Methodology of Exchange Design

1
Methodology of Exchange Design

• John Ledyard and Preston McAfee
• Caltech and Yahoo!

2
Introduction

• There is a large literature on the design of
  selling mechanisms.
• Builds on theory, experiment, and other practical
  tests.
• Has led to a practical methodology for the choice
  of selling methods
• Little has been written on practical exchange
  design.
• Exchange examples
• Flow a group of sellers sells a continuing
  sequence of differentiated goods to a group of
  buyers
• Stock A group of traders is continuously
  rebalancing asset positions

3
Introduction (continued)

• An exchange maps expressed preferences into
  allocations and
• It is a mechanism, which may be iterative and
  reactive
• The process of expressing preferences and the
  mapping into allocations is exchange design
• Different from auction design
• Competition between multiple sellers
• Goals efficiency not revenue, exchange profits
• Auction methods (e.g. ascending prices) may not
  be applicable
• Replacing brokers network externalities
• A single seller can replace own brokers with
  auction.
• New issue who to charge?

4
Introduction (continued)

• Not much literature
• Theory
• VCG
• Myerson-Satterthwaite, Gresik-Satterthwaite,
• General equilibrium
• Experiment and Practice
• One-sided lessons
• We are putting together a bibliography
• Please send references
• Today
• - A rough state of the art commentary

5
Summary(In case I dont get to the end.)

• Expressively Easy
• Design a language for expressing desired trades
  that accommodates important distinctions.
• Understanding what participants actually value is
  critical to a successful design.
• Strategically Simple
• Design trading algorithms so that a
  straight-forward strategy performs reasonably
  well.
• Permitting iterative adjustment of bids can
  simplify strategies but should be binding.
• Functionally Fair
• The exchange design should not be tilted towards
  one type of participant.
• Exchange and traders must keep commitments.

6
A running example RECLAIM

• The Cap maximal pollution levels by year
• The Permits (year, cycle, zone)
• Years (initial) 1994-2010
• Cycle 1 Jan to Dec, Cycle 2 July to June
• Zone 1 inland, Zone 2 coastal
• Declining aggregate amount, total 50.
• Example To cover pollution in Feb 2008 an inland
  firm can use either (2008, 1, 1), (2007, 2,1),
  (2008, 1, 2), or (2007, 2,2)
• A traders problem is to decide whether to buy
  and sell permits or to install abatement
  equipment covering 20 years, one needs to
  negotiate over quantities and prices of 80
  different permits.

7
(No Transcript)
8
(No Transcript)
9
Converging slowly when thin
10
A little faster when much thicker
(N40)
11
(No Transcript)
12
(No Transcript)
13
What a CVM can do to a thin market!
(N12)
14
(No Transcript)
15
(No Transcript)
16
(No Transcript)
17
(No Transcript)
18
A real application - bonds

• Allowed more order types
• Downward sloping demand (diminishing MU)
• Upward sloping demand (quantity discounts)
• ORs of ANDS, ANDs of ORs, etc.
• Size and difficulty of the real problem
• 200,000 variables, 300,000 constraints
• 2,000 bonds
• 50,000 bids (many contingencies allowed 0,1)
• Relaxed algorithm (LP) took 20 minutes
• Needed a solution in 7 minutes
• Could get 85 of best known bound 90 of the time

19
Expressively Easy

• Intentionally design a language for expressing
  trades that accommodates important distinctions.
  
• Understanding what participants actually value is
  critical to a successful design.
• The exchange is replacing brokers who know their
  clients
• Different but similar products can be treated as
  identical to simplify
• Issue exogenous or endogenous?
• Different buyers can have different interfaces
  and bid formats
• Spot buyers vs. impression buyers
• Portfolio balancers vs. single issue speculator

20
Strategically Simple

• Design trading algorithm so that a
  straight-forward strategy performs reasonably
  well
• Dominant strategy is simple but may cost in
  efficiency
• VCG vs McAfee vs Uniform Price Call
• Algorithmic complexity can make sensible
  participation difficult and should be minimized
• Generalized Uniform Price Call Market works very
  well with single-minded traders.
• Open question what if they are not
  single-minded? Conjecture from BFL still ok.
• If prices depend primarily on the marginal
  traders then most have incentive to honestly
  report willingness to pay and accept.
• Pay what you bid is not a particularly good
  approach.
• Prices can be set in a relatively coarse manner
  without significant efficiency loss
• Permitting iteration of bids simplifies but bids
  should be binding
• Information
• Generally want individual bid information not
  available
• Do want aggregate information, like prices,
  available
• With combinatorics, fitting in is important so
  providing individual information can be valuable.
  Endogenous sunshine seems to work here.

21
Functionally Fair

• Exchange neutrality
• Exchange design should not be tilted towards one
  type of participant.
• Example Max stated surplus and not sellers
  surplus
• Commitments
• It is crucial that commitments be filled.
• Traders Deliver promised assets and cash.
• Can enforce with escrow, etc.
• Exchange Stick to the stated rules.
• Bad examples Enron, ACE, ..
• Balanced revenue model
• Modest levels of revenue can be raised with a
  straight percentage charge (and can be
  incorporated in pricing information).
• Large revenue should be collected with value-add
  pricing to cause less damage to efficiency.

22

• END