2.5 The CAT Competition
2.5.1 Basic Structure and Rules
The CAT competition was first introduced as part of TAC in 2007, in order to promote research on efficient and effective competing marketplace design. The underlying platform, JCAT, allows multiple marketplaces to compete against each other and allows marketplaces to be evaluated in a uniform way. Furthermore, this competition has been run successfully in 2007, 2008, 2009 and 2010, and can provide an international benchmark for evaluating the competing marketplace design. Given this, in this thesis, we will evaluate our design of a competing marketplace based on this platform.
As depicted in Figure 2.10, the CAT competition consists of traders, i.e. buyers and sellers, and specialists. Traders can buy or sell goods in one of the available marketplaces which are operated by specialists. In the competition, the traders are provided by the CAT organisers, and specialists (with the market policies and charging strategies) are designed by the competition entrants. Each entrant can only operate a single marketplace in the competition.
FIGURE2.10: Architecture of CAT competition.
Each CAT competition lasts for a number of trading days7, and each day consists of a fixed number of rounds, during which traders submit shouts to the specialist they are registered with.
Each round lasts for a known constant length of time. In the competition, each trader is assigned a private value for the goods it wishes to buy or sell. For the buyer, the private value is its limit price, and for the seller, the private value is its cost price. In the competition, traders’ private values remain constant during a day, but may change from day to day, depending on the config-uration of the game parameters. Traders need to enter marketplaces to make transactions. Here, each trader can only register with one marketplace on a particular day, i.e. only single-home
7A trading day is a virtual day, not an actual tournament day in real time.
trading is considered. Thus bidding strategies in isolated marketplaces can be used. Typically, ZI-C, ZIP, GD and RE strategies (see Section 2.3.2.1 for details) are permitted in this setting.
In addition to the bidding strategies, traders have one of several market selection strategies to decide which marketplace to register with. These market selection strategies implemented in JCAT are as follows:
• random strategy: the trader randomly chooses a marketplace to participate in.
• -greedy exploration strategy: the trader treats the choice of marketplace as an n-armed bandit problem8which it solves using an -greedy exploration strategy where 0 ≤ ≤ 1.
With such a strategy, a trader updates its value function according to its recent profits, and then it chooses the most profitable marketplace with probability 1 − , and randomly chooses one of the remaining marketplaces with probability . The probability can remain constant or can vary over time, depending on the value of a parameter α. If α is 1,
remains constant, while if α takes any value in (0, 1), then will decrease over time.
• softmax exploration strategy: This market selection strategy is similar to the above -greedy exploration strategy except that it uses a softmax exploration strategy in the n-armed bandit algorithm. This means the trader does not treat all marketplaces, except the best marketplace, as the same. If this strategy does not choose the most profitable market-place, it weighs the choice of remaining marketplaces in order to choose more profitable marketplaces. A parameter τ, which is similar to , controls the relative importance of the weights a trader assigns to marketplaces. It can also be fixed or have a variable value determined by α.
Specialists facilitate transactions by matching bids and asks and determining the transaction price. As discussed in Section 1.1, a double auction marketplace consists of the following market policies: timing, matching, pricing and shout accepting policies. Furthermore, the marketplace also needs to specify its charging strategy. An overview of the policies implemented in JCAT is listed in Table 2.4. In the pricing policy, in addition to k-pricing policy and n-pricing policy we introduced in Section 2.3.2.2, JCAT implements two other pricing policies:
• Uniform pricing policy: sets the transaction prices for all matched ask-bid pairs at the same point.
• Side-biased pricing policy: is a kind of k-pricing policy with varying k, where k is set to split the profit in favour of the side where fewer shouts exist.
In addition to the quote-beating accepting and equilibrium accepting policy introduced in Sec-tion 2.3.2.2, JCAT implemented four other accepting policies:
8This is a machine learning problem, similar to the slot machine, where there are n levers, and each lever provides
a reward drawn from a distribution associated with that lever Sutton and Barto (1998). The objective is to maximize the reward sum through iterative pulls. There is no initial knowledge about the levers. In each pull, a trade-off is made between “exploitation” of the lever that has the highest expected reward and “exploration” to get more information about the expected rewards of the other levers.
• Always accepting policy: accepts any shouts submitted by traders.
• Transaction-based accepting policy: records the most recently matched bids and asks, and then uses the lowest matched bid and the highest matched ask as thresholds to restrict the allowed shouts.
• History-based accepting policy: is derived from the GD bidding strategy. GD forms the belief about how likely a given shout is to be matched based on the history of previous shouts. Then, according to this belief, the marketplace accepts shouts that can be matched with the probability no lower than a specified threshold.
• Self-beating accepting policy: accepts all shouts of traders who have not submitted shouts yet, but then only allows traders to modify their standing shouts with more competitive prices.
For the charging strategy, JCAT provides five types of fees which can be set:
• Registration fee: a flat fee charged for participating in a marketplace.
• Information fee: a flat fee charged by the specialist to traders and specialists who require market information from the marketplace. This information consists of the entire history of the accepted shouts and the transactions in that marketplace.
• Shout fee: a flat fee charged for successfully submitted bids or asks.
• Transaction fee: a flat fee charged on each successful transaction.
• Profit fee: a share of the observed profit made by traders, where a trader’s observed profit is calculated as the difference between the shout and transaction price. Note that a trader’s observed profit is different from its actual profit (which is the difference between its private value and transaction price).
Given this, JCAT comes with five pre-set charging strategies. The five pre-set charging strategies in the JCAT platform are as follows:
• Fixed charging: imposes fees at a specified fixed level.
• Bait-and-switch charging: makes a specialist reduce its charges until it captures a certain market share, and then it slowly increases charges to increase profit. The specialist adjusts its charges downward again when the market share drops below a certain level.
• Charge-cutting charging: sets the charges to the lowest charges set by the other market-places on the previous day. This will attract many traders, but may sacrifice profits.
• Learn-or-lure-fast charging: adapts charges towards some desired target according to the scheme used by the ZIP bidding strategy.
Timing Policy (Section 2.3.2.2) Matching Policy (Section 2.3.2.2)
TABLE2.4: Market policies implemented in the JCAT.
Although JCAT has provided some pre-set market policies and charging strategies, specialists can design their own in order to effectively compete with one another in making profit, attracting traders and ensuring shouts submitted in the marketplace result in transactions. To evaluate their performance, the marketplaces are scored based on a combination of three different metrics:
• Profit Share: the profit obtained by the specialist on a particular day as a percentage of the total profits obtained by all specialists on that same day. The profit share score is a number between 0 and 1 for each specialist for each day.
• Market Share: the percentage of traders who have registered with that specialist on a given day. The market share score is between 0 and 1.
• Transaction Success Rate: the percentage of shouts accepted by the marketplace result-ing in transactions. The transaction success rate score is between 0 and 1.
These metrics are weighted equally (i.e. weighted one-third each) and added together to produce a combined score for each specialist for each Assessment Day11. Scores are then summed across all Assessment Days to produce a final game score for each specialist. The specialist with the highest final score is the winner of the competition.