1. A prediction market A random variable, e.g. Turned into a financial instrument payoff = realized value of variable $1 if$0 if I am entitled to: Will US go into recession in 2012? (Y/N) Recession in 2012

2. 2012 September 30 8:02 p.m. ET

3. Between 6.0% and 8.0% chance

4. Kelly betting You have $1000 Market price mp is 0.05 You think probability p is 0.10 Q: How much should you bet?

5. Kelly betting You have $1000 Market price mp is 0.05 You think probability p is 0.10 Q: How much should you bet? Wrong A: $1000 P(bankruptcy)  1

6. Kelly betting You have $1000 Market price mp is 0.05 You think probability p is 0.10 Q: How much should you bet? Wrong A: fixed $1 No compounding magic

7. Kelly betting You have $1000 Market price mp is 0.05 You think probability p is 0.10 Q: How much should you bet? Optimal A: $52.63 or f* fraction of your wealth where f* = (p-mp)/(1-mp) “ edge/odds ”

8. Why? Kelly betting  maximizing log utility – Maximizes compounding growth rate Maximizes geometric mean of wealth – Minimizes expected doubling time Minimizes exp time to reach, say, $1M – Does not maximize expected wealth ( “ All in ” does, but ensures bankruptcy)

9. Compounding You put $10,000 into the bank at 5% rate – Compounded annually: 10,000*1.05 = 10,500 – Monthly: 10,000*(1+.05/12) 12 = $10,511.62 – “Secondly”: 10,000*(1+.05/31.5M) 31.5M =$10,512.71 – Compounded continuously: 10,000*e.05 =$10,512.71 Good way to compare rates: – W t = W 0 e rt r = 1/t*ln(W t /W 0 )

10. Choosing a bank r =

11. Choosing an investment S&P 500, 1988-2011 r=18.45%

12. Choosing an investment S&P 500, 1988-2011 Year Annual Return 198816.61% 198931.69% 1990−3.10% 199130.47% 19927.62% 199310.08% 19941.32% 199537.58% 199622.96% 199733.36% 199828.58% 199921.04% 2000−9.10% 2001−11.89% 2002−22.10% 200328.69% 200410.88% 20054.91% 200615.79% 20075.49% 2008−37.00% 200926.46% 201015.06% 20112.05% Average18.45%

13. Kelly = maximize E[r] Initial wealth W 0  Final wealth W 1 W 1 = W 0 e r Kelly: Don’t maximize E[W 1 ] Instead, maximize E[r] = E[ln(W 1 /W 0 )] = E[ln W F ] - ln W 0

14. The Long Run Ignores variance of r! But in the long (enough) run, variance doesn’t matter W T = W 0 e r1 e r2 e r3... lnW T = lnW 0 +r1+r2+r3+... lnW T ≈ lnW 0 +t*E[r] (law of large #s) W T ≈ W 0 e t*E[r] Year Annual Return 198816.61% 198931.69% 1990−3.10% 199130.47% 19927.62% 199310.08% 19941.32% 199537.58% 199622.96% 199733.36% 199828.58% 199921.04% 2000−9.10% 2001−11.89% 2002−22.10% 200328.69% 200410.88% 20054.91% 200615.79% 20075.49% 2008−37.00% 200926.46% 201015.06% 20112.05% Average18.45%

15. The Long Run Suppose you could invest for 1000s or millions of years if need be Or imagine blackjack – new indep bet every few seconds Strategy with higher E[r] will win out with high probability Year Annual Return 198816.61% 198931.69% 1990−3.10% 199130.47% 19927.62% 199310.08% 19941.32% 199537.58% 199622.96% 199733.36% 199828.58% 199921.04% 2000−9.10% 2001−11.89% 2002−22.10% 200328.69% 200410.88% 20054.91% 200615.79% 20075.49% 2008−37.00% 200926.46% 201015.06% 20112.05% Average18.45%

16. Decision making under uncertainty, an example ABC TV’s “Who Wants to be a Millionaire?”

17. Decision making under uncertainty, an example v 15 = $1,000,000if correct $32,000 if incorrect $500,000 if walk away

18. Decision making under uncertainty, an example if you answer: E[v 15 ] = $1,000,000 *Pr(correct) + $32,000 *Pr(incorrect) if you walk away: $500,000

19. Decision making under uncertainty, an example if you answer: E[v 15 ] = $1,000,000 *0.5 $32,000 *0.5 = $516,000 if you walk away: $500,000 you should answer, right?

20. Decision making under uncertainty, an example Most people won ’ t answer: risk averse U($x) = log($x) if you answer: E[u 15 ] = log($1,000,000) *0.5 +log($32,000) *0.5 = 6/2+4.5/2 = 5.25 if you walk away: log($500,000) = 5.7

21. Decision making under uncertainty, an example Maximizing E[u i ] for i<15 more complicated Q7, L={1,3} walkanswerL1  0.4 X 0.6 log($2k) Q7, L={3} Q8, L={1,3} log($1k) walkanswerL3  0.8 X 0.2 Q8, L={3} log($1k) log($2k) L3

22. Decision making under uncertainty, in general  =set of all possible future states of the world

23. Decision making under uncertainty, in general  are disjoint exhaustive states of the world  i : rain tomorrow & Bush elected & Y! stock up & car not stolen &...  j : rain tomorrow & Bush elected & Y! stock up & car stolen &... 11 22 33  ii ||||

24. Decision making under uncertainty, in general Equivalent, more natural: E i : rain tomorrow E j : Bush elected E k : Y! stock up E l : car stolen |  |=2 n  E1E1 E2E2 EiEi EjEj EnEn

25. Decision making under uncertainty, in general Preferences, utility  i >  j  u(  i ) > u(  j ) Expected utility E[u] =   Pr(  )u(  ) Decisions (actions) can affect Pr(  ) What you should do: choose actions to maximize expected utility Why?: To avoid being a money pump [de Finetti ’ 74], among other reasons...

26. Preference under uncertainty Define a prospect,  = [p,  1 ;  2 ] Given the following axioms of  : orderability:(  1    2 )  (  1    2 )  (  1 ~   2 ) transitivity:(  1    2 )  (  2    3 )  (  1    3 ) continuity:  1    2    3   p.  2 ~ [p,  1 ;  3 ] substitution:  1 ~   2  [p,  1 ;  3 ] ~ [p,  2 ;  3 ] monotonicity:  1    2  p>q  [p,  1 ;  2 ]   [q,   ;  2 ] decomposability: [p,  1 ; [q,  2 ;  3 ]] ~ [q, [p,  1 ;  2 ]; [p,  1 ;  3 ]] Preference can be represented by a real-valued expected utility function such that: u([p,  1 ;  2 ]) = p u(  1 ) + (1–p)u(  2 )

27. Utility functions (  a probability distribution over  ) E[u]:  represents preferences, E[u](  )  E[u](  ) iff    Let  (  ) = au(  ) + b, a>0. – Then E[  ](  ) = E[au+b](  ) = a E[u](  ) + b. – Since they represent the same preferences,  and u are strategically equivalent (  ~ u).

28. Utility of money Outcomes are dollars Risk attitude: – risk neutral: u(x) ~ x – risk averse (typical): u concave (u  (x) < 0 for all x) – risk prone: u convex Risk aversion function: r(x) = – u  (x) / u(x)

29. Risk aversion & hedging E[u]=.01 (4)+.99 (4.3) = 4.2980 Action: buy $10,000 of insurance for $125 E[u]=4.2983 Even better, buy $5974.68 of insurance for $74.68 E[u] = 4.2984  Optimal  1 : car stolen u(  1 ) = log(10,000)  2 : car not stolen u(  2 ) = log(20,000) u(  1 ) = log(19,875)u(  2 ) = log(19,875) u(  1 ) = log(15,900)u(  2 ) = log(19,925)

30. What utility function? Economist: Only you can choose Engineer: Let’s find out

31. Performance of Prediction Markets with Kelly Bettors David Pennock, Microsoft* John Langford, Microsoft* Alina Beygelzimer (IBM) *work conducted at Yahoo!

32. Wisdom of crowds More: http://blog.oddhead.com/2007/01/04/the-wisdom-of-the-probabilitysports-crowd/ http://www.overcomingbias.com/2007/02/how_and_when_to.html When to guess: if you ’ re in the 99.7th percentile

33. Can a prediction market learn? Typical Mkt >> avg expert Static This paper Mkt >? best expert Dynamic/Learning (experts algos) Assumption: Agents optimize growth rate (Kelly)

34. $1 if E 1 $1 if E 2 $1 if E S Market Model p 1, u 1 subjective probability utility for money p i, u i p n, u n Competitive equilibrium prices mp, mp,…  Σq i (mp)=0

35. $1 if E 1 $1 if E 2 $1 if E S Market Model p 1, ln subjective probability utility for money p i, ln p n, ln Competitive equilibrium prices mp, mp,…  Σq i (mp)=0 q i (mp)=w i /mp*(p i -mp)/(1-mp)

36. Wealth-weighted average Assume all agents optimize via Kelly and are “ price takers ” Then mp = ∑ w i p i

37. Fractional Kelly betting Bet f* fraction of your wealth,  [0,1] Why? Ad hoc reasons – Full Kelly is too risky (finite horizon) – I ’ m not confident of p (2nd-order belief) Our (new) reason – fraction Kelly behaviorally equiv to full Kelly with revised belief p+(1- )mp – Has Bayesian justification

38. Wealth-weighted average Assume all agents optimize via Kelly and are “ price takers ” Then mp = ∑ w i p i Fractional Kelly: mp = ∑ i w i p i

39. Wealth dynamics Agents trade in a binary pred market Event outcome is revealed Wealth is redistrib. (losers pay winners) Repeat After each round If event happens: w i  w i p i /pm If event doesn ’ t: w i  w i (1-p i )/(1-pm) Wealth is redistributed like Bayes rule!

40. Wealth dynamics Single security Multiperiod market Agents with log util Fixed beliefs belief time successes:0 trials:0successes:0 trials:1successes:1 trials:2successes:2 trials:3successes:2 trials:4successes:3 trials:5successes:5 trials:10successes:10 trials:20successes:17 trials:30successes:24 trials:40successes:30 trials:50 price frequency wealth Beta(1,1) wealth Beta(1,2) wealth Beta(2,2) wealth Beta(3,2) wealth Beta(3,3) wealth Beta(4,3) wealth Beta(6,6) wealth Beta(11,11) wealth Beta(18,14) wealth Beta(25,17) wealth Beta(31,21)

41. Wealth dynamics

42. Wealth dynamics: Regret Theorem: Mkt log loss < min i agent i log loss - ln w i Applies for all agents, all outcomes, even adversarial Same bound as experts algorithms  No “ price of anarchy ”

43. Learning the Kelly fraction Competitive equilibrium: = 1 Rational expectations equil: = 0 Practical (wisdom of crowds): =  A proposal: Fractional Kelly as an experts algorithm btw yourself and the market – Start with = 0.5 – If right, increase ; If wrong, decrease – Won ’ t do much worse than market (0) Won ’ t do much worse than original prior p

44. Summary and Questions Self-interested Kelly bettors implement Bayes ’ rule, minimize market ’ s log loss If mkt & agents care about log loss: win-win Questions – Can non-Kelly bettors be induced to minimize log loss? – Can Kelly bettors be induced to minimize squared loss, 0/1 loss, etc.? – Can/should we encourage Kelly betting?

45. Recommended Read Great science writing (Pulitzer prize winner) Fun read Accessible yet accurate Well researched Mob meets politics meets Time Warner meets Wall Street meets Vegas meets Bell Labs scientists