Titelfolie Partition Dependence in Prediction Markets Evidence from the Lab and from the Field Ulrich Sonnemann Muenster Colin Camerer CalTech Craig Fox Thomas Langer UCLA Muenster Outline 1. Motivation & Research Question 2. Experimental/Empirical Evidence • Lab Study: DAX index, (temperature, sports) • Field Experiment (NBA Playoffs, Soccer World Cup) • Field Data (Economic Derivatives) 3. Conclusions [email protected] +49 (0) 251 83-22026 ESA 2007 World Meeting, Rome 1 1. Motivation & Research Question Prediction Markets and Partition Dependence Prediction Markets Markets where participants trade in contracts whose payoff depends on unknown future events. Iowa Electronic Markets TradeSports HP printer sales (Chen, Plott 02) Siemens deadline (Ortner 98) Partition Dependence Individual probability judgments depend on specific categorical „partition“, i.e. there is a bias towards the ignorance prior (bias toward 1/n). Different partitions of the event space can lead to different probability judgments for the same event e.g. Fox, Clemen Mgt Sci 05: Survey: Wolfers, Zitzewitz JEP 04 „.. market forecasts are typically fairly accurate...“ Subjective Probability Assessment in Decision Analysis: Partition Dependence and Bias Toward the Ignorance Prior ESA 2007 World Meeting, Rome 2 1. Motivation & Research Question Research Question Are prices in prediction markets (and therefrom inferred probabilities) influenced by the partition of the event space? Do markets eliminate (or at least mitigate) the partition dependence bias which is observed in individual judgment (and under what circumstances)? ESA 2007 World Meeting, Rome 3 2. Experimental/Empirical Evidence Lab Study: DAX index, (temperature, sports) Study 1 Laboratory Study: DAX index, (temperature, sports) ESA 2007 World Meeting, Rome 4 2. Experimental/Empirical Evidence Lab Study: DAX index, (temperature, sports) Construction of the assets (for DAX index stimulus) Underlying event: A claim will pay 100 cts if the DAX closes within the respective interval, otherwise the claim will pay nothing „DAX close two weeks in the future“ DAX 7497 7328 Asset 1 DAX 7327.99 Asset 2 7328 DAX 7496.99 7647 Asset 3 7497 DAX ESA 2007 World Meeting, Rome 5 2. Experimental/Empirical Evidence Lab Study: DAX index, (temperature, sports) DAX 7497 7328 7647 Treatment 1 Asset 1 DAX 7327.99 Asset 2 7328 DAX 7496.99 Asset 3&4 7497 DAX Treatment 2 Asset 1&2 DAX 7496.99 Asset 3 7497 DAX 7646.99 ESA 2007 World Meeting, Rome Asset 4 7647 DAX 6 2. Experimental/Empirical Evidence Lab Study: Main Result Bias in Equilibrium Market Prices (DAX index stimulus) intervals Treatment 1: N = 12 mean equil. prices 1 2 3&4 0.152 0.561 0.336 0.713 p<0.0001 Wilcoxon N = 12 mean equil. prices Treatment 2: 12 34 0.245 p<0.0001 Wilcoxon 0.581 0.289 0.424 0.404 0.177 1&2 3 4 intervals ESA 2007 World Meeting, Rome 7 2. Experimental/Empirical Evidence Field Experiment (NBA Playoffs, Soccer World Cup) Study 2 Field Experiment: NBA Playoffs, Soccer World Cup ESA 2007 World Meeting, Rome 8 2. Experimental/Empirical Evidence Field Experiment (NBA Playoffs, Soccer World Cup) Engaging events! NBA Basketball Playoffs 2006 FIFA Soccer World Cup 2006 Large scale study: N = 317 (Germany) + 139 (UCLA) = 456 Max. 20 traders per single market Long span of continuous open markets (24/7) Playoffs April 22 through June 22, 2006 (~9 weeks) World Cup May 24 through July 9, 2006 (~6½ weeks) Main Finding: Evidence for Partition Dependence! ESA 2007 World Meeting, Rome 9 2. Experimental/Empirical Evidence Field Data (Economic Derivatives) Study 3 Field Data: Economic Derivatives ESA 2007 World Meeting, Rome 10 2. Experimental/Empirical Evidence Field Data (Economic Derivatives) Economic Derivatives (by Goldman Sachs and Deutsche Bank) Field Prediction Markets for macroeconomic derivatives such like growth in non-farm payrolls, retail sales, etc. Highly professional traders / High stakes ($) Allows derivation of probability density function from prices of digital options: Gürkaynak, Wolfers 06 ESA 2007 World Meeting, Rome 11 2. Experimental/Empirical Evidence Field Data (Economic Derivatives) We assume the observed probability distribution to be actually a mixture between • • the true probability and an ignorance prior distribution 0.1 0.1 0.1 0.09 0.08 0.09 0.09 0.07 0.08 0.07 0.08 0.07 0.06 0.06 0.06 0.05 0.04 0.05 0.05 F(x) B(x) Observed distribution =* 0.04 0.03 0.02 0.01 1. 35 1. 15 0. 95 0. 75 0. 55 0. 35 0. 15 -0 .0 5 0 -0 .3 60 69 02 1. 35 1. 15 0. 95 0. 75 0. 55 0. 35 0 0. 15 1. 35 1. 15 0. 95 0. 75 0. 55 0. 35 0. 15 -0 .0 5 -0 .3 60 69 02 0 + (1- ) * 0.01 -0 .0 5 0.02 0.01 0.03 0.02 -0 .3 60 69 02 =* 0.03 0.04 I(x) True (i.e. unbiased) + (1- ) * distribution Ignorance prior (1/n) distribution Are bias-corrected B(x) more accurate than observed F(x)? Yes: =.6 mean abs error .673 ( =1 .680) Slight correction w/ =.99 improves forecasts (mean forecast error) 53%, 56%, 70%, 54% in four markets (overall n=153, p<.01) ESA 2007 World Meeting, Rome 12 3. Conclusions Conclusions Partition-dependence exists in prediction markets Evidence from: Lab: Around 25% in judgment and equilibrium prices Field experiments: Virtual arbitrage 1-6% Field markets: 40% weight on (1/n)? Market forces seem not to be able to mitigate the systematic distortion in individual judgment! Thank you for your attention! ESA 2007 World Meeting, Rome 13
© Copyright 2026 Paperzz