Humans, Automatons and Markets

1
Humans, Automatons and Markets

• Shyam Sunder
• Yale University
• Center for Analytical Research in Technology
• Tepper School of Business, Carnegie Mellon
  University October 10, 2007

2
An Overview

• Until recent decades, economics had been
  constrained by
• Mathematical solvability of models
• Controlling or observing strategies that generate
  empirical observations
• Computers allow us to simulate complex markets
  and specified strategies
• Confluence of psychology, computer science, and
  economics now allows us to
• Engineer alternative models of trader behavior
  and map them to market outcomes
• Find out if, and to what degree, the limited
  human cognition stands in conflict with market
  equilibria derived from agent optimization
• Design markets with specified outcome properties
  through study of statistical interactions among
  traders
• Examine investor attempts to gain competitive
  advantage through algorithmic trading

3
Designing Trading Automatons

• Psychologists have long questioned the validity
  of economic theories predicated on optimizing the
  behavior of agents
• Computers allow us to populate markets with
  various kinds of models (including psychological)
  of individual behavior, and to observe their
  aggregate level outcomes
• Simulations allow us to assess the strengths and
  weaknesses of human vs. automaton traders, and to
  engineer hybrid strategies that might combine
  their strengths

4
Human Traders

• Simon proposed and empirically validated a
  coherent theory of intuitive human behavior
• Bounded rationality as a theory of how the mind
  works, and not optimal costly search
• Economics, decision theory, psychology, and
  sociology inform us about trading motivations,
  opportunities, information, cognition, and
  learning
• Trading complexity limits individually optimal
  decisions through hot intuitions
• Reading, instruction, and experience can help
  modify beliefs about opportunity sets, behavior
  of others, interrelationships among variables,
  and response and outcome functions in a market
  setting
• Human learning by experience tends to land on a
  plateau
• Can cyborgs help?

5
Can Cyborgs Help?

• What are the comparative advantages of human and
  automaton traders?
• The boundary between computer and human traders
  has become less clear
• To what degree can social, cognitive and brain
  sciences inform the engineering of automatons and
  their work with humans
• Parallels from other fields birds flight,
  internal combustion engine, car, electricity,
  chess
• There appear to be only a few useful parallels
• Combining human advantages (abstraction, pattern
  recognition, hypothesizing, robustness,
  versatility and imagination) with automaton
  advantages (fast in simple steps, large memory,
  and repetition without getting tired, bored,
  discouraged or frustrated

6
Cognitive Limitations of Humans

• How relevant is the study of these limitations to
  the design of trading automatons?
• May help design traders who take advantage of
  their human counterparties
• However, building such limitations into
  automatons would defeat the very purpose of
  building better traders
• Difference between doll-making and engineering
• Building automatons that imitate human
  limitations may have scientific value, but their
  engineering value is unclear

7
Demand for Trading Automata

• Traders in every market strive to gain
  competitive advantage through every available
  means (Reuters, trans-Atlantic cable)
• ICT has enabled detailed instructions to
  implement complex, state-contingent, even
  learning strategies, and communicate them over
  long distances for rapid execution at remote
  servers
• Four environments for trading automata
• When optimal behavior of trader is specified by
  theory as a simple function of information
  available to trader faithful execution without
  errors and elegant variations, e.g., Vickrey
  auction
• Optimal strategy is known but is computationally
  demanding, e.g., arbitrage trading to dredge the
  pennies
• Optimal strategy is unknown calling for execution
  of heuristics, progressive revision, and learning
  from experience, e.g., DA
• Mutual dependence of expectations and strategies
  with questionable existence and uniqueness of
  equilibria most difficult environment for
  automata

8
Does the Knowledge of Equilibria Help Design
Automata?

• Auction theory literature focuses on identifying
  equilibrium trading strategies and their market
  level outcomes (Nash)
• When Nash does not exist, there is no obvious
  candidate strategy
• The designer might make a guess about others
  strategies and design a good response
• Or an automaton may be designed to find the best
  response to the designers beliefs about others
  strategies
• Or the designer may have automata form beliefs
  (how?)

9
Does Nash Help the Designer?

• The existence of Nash hardly simplifies the
  problem of designing automata
• Existence insufficient to convince people that
  others will use them
• What should a trader do if he does not believe
  the others will use Nash?
• What should the automata do in off-equilibrium
  paths?
• Is there a way for one to use the knowledge of
  equilibrium to make money?
• In Santa Fe DA Tournament, BGAN performed 10 s.d.
  under the winner
• DA with truth telling attains equilibrium but
  loses out in a heterogeneous market to non-truth
  tellers

10
Number of Traders

• Theory and experiments establish a rapid
  asymptotic approach to theoretical equilibria as
  N increases, 95 with a mere 5-6 traders in DA
  and some other auctions
• Greater the competition ? closer to equilibrium ?
  smaller absolute gain from better strategies
• Traders equilibrium share of surplus is a rent
  which is likely to be bid away, leaving only the
  amount earned above the equilibrium level for the
  trader

11
Is Speed an Advantage?

• Speed of computation, decision entry, and
  awareness of market events are pluses
• Speed expands the opportunity set of the trader
• Can does not imply should act faster. Long
  (strategic?) pauses in continuous human trading
  (not cognitive)
• Is pausing a good strategy for automatons? We do
  not know. Wilsons (1987) Waiting Game Dutch
  Auction is the only model of this type
• Strategic use of timing requires forming
  expectations of what others might do and when
• Oft-observed flurry of activity before closing in
  both human as well as automaton markets
• In the Santa Fe tournament, more than 50 percent
  of efficiency losses arose from untraded units
  among intra-marginal traders
• Does the strategic use of timing confer any
  systematic advantage? How do we learn to have our
  automatons make such use of time?

12
Cognition and Automata

• As calculating machines, humans act by intuition
  and stripped of their learned algorithms, do not
  perform well and commit systematic errors, not
  all of which are attenuated by experience
• Automatons should exploit such weaknesses in
  others when possible, but not be subject to them
• Difference between markets with reservation
  values inherent in traders vs. market dependent
  values (e.g., stock markets)

13
Learning Strategies

• Level zero the data, initial opportunity sets,
  rules of the market, and mapping from trader
  actions and market events to payoff functions to
  start with
• Tempting to include optimal decision rules, but
  the parameters needed to apply or condition such
  rules on are rarely available
• In some markets (one-shot sealed bid auction)
  learning does not proceed beyond this basic level
  and automatons enjoy an inherent advantage

14
Level 2

• Forming expectations about parameters,
  opportunity sets, and payoff relevant
  consequences of ones own actions and of market
  events
• Humans are naturally equipped to do form and
  adjust such expectations instantaneously without
  apparent effort
• Ill-defined problems rarely stop people from
  offering answers, even wrong answers they learn
  from experience and go on to devise better
  answers
• Fluidity of the human brain is an advantage at
  this level better than what machines have been
  able to do so far
• Building Bayesian adjustments in automata still
  requires endowing them with priors and
  likelihoods appropriate for the environment

15
Higher Levels

• Detecting changes in trading environment
• Whether the change is endogenous (learning and
  behavior) or exogenous
• Humans rapidly form, test, and reject many
  hypotheses
• Building automata with this level of learning is
  a challenge
• Versatility of humans beyond unstructured tasks
• How will automata do in expectation formation,
  and in competing against their own clones?

16
Markets Prone to Indeterminacy

• In private value markets, prices are determinate
  and the consequences of ones actions are known,
  albeit with noise
• In security markets with short term investors,
  values depend on beliefs we impute to unknown
  others, and their beliefs ? price indeterminacy
  and bubbles (Keynes beauty contest, Hirota and
  Sunder 2006)
• Even if the trader knows the fundamental value,
  he cannot benefit from trading on that value
  unless the market prices reflect that value
  before his investment horizon
• Such markets present the most difficult
  challenges for designers of automata

17
Three Simple Designs

• A Automaton ignores all but its private
  information and the fundamental value based on
  this private information buys below and sells
  above this value
• B Automaton assumes that the next transaction
  price will be equal to the most recent price
• C Automaton uses all past data to search over a
  set of forecast functions for price prevailing at
  the investment horizon and trades relative to
  this price
• No automaton can beat its own clones
• Singleton A against many Bs will not beat B
• Whether A and B will do well in a market
  dominated by Cs depends on the set of forecast
  functions used

18
Is Stationarity a Problem?

• It is possible that genetic algorithms or neural
  networks may come up with occasional winners
  against some alternatives
• What happens in non-stationary environments?
• Neural networks need training and data and a
  stationarity assumption at some level
• Will they dominate human traders in nonstationary
  environments?
• Perhaps computer scientists and mathematicians
  already know the answer

19
Design of Trading Automata

• Trade off between the speed and depth of decision
  making
• In a fast moving DA, advantages of deep
  calculations are erased by obsolescence
• Relative, not absolute speed, counts, generating
  profits for early adopters of fast computers
  against humans and slower machines
• Depth of analysis is a decision of the trader
  subject to trading environment automaton should
  be able to conduct its own Turing test (whether
  it is trading against other machines, and assess
  their level of sophistication)

20
What about Market Psychology

• Market psychology or animal spirits formalized in
  economics as higher order beliefs
• It has been difficult to build such abilities in
  automatons (and we are also unclear about how
  humans form higher order beliefs)
• Little theory, evidence, or laws to govern higher
  order belief formation
• Will humans do consistently better or worse than
  Data (of Star Trek)?

21
Markets for People and for Machines

• Humans minimize chances of failure by gradual
  adaptation of their systems
• AURORA of CBT visually reproduces the trading
  pits still rejected by traders
• Such systems were designed for human traders
  assisted by computers for input, output, storage,
  record keeping, communication, and rule
  enforcement, not for automatons
• In a market designed for machines, speed is a
  pre-requisite, not a choice
• Absolute competitive advantage of speed will
  diminish over time, but the relative advantage
  will remain

22
To Summarize This Part

• Productive use of automata in scientific research
• Allow us to fix behavior and explore properties
  of their environmentsuseful ceteris paribus
  approach (difficult for automata to modify
  themselves, difficult for humans to stand still)
• Automata used to supplement humans with speed,
  memory, and computation (arbitrage)
• But dreams are built not on science or
  labor-saving but on the sci-fi versions of
  self-learning automatons that can humiliate the
  masters of the universe on the Wall Street
• Whether this can happen depends on which side of
  the Chinese Room debate you are on
• I do not know enough AI to give an answer

23
Economics Suggests

• Either there will be one winner which will drive
  out all others and thus close the market, or
  there will be no stationary equilibrium among
  strategies
• If competing automatons coexist, only about one
  half of them will perform above average, just as
  naïve traders and expert fund managers do
• In deeper markets, net returns to investors are
  about the same whether they use their own naïve
  random strategies, or pay experts to manage their
  money
• Any extra returns earned by experts are captured
  by them
• Any profits earned by smart automatons, too, will
  end up in the pockets of their designers
• Will having smart traders doing all the trading
  change allocative efficiency?

24
Importance of Being Intelligent?

• Computer simulations reveal the robustness of
  certain market outcomes, and sensitivity of
  others, to trader intelligence
• These simulations, and the analyses that follow,
  help address critical questions of why some
  markets, populated by limited cognition human
  traders, approximate the predictions based on
  optimization (while others exhibit systematic
  deviations from such predictions)
• Are social institutions built (or have they
  evolved) to minimize the importance of our
  intelligence for their efficiency?
• Buy stock in a company thats so good that an
  idiot can run it, because sooner or later one
  will. Peter Lynch

25
Designing Market Institutions

• Institutions are defined by their rules
• Designing a market consists of specifying its
  rules so that it yields outcomes with known
  characteristics under a range of trader behaviors
• Computer simulations help us understand how
  market rules determine the statistical properties
  of interactions among traders

26
Summary

• We can engineer trading algorithms that embody
  their models and conjectures and map them to
  market outcomes
• Bridging the chasm between psychology (individual
  behavior) and economics (aggregate outcomes)
• Designing market rules with specified outcome
  properties by study of statistical interactions
  among traders
• Prospects and consequences of investors to build
  algorithms to try to gain competitive advantage
  over human and other algorithmic traders
• Mostly open questions, few answers yet

27
Thank You!

• shyam.sunder_at_yale.edu
• www.som.yale.edu/faculty/sunder