics:
(1) make a non-oblivious program that beats up on oblivious ones
(2) choose the most robust oblivious strategy you can manage.
I chose the second option, mainly because it's easier.
There's a tricky tradeoff: bid quickly so as to put yourself in the drivers'
seat and make all the trades possible, or bid slowly to pull equilibrium
prices downwards. I try to strike a middle ground, by not bidding at all,
but by snatching transactions so as not to be left out.
Rank in 1990 SFI DA tournament: 2nd out of 29
-------------------------------------------------------------------------
SILVER BUFFALO
Greg Fullerton 4771 Tantra Boulder CO. 80303
Mark Cronshaw CU, Dept of Econ. Box 256, Boulder CO. 80309-0256
BJ Lee CU, Dept of Econ. Box 256, Boulder CO. 80309-0256
Jamie Kruse CU, Dept of Econ. Box 256, Boulder CO. 80309-0256
This program bids based on risk calculated by time and tokens left.
It starts by giving a random increment above the last token or cbid
then when available it uses the past increment and estimated
equilibrium calculated by a regression on past bids and offers.
Rank in 1990 SFI DA tournament: 15th out of 29
-------------------------------------------------------------------------
SKELETON
Richard Palmer, Duke University
John Miller, Carnegie Mellon University
This strategy is the simple example strategy distributed to all potential
entrants to the 1990 SFI double auction tournament. A block diagram
of this strategy appears in the paper "Characterizing Effective Trading
Strategies: Insights from a Computerized Double Auction Tournament"
forthcoming in the Journal of Economic Dynamics and Control 1993 and
also available by anonymous ftp from /pub/azte/papers at
fido.econ.arizona.edu.
The bidding strategy of the Skeleton program (the strategy for offering is
symmetric) depends on whether or not there is a current bid on the table.
If cbid>0, skeleton bids a randomly weighted average of cbid+1 and the
variable MOST defined as the minimum of the current ask and the program's
redemption value minus one (provided the current ask > 0) or the
redemption value minus one otherwise. If there is no current bid (cbid=0)
then skeleton bids a value of MOST less U(first token -last token) where
U is a uniform (0,1) random variable and first token is the
redemption value of Skeleton's most valuable token and last token
is the redemption value of skeleton's least valuable token. When cbid=0
MOST is set to the minimum of the current ask and the redemption value
of the least valuable token minus (provided the current ask > 0),
or just the value of the least valuable token less one otherwise.
The buy/sell decision is based on target value defining a "good price"
which is raised towards Skeleton's redemption value as the time since the
last trade increases. The current offer is only accepted if it is below
this target value. The target value is initially set at 1.3 times the
least valuable token less .3 times the most valuable token. Subsequent
target values are given by a*target + (1-a)*next token value, where
a = 1/(t - lasttime) where t is the current time and lasttime is the
last time any trade occurred. Skeleton's selling strategy is symmetric.
Skeleton did not participate in the 1990 SFI DA tournament, but performed
above average in subsequent "scientific tournaments".
-------------------------------------------------------------------------
SLIDE
Mark Van Boening, Shawn LaMaster
Economic Science Laboratory, University of Arizona, Tucson, Arizona
This program collects bids/offers and computes a prediction of
price (bid or offer) based on a OLS regression of a sliding window of 15
observations. This strategy is basically computing a trend,
estimating the coefficients from the regression and computing a prediction
to bid or offer.
CLASSIFICATION: simple,nonadaptive,stochastic,predictive,optimizing
Rank in 1990 SFI DA tournament: 16th out of 29
-------------------------------------------------------------------------
STAECKER
MARK STAECKER
UNIVERSITY OF WESTERN ONTARIO
LONDON ONTARIO CANADA
Strategy: The basic strategy of this program is to predict what will
occur in the next step, then attempting to maximize my profits by
using this information. The program attempts to predict the next
low offer, the next high bid, and the equilibrium trading price by using
general statistrics on past behavior. Then, using these three
predictions, it decides if a transaction is likely to occur on the
next step. If it is, it attempts to beat the best price, otherwise, it
does not attempt to beat the best price. Further details on Staecker's
strategy can be found in paper "Characterizing Effective Trading STrategies:
Insights from a Computerized Double Auction Tournament" by Rust, Miller
and Palmer (forthcoming Journal of Economic Dynamics and Control, 1993)
and available via anonymous ftp from fido.econ.arizona.edu in
/azte/papers/da2.ps
Classification: complex, adaptive, non-stochastic, predictive, optimizing
Rank in 1990 SFI DA tournament: 2nd out of 29
-------------------------------------------------------------------------
TERMINATOR
Shawn LaMaster, Mark Van Boening, Stephen Rassenti, Robert Dorsey
Original design by Dr. Robert Dorsey, University of Mississippi
(see "Estimation and Simulation of the Double Auction" a discussion
paper for the Economic Science Association meetings, 1989,
available from Robert Dorsey, Dept. of Economics, University
of Mississippi, University MS 38677.)
The Terminator strategy is derived from an empirical analysis of
human behavior in experimental DA markets.
Terminator uses a 3 part linear decay rule for
setting bids and asks. As a starting
point, it uses the minimum seller value/maximum buyer value computed
from the game type. Current parameters have Terminator decay 30 percent
over the first .5 ntimes, 50 percent over the next .3 ntimes, and
20 percent over the last .2 ntimes.
The buy/sell decision is based on a straight linear decay of
a measure of acceptable profit.
The original design uses an exponential decay. We have modified the
design to use a piecewise linear decay.
Rank in 1990 SFI DA tournament: 11th out of 29
-------------------------------------------------------------------------
TRUTHTELLER
John Rust
University of Wisconsin
Madison, WI 53706
This program implements naive truthtelling strategy in which the
program bids a value equal to its current token value and accepts any
current ask for which it can make a profit. The selling strategy is
symmetric.
Classification: simple, nonadaptive, nonstochastic, nonoptimizing,
nonpredictive
Program did not participate in 1990 SFI DA tournament, but did very poorly
in subsequent scientific tournaments.
-------------------------------------------------------------------------
UTGOFF
Paul E. Utgoff
Lederle Graduate Research Center
University of Massachusetts
Amherst, MA 01003
Strategy is based on fitting two quadratics (one for bids, other for
offers) to the three most recent bids/offers, using time as the x-axis.
One can then extrapolate, drawing inferences about whether, when, and where
the two curves might intersect in the future, or (more likely) when and
where they come closest. See comments below for details. This could be
improved quite a bit with some more effort.
Classification: simple, nonadaptive, stochastic, predictive, nonoptimizing
Rank in 1990 SFI DA tournament: 21st out of 29
-------------------------------------------------------------------------
WENDROFF
Harvey Rose and Burton Wendroff
Los Alamos National Laboratory
Los Alamos NM 87545
Bid-Offer: We determine an equilibrium price by averaging over
previous trades. Our bid-offers are constrained to be within
two standard deviations of that price (random in that interval).
Buy-Sell: If offer is too much
higher than the equilibrium we reject it, otherwise we accept.
CLASSIFICATION: simple, adaptive, nonstochastic, predictive, nonoptimizing
Rank in 1990 SFI DA tournament: 25th out of 29
-------------------------------------------------------------------------
ZI
This is a version of the "Zero Intelligence" strategy of Gode
and Sunder (see chapter 7 by Gode and Sunder in "The Double
Auction Market: Institutions, Theories and Evidence" Friedman
and Rust (eds.)
The ZI program simply bids a random amount below its current
redemption value and MINPRICE, and a ZI seller asks a random
amount abouve its token cost and MAXPRICE. A ZI trader will
accept any bid or ask if it can make a positive profit.
CLASSIFICATION: simple, nonadaptive, stochastic, nonpredictive, nonoptimizing
Program did not participate in 1990 SFI DA tournament, but
did poorly in subsequent "scientific tournaments".
-------------------------------------------------------------------------