1 Computation in a Distributed Information Market Joan Feigenbaum (Yale) Lance Fortnow (NEC Labs) David Pennock (Overture) Rahul Sami (Yale)

2 Markets Aggregate Information! Evidence indicates that markets are good at combining information from many sources: –Markets like the Iowa Electronic Market predict election outcomes better than opinion polls [Forsythe et al. ’99]. –Futures and options markets provide accurate forecasts of their underlying commodities/securities [Jackwerth et al. ’96]. –Sports betting markets provide unbiased forecasts of game outcomes [Gandar et al. ’98; Debnath et al. ’03] –Laboratory experiments confirm information aggregation [Plott et al ’88, Plott et al. ’97] –Markets sometimes deployed primarily for information aggregation (e.g., IEM, Hollywood Stock Exchange)

3 Market as a Computation Device x1x1 xnxn initial info. market buy/sell orders price x 1,p 1 x n,p 1 updated information new orders p1p1 equilibrium price p* market equilibrium price p*  aggregate f (x 1, x 2, …, x n ) Questions: What aggregate functions f ( ) can be computed? How many securities must be traded? How fast does the market price converge?

4 Simplified Market Model – Each trader i has a single bit of information x i – Desired aggregate is a Boolean function f (x 1, x 2, …, x n ). Study Boolean functions Trade in a single security F with payoff contingent on f : If f (x 1, x 2, …, x n ) turns out to be 1, F eventually pays off $1; otherwise, F eventually pays off $0. Use multiperiod Shapley-Shubik model of the market – Trading occurs in a sequence of rounds. – In each round, trader i brings a “money supply” b i and a “securities supply” q i to the market. – Clearing price is p =  b i /  q i. Trader behavior: common-prior Bayesians, truthful (non-strategic) bidding. Simplifying assumption: q i = 1 (forced trade)

5 Example: OR function Two traders, who initially know x 1, x 2 respectively. Uniform prior distribution on (0,0), (0,1), (1,0), (1,1). Single security F, based on f(x 1,x 2 ) = x 1 V x 2. F has value $1 if x 1 V x 2 =1 ; value $0 otherwise. x 1 = 1 x 2 = 0 F market clearing price p 1 = 0.75 b 1 = 1 b 2 = 0.5 bids initial info. Equilibrium price reveals the value of f(x 1,x 2 ) in this market. x 1 = 1, x 2 = 0, F market p * = 1 b 1 = 1 b 2 = 1 bids  x2 = 0  x2 = 0 p 1 =0.75  x1 = 1  x1 = 1 equilibrium price updated info.

6 Example: XOR function Two traders, who initially know x 1, x 2 respectively. Uniform prior distribution on (0,0), (0,1), (1,0), (1,1). Single security F, based on f(x 1,x 2 ) = x 1  x 2. F has value $1 if x 1  x 2 =1 ; value $0 otherwise. x 1 = 1 x 2 = 0 F market clearing price p 1 = 0.5b 1 = 0.5 b 2 = 0.5 bids initial info. Clearing price reveals no new information to either agent  equilibrium reached! Price p* = 0.5. X Equilibrium price does not reveal the value of f(x 1,x 2 ) here.

7 Theorem: Computable Functions If f can be expressed as a weighted threshold function f (x 1, x 2, …, x n ) = 1 if  w i x i  1 = 0 if  w i x i < 1, then, for any prior distribution, the market price of F converges to the true value of f (x 1, x 2, …, x n ). e.g., OR function: x 1 V x 2 V … V x n = 1 iff  x i  1 Proof uses McKelvey-Page theorem on common knowledge of aggregates, combined with a counting argument.

8 Converse: Non-computable Functions If f cannot be expressed as a weighted threshold function, then there exist prior distributions for which the price of F does not converge to the true value of f (x 1, x 2, …, x n ). Proof idea: Construct a probability distribution over x such that By Bayes’ law,  Each agent i ’s bid is independent of x i !  i P(x i = 1 | f(x) = 1) = P(x i = 1 | f(x) = 0)  i P(f(x) = 1 | x i = 1) = P(f(x) = 1| x i = 0)

9 Converse Proof (1) f is not a weighted threshold function  Convex hulls of f -1 (0) and f -1 (1) intersect in R n x*x* There is a point x* in intersection of f -1 (0) and f -1 (1).

10 Converse Proof (2) Can find probabilities 1, 2, … and  1,  2, … such that E(x | f(x) = 1) = x*E(x | f(x) = 0) = x* = = 0.5  2 = 0.5  1 = 0.5 Take the mean of the two distributions:  i P(f(x) = 1 | x i = 1) = P(f(x) = 1 | x i = 0) = 0.5 ! 1 /2 = /2 = 0.25  2 /2 = 0.25  1 /2 = 0.25 x*

11 Convergence Time Bounds Upper bound: For any function f, and any prior distribution, the market reaches the equilibrium price p* after at most n rounds of trading. Lower bound: There is a family of weighted threshold functions C n (the “carry-bit” functions) with 2n inputs, and corresponding prior distributions, such that it takes n rounds in the worst case to reach equilibrium. Bounds are tight up to a factor of 2.

12 Lower Bound Proof (1) Function C n : Carry bit of adding two n-bit numbers x1x1 y2y2 x2x2 y1y1 y n-1 ynyn x n-1 xnxn + C1C1 C2C2 C n-1 CnCn Constructing the probability distribution P n − (x 1, y 1 ) : uniformly distributed − (x 2, y 2 ) : distribution conditioned on (x 1, y 1 ), such that C 2 is independent of (x 1, y 1 ) − (x i, y i ) : distribution conditioned on (x 1, …, x i-1, y 1, …, y i-1 ), such that C n is independent of (x 1, …, x i-1, y 1, …, y i-1 )

13 Lower Bound Proof (2) First round: Except x n and y n, no other bit influences its owner’s expectation of C n.  only x n + y n revealed. If x n + y n is revealed to be 1, remaining problem is equivalent to computing C n-1. x1x1 y2y2 x2x2 y1y1 y n-1 ynyn x n-1 xnxn + C1C1 C2C2 C n-1 CnCn

14 Summary Analysis of a simple market model shows limits on what a market can compute and how fast it can compute. → More realistic market model Strategic models, richer information, generic priors, complexity of traders’ computation, etc. → Application to information market design Design securities for certain convergence and faster convergence. Future directions: