25 minute presentation! An Experimental Test of Information & Decision Markets Robin Hanson, Takashi Ishikida and John Ledyard Caltech 2/4/2005 Information Markets • Standard Information Markets seem to work. – Small but complete set of securities – Many informationally small and unbiased traders. • Theory and evidence from experiments and applications are all positive. • But all assume, require, use, ….. – Straight-forward behavior • Price taking, honest revelation, etc. – Complete set of state dependent contracts – Common knowledge of all priors, … • Even then we see “failures to fully aggregate information” – Incomplete Bayesian updating – Incompletely revealing Rational Expectations Equilibrium DIMACS 2 Decision vs Prediction • A policy maker does not just want to know the probability that “terrorist attacks in the US will increase in 2005.” • They want to know the probability that “terrorist attacks in the US will increase in 2005” if “US troops remain in Iraq for 2005.” • With N events (attacks, troop size, …) and S outcomes for each (increase from 10-20%, decrease, …), a complete set of state dependent contracts requires N^S - 1 contracts. S = 2, N = 8 => 255 contracts DIMACS 3 Remember PAM? Goal: Collect accurate predictive information on political and economic stability in the middle east. DIMACS 4 Remember PAM? Every nation*quarter: -Political stability -Military activity -Economic growth -US $ aid -US military activity & all combinations & ……… 8 nations, 5 indices, 4 quarters DIMACS 5 Remember PAM? Every nation*quarter: -Political stability -Military activity -Economic growth -US $ aid -US military activity & all combinations & ……… 8 nations, 5 indices, 4 quarters (N = 180) •Even if we only use up-down questions, completeness requires 2^180 =1.5*(10^54) contracts. DIMACS 6 Using Conditional Contracts • The good news – There may more overall trading. • Traders may know more about and be more willing to trade on the relatively more precise event {terrorism up | troops up} as opposed to the less precise {terrorism up}. • The bad news – There may be thinner trading per security. • Too many markets to pay attention to. – Thinner trading => bad price discovery and incomplete arbitrage => prices do not aggregate information. DIMACS 7 Decision Markets • Markets for Decision Analysis will be thin. – Large and possibly incomplete set of securities – Few informationally large and biased traders • Theory is unlikely to be a good predictor of behavior. • Current applications and experiments may not be applicable to the thinner situation. • How can we know what will actually work? DIMACS 8 Experimental Test Beds • Create an environment that”captures” as much of the problem as possible (the econ wind tunnel) – Three traders, three events with 2 outcome each (8 states) – Common prior with asymmetric information • 10 draws from one urn of 6 equally likely => (1,0,1), (1,1,0),….. • Each trader sees only two entries of each draw: (1,0,x), (1,1,x),… • Run different mechanisms and market designs • Measure performance – How close are final prices to the fully informed posterior? DIMACS 9 Theory Benchmarks - 3 events priors 1 0.9 posteriors uniform 0.8 0.7 CDF 0.6 3 Variable CDFs 0.5 Prior Distribution 0.4 0.3 Individual Posterior 0.2 Uniform Disribution 0.1 0 0 0.05 0.1 0.15 0.2 0.25 KL Distance from Group Posterior DIMACS 0.3 0.35 0.4 10 Individual Scoring Rule priors 1 posteriors 0.9 uniform 0.8 0.7 CDF 0.6 3 Variable CDFs 0.5 Individual (72) 0.4 Prior Distribution 0.3 Individual Posterior 0.2 Uniform Disribution 0.1 0 0.00E+00 5.00E-02 1.00E-01 1.50E-01 2.00E-01 2.50E-01 DIMACS KL Distance from Group Posterior 3.00E-01 3.50E-01 4.00E-01 11 Standard Markets priors 1 0.9 posteriors uniform 0.8 0.7 CDF 0.6 3 Variable CDFs 0.5 Independent DA (24) 0.4 Individual (72) 0.3 Prior Distribution 0.2 Individual Posterior Uniform Disribution 0.1 0 0 0.05 0.1 0.15 0.2 0.25 KL Distance from Group Posterior DIMACS 0.3 0.35 0.4 12 Design Matters • Asking is not enough. • “Let there be markets” is not enough. • Conjecture: An IM will work better in thin situations, if we use (to “thicken” trading) – Conditional contracts and – a Combinatoric (package bid) Call Market • Includes “no arbitrage” pricing but is intermittent • Does not directly address “monopolistic agents” DIMACS 13 Combinatoric Call Market 1 0.9 uniform posteriors 0.8 0.7 CDF 0.6 3 Variable CDFs 0.5 Independent DA (24) priors 0.4 Combined Value (22) Individual (72) 0.3 Prior Distribution 0.2 Individual Posterior Uniform Disribution 0.1 0 0 0.05 0.1 0.15 0.2 0.25 KL Distance from Group Posterior DIMACS 0.3 0.35 0.4 14 Design Matters • We are not yet at complete aggregation. • Conjecture: An IM will work even better in thin situations, if we use (to “thicken” trading) – Conditional contracts and – A Combinatoric Sequentially Shared (Market) Scoring Rule • Is continuous and directly addresses report manipulation • But it involves a subsidy to traders. DIMACS 15 Shared Scoring Rule - w/CC 1 0.9 posteriors 0.8 uniform 0.7 CDF 0.6 3 Variable CDFs 0.5 Independent DA (24) 0.4 Combined Value (22) priors Market Maker (8) 0.3 Individual (72) Prior Distribution 0.2 Individual Posterior Uniform Disribution 0.1 0 0 0.05 0.1 0.15 0.2 0.25 KL Distance from Group Posterior DIMACS 0.3 0.35 0.4 16 Tentative Conclusions • Standard markets and surveys do not work will in thin situations. • Using conditional contracts and assuming some self - selection, either combinatoric call markets or combinatoric sequentially shared scoring rules significantly improve performance over standard markets. DIMACS 17 Open Questions • There are many others we did not test – Pari-mutuel mechanisms • Economides, Lange, and Longitude (some combinatorics) • Pennock - Dynamic Pari-mutel Market • Plott - Auction then Pari-mutuel DIMACS 18 Open Questions • There are many others we did not test – Pari-mutuel mechanisms • Economides, Lange, and Longitude (some combinatorics) • Pennock - Dynamic Pari-mutel Market • Plott - Auction then Pari-mutuel – Others • HP • ………. DIMACS 19 Some Open Questions • There are many other mechanisms we did not test. – Pari-mutuel mechanisms • Economides, Lange, and Longitude (some combinatorics) • Pennock - Dynamic Pari-mutel Market • Plott - Auction then Pari-mutuel – Others • HP • ………. • There are many other environments we did not test in. – Information monopolist – External incentives to manipulate internally – And for PAM -- do these results survive in an ultra-thin world? DIMACS 20 A Force 12 Storm • Create an environment that really stresstests the mechanisms – Six traders, 8 events w/ two outcomes each (256 states) – Common prior with asymmetric information • 10 draws from one urn of 8! equally likely: – (1,0,1,0,1,1,0,0), (1,1,0,0,0,0,0,0),….. • Each trader sees only 4 different entries: – (1,0,x,0,x,1,x,x), (1,0,x,0,x,1,x,x), … DIMACS 21 1 posteriors 0.9 0.8 priors 0.7 8 Variable CDFs CDF 0.6 0.5 Independent DA (18) Combined Value (12) 0.4 Market Maker (17) 0.3 Individual (144) Prior Distribution 0.2 uniform Individual Posterior Uniform Distribution 0.1 0 0 0.2 0.4 0.6 0.8 1 1.2 DIMACS KL Distance from Group Posterior 1.4 1.6 1.8 2 22 Summary of Testing • Thin: 3 traders, 3 events • 7 independent prices from 3 people in 12 minutes • Markets < Individual Scoring Rule < Call < SSSR • SSSR ~ Call given that the group beats the prior – With selection, SSSR and Call Market do best. • Ultra-Thin: 6 traders, 8 events • 255 independent prices from 6 people in 12 min. • Markets ~ Individual Scoring Rule ~ Call < SSSR – SSSR beats the priors at the top (60%) – Nothing else even beats the priors – SSSR is only one with any aggregation DIMACS 23 Final Thoughts • Information Markets are possible and desirable. – Can improve our ability to identify and deal with uncertainty. • Many policy applications will be in thin situations. • Traditional market designs do not work in thin situations. – Information monopolies, adverse decisions, partial updating • The SSSR (w/conditionals) definitely sharpens the signal/noise ratio in thin and ultra-thin markets over traditional markets. • Can we do better? Undoubtedly. DIMACS 24 DIMACS 25