Title: Maximizing Information, Optimizing Risk, and Leveraging Forecasts in Securities Markets
1Maximizing Information, Optimizing Risk, and
Leveraging Forecasts in Securities Markets
- David M. PennockNEC Research Institute
Contributors Michael P. Wellman, University
of Michigan Steve Lawrence, NEC Research
Institute C. Lee Giles, Pennsylvania State
University
2Securities marketsstocks, options, futures,
insurance, ..., sports bets, ...
- Allocate risk
- insured transfers risk to insurer, for
- farmer transfers risk to futures speculators
- put option buyer hedges against stock drop
seller assumes risk - sports bet may hedge against other stakes in
outcome
- Aggregate information
- price of insurance? prob of catastrophe
- OJ futures prices yield weather forecasts
- prices of options encode prob dists over stock
movements - market-driven lines are unbiased estimates of
outcomes - IEM political forecasts
3Example IEM
http//www.biz.uiowa.edu/iem
1 if Hillary Clinton wins
1 if Rick Lazio wins
1 if Rudy Giuliani wins
4Talk overview
- (1) Compact securities markets
- (2) Web market games
- Maximizing Information
- Optimizing Risk
- Leveragingforecasts
5Talk overview
- (1) Compact securities markets
- (2) Web market games
- Maximizing Information
- Optimizing Risk
- Leveragingforecasts
6Classical theoryAtomic securities
- Events
- e.g house floods 100ltNEClt110 Bush wins
- Security
- pays off 1 if E1 occurs, nothing otherwise
- 1 iff flood iff 100ltNEClt110 iff Bush wins
- Conditional security
- pays off 1 if E1 E2 occur
- lose price paid if E1 E2 bet called off if
E2 - 1 iff 100ltNEClt110 given that Bush wins
E2
E1
E3
1 if E1
1 if E1E2
7Classical theoryOptimal risk sharing maximal
information
- Requires ?-1 linearly indep securities, where ?
E1?E2?E3? all world states - Benefits
- Pareto optimal allocation of risk
- max info collective probability forecasts
available for all ??? (state prices) - Major practical hurdle
- ?2events intractable!
8Compact securities markets
- Classical complete market
- enough secs to span joint space of events
- Pareto optimal reallocation of risk
- market probabilities defined over full joint
- New compact market
- structured as a Bayesian network
- exponentially fewer securities may suffice
- still Pareto optimal
- market probabilities still defined over full
joint - P W, UAI 2000
9Compact securities markets Intuition
classical
Interestrate ?
1 if YNDI
1 if YNDI
1 if YNDI
1 if YNDI
1 if YNDI
1 if YNDI
1 if YNDI
1 if YNDI
1 if YNDI
1 if YNDI
1 if YNDI
1 if YNDI
Dow avg ?
1 if YNDI
1 if YNDI
1 if YNDI
1 if YNDI
- Y is cond indep of D,I, given N(and everyone
agrees, and more )
Nasdaq avg ?
Yahoo ?
10Extended exampleDrug development process
- Cost of one new drug to market 359M avg, up to
500M - 5 in 5000 drugs pass pre-clinical testing
- 1 in 5000 approved by FDA
- Huge risk
11Extended exampleDrug development process
1 if FDA approves drug X
12Extended exampleDrug development process
1 if drug passes pre-clinical testing
1 if drug passes Phase I testing pre
1 if drug passes Phase II testing PI
1 if drug passes Phase III testing PII
1 if NDA approved PIII
13Extended exampleDrug development process
1 if drug passes pre-clinical testing
1 if drug passes Phase I testing pre
1 if drug passes Phase II testing PI
1 if drug passes Phase III testing PII
1 if NDA approved PIII
14Extended exampleDrug development process
1 if drug passes pre-clinical testing
1 if drug passes Phase I testing pre
1 if drug passes Phase II testing PI
1 if drug passes Phase III testing PII
1 if NDA approved PIII
15Extended exampleDrug development process
1 if lab testing ?
1 if animal testing ?
1 if drug passes pre-clinical testing l, a
1 if drug passes Phase I testing pre
1 if drug passes Phase II testing PI
1 if drug passes Phase III testing PII
1 if NDA approved PIII
16Extended exampleDrug development process
1 if lab testing ?
1 if animal testing ?
1 if drug passes pre-clinical testing l, a
1 if safe in humans pre
1 if drug passes Phase I testing pre, safe
1 if drug passes Phase II testing PI
1 if drug passes Phase III testing PII
1 if NDA approved PIII
17Extended exampleDrug development process
1 if lab testing ?
1 if animal testing ?
1 if drug passes pre-clinical testing l, a
1 if safe in humans pre
1 if drug passes Phase I testing pre, safe
1 if effective in small trials PI
1 if drug passes Phase II testing PI, eff
1 if drug passes Phase III testing PII
1 if NDA approved PIII
18Extended exampleDrug development process
1 if lab testing ?
1 if animal testing ?
1 if drug passes pre-clinical testing l, a
1 if safe in humans pre
1 if drug passes Phase I testing pre, safe
1 if effective in small trials PI
1 if drug passes Phase II testing PI, eff
1 if safe effectivein large trials PII
1 if drug passes Phase III testing PII, se
1 if NDA approved PIII
19Extended exampleDrug development process
1 if lab testing ?
1 if animal testing ?
1 if drug passes pre-clinical testing l, a
1 if safe in humans pre
1 if drug passes Phase I testing pre, safe
1 if effective in small trials PI
1 if drug passes Phase II testing PI, eff
1 if safe effectivein large trials PII
1 if drug passes Phase III testing PII, se
1 if NDA approved PIII, IND
1 if IND approved life
compact 31 vs classical 639
1 if treats life-threatening illness
20Bayesian networks
Bayesiannetwork
DecomposableBayesian network
Pr(E6E3E5) Pr(E6E3E5) Pr(E6E3E5) Pr(E6E3E5)
(moralized, triangulated)
21Structured markets
- Securities markets can be structured analogously
to a BN - One (conditional) security for each CPT entry
- Fully connected BN ? complete market
E1
E2
E4
E5
E3
1 if E6E3E5
E6
Pr(E6E3E5) Pr(E6E3E5) Pr(E6E3E5) Pr(E6E3E5)
1 if E6E3E5
1 if E6E3E5
1 if E6E3E5
22Compact markets Take one
- Natural structure market according to
unanimously agreed-upon independencies
E5
E4
E6
E1
E2
E3
23Example GLU
E1
E2
CIE1,Ø,E2
1 if E1
1000?
1 if E2
-800?
u(y) ln(y b) Generalized Logarithmic Utility
for money
24Risk-Neutral Probability
- Behavior is the product of Pr and u
- maxa ?? Pr(?) ? u(a, ?)
- An observer cannot determine Pr or u
- Agent A with Pr ? f(?) and u/f(?) is equivalent
to agent A with Pr and u - PrRN ? ? Pr u?
- uRN ??u/u?
25Risk-Neutral Probability
- A RN agent would buy if pltEgt lt Pr(E)
- Any agent would buy if pltEgt lt PrRN(E)
- Any agent would sell if pltEgt gt PrRN(E)
- If PriRN(E) ? PrjRN(E) then i and j would desire
to trade - At equilibrium, all agents risk-neutral
probabilities agree, equal prices
1 if E
1 if E
1 if E
26Compact markets take twoRisk-neutral
independency markets
- Instead structure market according to agreed
upon risk-neutral independencies - If, in equil, all RN indep agree with market
structure ? mkt is operationally complete - Pareto optimal allocation of risk
- probabilities for all states (state prices)
inferable - But RN independencies change out of equilibrium
perhaps more arguable basis for agreement on true
independencies
27Example CARA
E1
E2
CIE1,Ø,E2
1 if E1
1000?
1 if E2
-800?
u(y) -e-cy Constant Absolute Risk Aversion
28Example CARA
1 if E2E1
1 if E3E2
1 if E1
u(y) -e-cy Constant Absolute Risk Aversion
29Example CARA
1 if E3E1E2
1 if E2
1 if E1
u(y) -e-cy Constant Absolute Risk Aversion
30Example CARA
1 if E6E3E5
moralized,triangulated graph
1 if E5E3E4
u(y) -e-cy Constant Absolute Risk Aversion
31Compact marketsTake three Independency markets
- CARA Markov indep ? risk-neutral indep
- If all agents have CARA, then market structured
as TRIANGULATE?ni1 MORALIZE(Di) is op complete - Can yield exponential savings
- This example 19 vs. 63
32Inherent limitations
- CARA, non-Markov structure ? not op complete
- GLU, Markov structure ? not op complete
- Impossibility theorems severely restrict
independence preserving functionsGenest
Wagner 86Pennock and Wellman UAI-99 - Market aggregation function subject to same
- Structured markets may yield approximately
optimal allocations in more general settings
33Computational complexityarbitrage
- If redundant securities are inconsistently
priced, arbitrage is possible - In a risk-neutral independency market, correct
prices of redundant securities are computable
given other prices and market structure - Correct pricing ? Bayesian inference
- P-complete
34Compact markets summary
- Under certain conditions, structured markets are
optimal, with exponentially fewer securities than
would otherwise be required - Applications new derivatives markets that allow
agents to hedge more of their risks, w/o combin.
explosion of fin. instruments - drug development
- energy
- financials, etc
35Talk Overview
- Maximizing Information
- Optimizing Risk
- Leveragingforecasts
- (1) Compact securities markets
- (2) Web market games
36Example IEM
http//www.biz.uiowa.edu/iem
1 if Hillary Clinton wins
1 if Rick Lazio wins
1 if Rudy Giuliani wins
37Play-money market gamesHollywood Stock Exchange
http//www.hsx.com/
38Play-money market gamesForesight Exchange
http//www.ideosphere.com/
1 iff Cancer curedby 2010
Canada breaks upby 2020
Machine Go championby 2020
39EfficiencyHollywood Stock Exchange
- Prices of movie stocks and options adhere to
put-call parity, as in real markets - Arbitrage loopholes disappear over time, as in
real markets
40EfficiencyIowa electronic market
- Qualitatively similar to HSX, though
quantitatively more efficient
41Information incorporationHollywood Stock Exchange
- Market probabilities can be evaluated according
to log score compared w experts - Log score - surprise in info-theoretic sense
- Can be considered info rate in bits
42Information incorporationIowa electronic market
- Qualitatively similar to HSX
43Forecast accuracyHSX movie box office predictions
- 0.94 correlation
- Comparable to expert forecasts at Box Office Mojo
44Probabilistic forecastsEntertainment awards
- Bins of similarly-priced options
- Observed frequency? average price
- Analysis similar for horse racing markets
- Error bars 95 confidence intervals assuming
events are indep Bernoulli trials
45Probabilistic forecastsFX science and
technology outcomes
- Prices 30 days before expiration
- Similar results
- 60 days before
- specific date
46Market games summary
- Market games are easier to set up(few
regulations) - Web market games popular and growing
- Q Are incentives strong enough?A Yes (to a
degree) - Even play-money markets can hold valuable
information - http//artificialmarkets.com
- Science 291 987-988, February 9 2001
- KDD 2001
47Applications
- Harvest information from existing games
- Build new games in areas of interest
- Otherwise difficult economic experiments
- Use game data as baseline boost with proprietary
data and algorithms for improved predictions - box office prediction(market chat, news,
experts, web, ...) - weather forecasting(futures, derivatives
experts, satellite images,...)
48Homepage
- http//www.neci.nec.com/homepages/dpennock