1. Research Computational Aspects of Prediction Markets David M. Pennock, Yahoo! Research Yiling Chen, Lance Fortnow, Joe Kilian, Evdokia Nikolova, Rahul Sami, Michael Wellman

2. Research Mech Design for Prediction Q: Will there be a bird flu outbreak in the US in 2007? A: Uncertain. Evidence distributed: health experts, nurses, public Goal: Obtain a forecast as good as omniscient center with access to all evidence from all sources

3. Research Mech Design for Prediction expert possible states of the world nurse citizen omniscient forecaster

4. Research A Prediction Market Take a random variable, e.g. Turn it into a financial instrument payoff = realized value of variable $1 if$0 if I am entitled to: Bird Flu Outbreak US 2007? (Y/N) Bird Flu US ’07

5. Research http://intrade.com Screen capture 2007/04/19

6. Research Mech Design for Prediction Standard Properties Efficiency Inidiv. rationality Budget balance Revenue Comp. complexity Equilibrium General, Nash,... PM Properties #1: Info aggregation Expressiveness Liquidity Bounded budget Indiv. rationality Comp. complexity Equilibrium Rational expectations Competes with: experts, scoring rules, opinion pools, ML/stats, polls, Delphi

7. Research Mech Design for Prediction Financial MarketsPrediction Markets PrimarySocial welfare (trade) Hedging risk Information aggregation SecondaryInformation aggregationSocial welfare (trade) Hedging risk

8. Research Outline Examples, research overview Some computational aspects of PMs Combinatorics Betting on permutations Betting on Boolean expressions Automated market makers Hanson’s market scoring rule Dynamic parimutuel market

9. http://tradesports.comhttp://intrade.com

10. http://www.biz.uiowa.edu/iem IPOIPOIPOIPO http://www.wsex.com/

11. Play money; Real predictions http://www.hsx.com/

12. http://us.newsfutures.com/ Cancer cured by 2010 Machine Go champion by 2020 http://www.ideosphere.com

13. Research Yahoo!/O’Reilly Tech Buzz Game http://buzz.research.yahoo.com/

14. Research Example: IEM 1992 [Source: Berg, DARPA Workshop, 2002]

15. Research Example: IEM [Source: Berg, DARPA Workshop, 2002]

16. Research Example: IEM [Source: Berg, DARPA Workshop, 2002]

17. Research Does money matter? Play vs real, head to head Experiment 2003 NFL Season ProbabilitySports.com Online football forecasting competition Contestants assess probabilities for each game Quadratic scoring rule ~2,000 “experts”, plus: NewsFutures (play $) Tradesports (real $) Used “last trade” prices Results: Play money and real money performed similarly 6 th and 8 th respectively Markets beat most of the ~2,000 contestants Average of experts came 39 th (caveat) Electronic Markets, Emile Servan- Schreiber, Justin Wolfers, David Pennock and Brian Galebach

18. Research

19. Research Does money matter? Play vs real, head to head Statistically: TS ~ NF NF >> Avg TS > Avg

20. Does it work? Yes... Evidence from real markets, laboratory experiments, and theory indicate that markets are good at gathering information from many sources and combining it appropriately; e.g.: –Markets like the Iowa Electronic Market predict election outcomes better than polls [Forsythe 1992, 1999][Oliven 1995][Rietz 1998][Berg 2001][Pennock 2002] –Futures and options markets rapidly incorporate information, providing accurate forecasts of their underlying commodities/securities [Sherrick 1996][Jackwerth 1996][Figlewski 1979][Roll 1984][Hayek 1945] –Sports betting markets provide accurate forecasts of game outcomes [Gandar 1998][Thaler 1988][Debnath EC’03][Schmidt 2002]

21. Does it work? Yes... E.g. (cont’d): –Laboratory experiments confirm information aggregation [Plott 1982;1988;1997][Forsythe 1990][Chen, EC-2001] –And field tests [Plott 2002] –Theoretical underpinnings: “rational expectations” [Grossman 1981][Lucas 1972] –Procedural explanation: agents learn from prices [Hanson 1998][Mckelvey 1986][Mckelvey 1990][Nielsen 1990] –Proposals to use information markets to help science [Hanson 1995], policymakers, decision makers [Hanson 1999], government [Hanson 2002], military [DARPA FutureMAP, PAM] –Even market games work! [Servan-Schreiber 2004][Pennock 2001]

22. Research Predicting Permutations Predict the ordering of a set of statistics Horse race finishing times Daily stock price changes NFL Football quarterback passing yards Any ordinal prediction Chen, Fortnow, Nikolova, Pennock, EC’07

23. Research Market Combinatorics Permutations A > B > C.1 A > C > B.2 B > A > C.1 B > C > A.3 C > A > B.1 C > B > A.2

24. Research Market Combinatorics Permutations D > A > B > C.01 D > A > C > B.02 D > B > A > C.01 A > D > B > C.01 A > D > C > B.02 B > D > A > C.05 A > B > D > C.01 A > C > D > B.2 B > A > D > C.01 A > B > C > D.01 A > C > B > D.02 B > A > C > D.01 D > B > C > A.05 D > C > A > B.1 D > C > B > A.2 B > D > C > A.03 C > D > A > B.1 C > D > B > A.02 B > C > D > A.03 C > A > D > B.01 C > B > D > A.02 B > C > D > A.03 C > A > D > B.01 C > B > D > A.02

25. Research Bidding Languages Traders want to bet on properties of orderings, not explicitly on orderings: more natural, more feasible A will win ; A will “show” A will finish in [4-7] ; {A,C,E} will finish in top 10 A will beat B ; {A,D} will both beat {B,C} Buy 6 units of “$1 if A>B” at price $0.4 Supported to a limited extent at racetrack today, but each in different betting pools Want centralized auctioneer to improve liquidity & information aggregation

26. Research Auctioneer Problem Auctioneer’s goal: Accept orders with non-negative worst-case loss (auctioneer never loses money) The Matching Problem Formulated as LP Generalization: Market Maker Problem: Accept orders with bounded worst-case loss (auctioneer never loses more than b dollars)

27. Research Example A three-way match Buy 1 of “$1 if A>B” for 0.7 Buy 1 of “$1 if B>C” for 0.7 Buy 1 of “$1 if C>A” for 0.7 A B C

28. Research Pair Betting All bets are of the form “A will beat B” Cycle with sum of prices > k-1 ==> Match (Find best cycle: Polytime) Match =/=> Cycle with sum of prices > k-1 Theorem: The Matching Problem for Pair Betting is NP-hard (reduce from min feedback arc set)

29. Research Subset Betting All bets are of the form “A will finish in positions 3-7”, or “A will finish in positions 1,3, or 10”, or “A, D, or F will finish in position 2” Theorem: The Matching Problem for Subset Betting is polytime (LP + maximum matching separation oracle)

30. Research Market Combinatorics Boolean Betting on complete conjunctions is both unnatural and infeasible $1 if A1&A2&…&An I am entitled to: $1 if A1&A2&…&An I am entitled to: $1 if A1&A2&…&An I am entitled to: $1 if A1&A2&…&An I am entitled to: $1 if A1&A2&…&An I am entitled to: $1 if A1&A2&…&An I am entitled to: $1 if A1&A2&…&An I am entitled to: $1 if A1&A2&…&An I am entitled to:

31. Research Market Combinatorics Boolean A bidding language: write your own security For example Offer to buy/sell q units of it at price p Let everyone else do the same Auctioneer must decide who trades with whom at what price… How? (next) More concise/expressive; more natural $1 if Boolean_fn | Boolean_fn I am entitled to: $1 if A1 | A2 I am entitled to: $1 if (A1&A7)||A13 | (A2||A5)&A9 I am entitled to: $1 if A1&A7 I am entitled to:

32. Research The Matching Problem There are many possible matching rules for the auctioneer A natural one: maximize trade subject to no-risk constraint Example: buy 1 of for $0.40 sell 1 of for $0.10 sell 1 of for $0.20 No matter what happens, auctioneer cannot lose money $1 if A1 $1 if A1&A2 trader gets $$ in state: A1A2 A1A2 A1A2 A1A2 0.60 0.60 -0.40 -0.40 -0.90 0.10 0.10 0.10 0.20 -0.80 0.20 0.20 -0.10 -0.10

33. Research Market Combinatorics Boolean

34. Research Complexity Results Divisible orders: will accept any q*  q Indivisible: will accept all or nothing Natural algorithms divisible: linear programming indivisible: integer programming; logical reduction? # eventsdivisibleindivisible O(log n)polynomialNP-complete O(n) co-NP-complete  2 p complete reduction from SAT reduction from X3C reduction from T  BF Fortnow; Kilian; Pennock; Wellman LP

35. Research Automated Market Makers A market maker (a.k.a. bookmaker) is a firm or person who is almost always willing to accept both buy and sell orders at some prices Why an institutional market maker? Liquidity! Without market makers, the more expressive the betting mechanism is the less liquid the market is (few exact matches) Illiquidity discourages trading: Chicken and egg Subsidizes information gathering and aggregation: Circumvents no-trade theorems Market makers, unlike auctioneers, bear risk. Thus, we desire mechanisms that can bound the loss of market makers Market scoring rules [Hanson 2002, 2003, 2006] Dynamic pari-mutuel market [Pennock 2004] [Thanks: Yiling Chen]

36. Research Automated Market Makers n disjoint and exhaustive outcomes Market maker maintain vector Q of outstanding shares Market maker maintains a cost function C(Q) recording total amount spent by traders To buy ΔQ shares trader pays C(Q+ ΔQ) – C(Q) to the market maker; Negative “payment” = receive money Instantaneous price functions are At the beginning of the market, the market maker sets the initial Q 0, hence subsidizes the market with C(Q 0 ). At the end of the market, C(Q f ) is the total money collected in the market. It is the maximum amount that the MM will pay out. [Thanks: Yiling Chen]

37. Research Hanson’s Market Maker I Logarithmic Market Scoring Rule n mutually exclusive outcomes Shares pay $1 if and only if outcome occurs Cost Function Price Function [Thanks: Yiling Chen]

38. Research Hanson’s Market Maker II Quadratic Market Scoring Rule We can also choose different cost and price functions Cost Function Price Function [Thanks: Yiling Chen]

39. Research Log Market Scoring Rule Market maker’s loss is bounded by b * ln(n) Higher b  more risk, more “liquidity” Level of liquidity (b) never changes as wagers are made Could charge transaction fee, put back into b (Todd Proebsting) Much more to MSR: sequential shared scoring rule, combinatorial MM “for free”,... see Hanson 2002, 2003, 2006

40. Research Computational Issues Straightforward approach requires exponential space for prices, holdings, portfolios Could represent probabilities using a Bayes net or other compact representation; changes must keep distribution in the same representational class Could use multiple overlapping patrons, each with bounded loss. Limited arbitrage could be obtained by smart traders exploiting inconsistencies between patrons A B C F E D H G [Source: Hanson, 2002]

41. Research 1 1 1 1 1 1 1 1 1 1 1 1 Pari-Mutuel Market Basic idea

42. Research.49.4.3 0.97.96.94.91.87.78.59.82 Dynamic Parimutuel Market C(1,2)=2.2 C(2,2)=2.8 C(2,3)=3.6 C(2,4)=4.5 C(2,5)=5.4 C(2,6)=6.3 C(2,7)=7.3 C(2,8)=8.2 C(3,8)=8.5 C(4,8)=8.9 C(5,8)=9.4

43. Research Share-ratio price function One can view DPM as a market maker Cost Function: Price Function: Properties No arbitrage price i /price j = q i /q j price i < $1 payoff if right = C(Q final )/q o > $1

44. Research Open Questions Combinatorial Betting Usual hunt: Are there natural, useful, expressive bidding languages (for permutations, Boolean, other) that admit polynomial time matching? Are there good heuristic matching algorithms (think WalkSAT for matching); logical reduction? How can we divide the surplus? What is the complexity of incremental matching?

45. Research Open Questions Automated Market Makers For every bidding language with polytime matching, does there exist a polytime MSR market maker? The automated MM algorithms are online algorithms: Are there other online MM algorithms that trade more for same loss bound?

46. Research Yahoo!,O’Reilly launched Buzz Game 3/05 @ETech Research testbed for investigating prediction markets Buy “stock” in hundreds of technologies Earn dividends based on actual search “buzz” API interface Exchange mechanism is dynamic parimutuel market Cross btw stock market and horse race betting http://buzz.research.yahoo.com

47. Research Yahoo!/O’Reilly Tech Buzz Game http://buzz.research.yahoo.com/

48. Research Technology forecasts iPod phoneWhat’s next? Google Calendar? Another Apple unveiling 10/12; iPod Video? search buzz price 9/8-9/18: searches for iPod phone soar; early buyers profit 8/29: Apple invites press to “secret” unveiling 8/28: buzz gamers begin bidding up iPod phone 9/7: Apple announces Rokr 9am 10/5

49. Analysis

50. Research Tech Buzz Game

51. Research An info market model: Computational properties From a computational perspective, we are interested in: What can a market compute? How fast? (time complexity) i.e., What mechanisms or protocols lead to faster convergence to the rational expectations equilibrium? Using how many securities? (expressivity and representational compactness) i.e., What market structures require a minimum of securities yet still aggregate information quickly and accurately?

52. Research Market computation [Feigenbaum EC-2003] General formulation Set up the market to compute some function f(x 1,x 2,…,x n ) of the information x i available to each market participant (e.g., we want the market to compute future interest rates given other economic variables) Represent f(x) as a circuit Questions How do we set up a market to compute f? How quickly can the market compute f? AND XOR OR x1x1 x2x2 x3x3 x4x4 f(x 1,x 2,x 3,x 4 )= (x 1  x 2 )  (x 3  x 4 )

53. Research Market model Each participant has some bit of information x i There is a security F that pays off $1 if and only if f(x)=1 at some future date, and $0 otherwise. Trading occurs in synchronous rounds In each round, participants bid their true expectation Clearing price is determined using a simplified Shapley-Shubik trading model, yielding mean bid Questions we ask/answer: Does the clearing price converge to a stable value? How fast does it converge (in how many rounds)? Does the stable price of F reveal the true value of f?

54. Research Theorems For any prior distribution on x, if f(x) takes the form of a weighted threshold function (i.e., f(x) = 1 iff  i w i x i > 1 for some weights w i ), then the market price will ultimately converge to the true value of f(x) in at most n rounds If f(x) cannot be expressed as a weighted threshold function (i.e., f(x) is not linearly separable), then there is some prior on x for which the price of F is stuck at $0.5 indefinitely, and does not reveal the true value of f(x)

55. Research Interpretation of theory: 1 security supports computation of threshold fn only More complex functions must utilize more securities: # of securities required = threshold circuit size of f In the example, with only a single security on f, the market may not converge Example and interpretation AND XOR OR x1x1 x2x2 x3x3 x4x4 f(x 1,x 2,x 3,x 4 ) $1 if (x 1  x 2 )  (x 3  x 4 ) $1 if x 3  x 4 With 2 additional securities it will converge in 4 rounds $1 if x 4

56. Research Extensions, future work Dynamic information revelation and changes Overcoming false information Obtaining incentive compatibility Modeling agent strategies Modeling overlapping information sources Characterizing in terms of work/round Bayesian network representation of prior Dealing with limited-precision prices

57. Research Open questions What is the relationship between our model and perceptron (neural network) learning? Perceptrons exactly compute threshold functions Could envision a system to learn smallest set of threshold functions to approximate desired function f, thereby minimizing the number of securities required Can alternate market protocols lead to faster convergence? Can subsidies speed convergence? What can other types of securities (e.g., nonbinary securities) compute?

58. Research Does it work? Yes... Evidence from real markets, laboratory experiments, and theory indicate that markets are good at gathering information from many sources and combining it appropriately; e.g.: Markets like the Iowa Electronic Market predict election outcomes better than polls [Forsythe 1992, 1999][Oliven 1995][Rietz 1998][Berg 2001][Pennock 2002] Futures and options markets rapidly incorporate information, providing accurate forecasts of their underlying commodities/securities [Sherrick 1996][Jackwerth 1996][Figlewski 1979][Roll 1984][Hayek 1945] Sports betting markets provide accurate forecasts of game outcomes [Gandar 1998][Thaler 1988][Debnath EC’03][Schmidt 2002]

59. Research Does it work? Yes... E.g. (cont’d): Laboratory experiments confirm information aggregation [Plott 1982;1988;1997][Forsythe 1990][Chen, EC-2001] And field tests [Plott 2002] Theoretical underpinnings: “rational expectations” [Grossman 1981][Lucas 1972] Procedural explanation: agents learn from prices [Hanson 1998][Mckelvey 1986][Mckelvey 1990][Nielsen 1990] Proposals to use information markets to help science [Hanson 1995], policymakers, decision makers [Hanson 1999], government [Hanson 2002], military [DARPA FutureMAP, PAM] Even market games work! [Servan-Schreiber 2004][Pennock 2001]

60. Research Catalysts Markets have long history of predictive accuracy: why catching on now as tool? No press is bad press: Policy Analysis Market (“terror futures”) Surowiecki's “Wisdom of Crowds” Companies: Google, Microsoft, Yahoo!; CrowdIQ, HSX, InklingMarkets, NewsFutures Press: BusinessWeek, CBS News, Economist, NYTimes, Time, WSJ,... http://us.newsfutures.com/home/articles.html