Information Aggregation: Experiments and Industrial Applications Kay-Yut Chen HP Labs Agenda • Lessons from HP Information Markets (Chen and Plott 2002) • Scoring Rules and Identification of Experts (Chen, Fine and Huberman 2004) (Chen and Hogg 2004) • Public Information (Chen, Fine and Huberman 2004) Experimental Economics Program HP Information Markets (Chen and Plott) • • Summary of Events – 12 events, from 1996 to 1999 – 11 events sales related – 8 events had official forecasts Methodology & Procedures – Contingent state asset (i.e. winning ticket pays $1, others $0) – Sales amount (unit/revenue) divided into (8-10) finite intervals – Web-based real time double-auction – 15-20 min phone training for EVERY subject – Market open for one week at restricted time of the day (typically lunch and after hours) – Market size: 10-25 people Experimental Economics Program Event 2 0.3 0.25 IAM Distribution P 0.2 Actual Outcome 0.15 0.1 HP Official Forecast 0.05 IAM Prediction 0 0 100 200 $ Experimental Economics Program 300 Results Abs % Errors of IAM Predictions Last Interval Ignored Event 2 3 4 5 6 7 8 9 Absolute % errors of HP forecasts 13.18% 59.55% 8.64% 32.08% 29.69% 4.10% 0.11% 28.31% T-test P-value Average last 60% trade 4.61% 57.48% 7.84% 30.93% 24.23% 7.33% 2.00% 23.85% Average last 50% trade 4.57% 55.72% 8.15% 31.57% 24.54% 7.02% 2.35% 24.85% Average last 40% trade 4.68% 54.60% 8.52% 31.83% 25.30% 6.71% 1.83% 24.39% Average last 60% trade 5.63% 59.25% 6.45% 29.74% 22.94% 5.35% 1.53% 17.55% 0.079 0.084 0.071 0.034 Random variable x=official error – market error H0: mean of x=0 Alternate: mean of x>0 Experimental Economics Program Last Interval Mass at Lower Bound Average Average last 50% last 40% trade trade 5.68% 5.80% 57.46% 56.32% 6.77% 7.13% 30.33% 30.48% 23.22% 23.93% 4.91% 4.55% 1.39% 1.00% 17.32% 16.54% 0.026 0.022 Business Constraints and Research Issues • Not allowed to “bet” players’ own money -> stakes limited to an average of $50 per person • Time horizon constraints -> 3 months to be useful • Recruit the “right” people • Asset design affects the results (How to set the intervals?) • Thin markets (sum of price ~ $1.11 to $1.31 over the dollar) – Few players – Not enough participation Experimental Economics Program Reporting with Scoring Rule Outcome A B C Reports of Probability Distribution p1 p2 Pays C1+C2*Log(p3) Experimental Economics Program p3 Information Aggregation Function If reports are independent, Bayes Law applies … Ps | I p s1 p s2 ... p s N p s s1 p s2 ... p s N Experimental Economics Program Two Complications • Non-Risk Neutral Behavior • Public Information Experimental Economics Program Dealing with Risks Attitudes: Two-Stage Mechanism Event 1 Stage 1: Information Market Event 2 Event 3 Call Market to Solicit Risk Attitudes Time Event 4 Event 5 Event 6 Event 7 Stage 2: Probability Reporting & Aggregation Individual Report of Probability Distribution Nonlinear Aggregated Function Event 8 Experimental Economics Program Second Stage: Aggregation Function Bayes Law with Behavioral Correction 1 1s 2 2s N Ns N Ns p p ...p Ps | I 1 2 p1s p2s ...p s Normalizing constant for individual risks i=r(V i / i)c Holding value/Risk - measure relative risk of individuals “market” risk ~sum of prices/winning payoff Experimental Economics Program Experiments: Inducing Diverse Information Outcome A B Box of Balls A C C B C C * In actual experiments, there are TEN states Experimental Economics Program Random Draws Provide Info Comparison To All Information Probability Kullback-Leibler = 1.453 0.900 0.9 Omniscient No Info 0.8 0.800 0.7 0.700 Probability 0.6 0.600 0.5 0.500 0.4 0.400 0.3 0.300 0.2 0.200 0.1 0.100 0.0 0.000 Series1 Series2 A 1 B 2 C 3 D 4 5 E 6 States 7 F G 8 Experiment 4, Period 17 No Information 9 H 10 I J Kullback-Leibler Measure • Relative entropy • Always >=0 • =0 if two distributions are identical • Addictive for independent events Experimental Economics Program Comparison To All Information Probability Kullback-Leibler = 1.337 0.900 0.9 Omniscient IA Mechanism 0.8 0.800 Probability 0.7 0.700 0.6 0.600 0.5 0.500 0.4 0.400 0.3 0.300 0.2 0.200 0.1 Series1 Series2 0.100 0.0 0.000 A 11 B 2 2 3 C 34 D 45 E F 65 States 7 6 8 Experiment 4, Period 17 1 Player G H I J 79 8 10 9 10 Comparison To All Information Probability Kullback-Leibler = 1.448 0.900 0.9 Omniscient IA Mechanism 0.8 0.800 Probability 0.7 0.700 0.6 0.600 0.5 0.500 0.4 0.400 0.3 0.300 0.2 0.200 0.1 0.100 0.0 Series1 Series2 A 0.000 1 B 2 C 3 4 D 5 E 6 F States 7 G 8 Experiment 4, Period 17 2 Players Aggregated 9 H 10 I J Comparison To All Information Probability Kullback-Leibler = 1.606 0.900 0.9 Omniscient IA Mechanism 0.8 0.800 Probability 0.7 0.700 0.6 0.600 0.5 0.500 0.4 0.400 0.3 0.300 0.2 0.200 0.1 Series1 Series2 0.100 0.0 A 0.000 1 B 2 C 3 4 D 5 E 6 F States 7 G 8 Experiment 4, Period 17 3 Players Aggregated H 9 10 I J Comparison To All Information Probability Kullback-Leibler = 1.362 0.900 0.9 Omniscient IA Mechanism 0.8 0.800 Probability 0.7 0.700 0.6 0.600 0.5 0.500 0.4 0.400 0.3 0.300 0.2 0.200 0.1 Series1 Series2 0.100 0.0 A 0.000 1 B 2 C 3 4 D 5 E 6 F States 7 G 8 Experiment 4, Period 17 4 Players Aggregated 9 H 10 I J Comparison To All Information Probability Kullback-Leibler = 0.905 0.900 0.9 Omniscient IA Mechanism 0.8 0.800 Probability 0.7 0.700 0.6 0.600 0.5 0.500 0.4 0.400 0.3 0.300 0.2 0.200 0.1 Series1 Series2 0.100 0.0 A 0.000 1 B 2 C 3 4 D 5 E 6 F States 7 G 8 Experiment 4, Period 17 5 Players Aggregated H 9 10 I J Comparison To All Information Probability Kullback-Leibler = 1.042 0.900 0.9 Omniscient IA Mechanism 0.8 0.800 Probability 0.7 0.700 0.6 0.600 0.5 0.500 0.4 0.400 0.3 0.300 0.2 0.200 0.1 0.100 0.0 Series1 Series2 A 0.000 1 B 2 C 3 4 D 5 E 6 F States 7 G 8 Experiment 4, Period 17 6 Players Aggregated 9 H 10 I J Comparison To All Information Probability Kullback-Leibler = 0.550 0.900 0.9 Omniscient IA Mechanism 0.8 0.800 Probability 0.7 0.700 0.6 0.600 0.5 0.500 0.4 0.400 0.3 0.300 0.2 0.200 0.1 Series1 Series2 0.100 0.0 A 0.000 1 B 2 C 3 4 D 5 E 6 F States 7 G 8 Experiment 4, Period 17 7 Players Aggregated H 9 10 I J Comparison To All Information Probability Kullback-Leibler = 0.120 0.900 0.9 Omniscient IA Mechanism 0.8 0.800 Probability 0.7 0.700 0.6 0.600 0.5 0.500 0.4 0.400 0.3 0.300 0.2 0.200 0.1 Series1 Series2 0.100 0.0 A 0.000 1 B 2 C 3 4 D 5 E 6 F States 7 G 8 Experiment 4, Period 17 8 Players Aggregated H 9 10 I J Comparison To All Information Probability Kullback-Leibler = 0.133 0.900 0.9 Omniscient IA Mechanism 0.8 0.800 0.7 0.700 Probability 0.6 0.600 0.5 0.500 0.4 0.400 0.3 0.300 0.2 0.200 0.1 0.100 0.0 0.000 Series1 Series2 A 1 B 2 C 3 4 D 5 E F G 6 States 7 8 9 Experiment 4, Period 17 9 Players Aggregated H 10 I J Comparison To All Information Probability 0.9 Probability 0.8 Omniscient 0.7 IA Mechanism 0.6 market 0.5 Best Individual 0.4 0.3 0.2 0.1 0.0 A B C D E F States Experiment 4, Period 17 G H I J KL Measures for Private Info Experiments No Information Market Prediction Best Player Nonlinear Aggregation Function 1.977 (0.312) 1.222 (0.650) 0.844 (0.599) 0.553 (1.057) 1.501 (0.618) 1.112 (0.594) 1.128 (0.389) 0.214 (0.195) 1.689 (0.576) 1.053 (1.083) 0.876 (0.646) 0.414 (0.404) 1.635 (0.570) 1.136 (0.193) 1.074 (0.462) 0.413 (0.260) 1.640 (0.598) 1.371 (0.661) 1.164 (0.944) 0.395 (0.407) Experimental Economics Program Group Size Performance Experimental Economics Program Did the Markets Pick out Experts? Group Exp 1 Exp 2 Exp 3 Exp 4 Exp 5 Random 1.36 0.93 1.18 1.12 1.15 Payoff 1.45 1.09 1.24 1.13 1.39 Value 0.72 0.91 0.94 1.13 1.22 Optimal 0.53 0.72 0.75 0.83 0.77 •KL measure of all query data •Pick groups of 3 Experimental Economics Program Did Previous Queries Pick out Experts? Group Exp 1 Exp 2 Exp 3 Exp 4 Exp 5 Random 1.15 0.92 1.18 1.07 1.21 Query 0.78 0.89 0.71 0.92 0.81 Optimal 0.60 0.59 0.69 0.72 0.72 •KL measure of second half of query data •Pick groups of 3 Experimental Economics Program Public Information • Information observed by more than one • Double counting problem Experimental Economics Program Information Aggregation with Public Information Kullback-Leibler = 2.591 1.2 1.2 120.00% Probability Probability Probability 11 100.00% 0.8 0.8 80.00% Omnicient Omnicient Omnicient Public Public Public IAM IAM No Info 0.6 0.6 60.00% 0.4 0.4 40.00% 0.2 20.00% 0.2 000.00% AA B AB C BC D D E C D E E FF F States States States G GG HHH Public Info Experiment 3, Period 9 11 Players Aggregated III JJJ Dealing with Public Information: Add a Game to the Second Stage Event 1 Stage 1: Information Market Event 2 Event 3 Call Market to Solicit Risk Attitudes Time Event 4 Event 5 Stage 2: Probability Reporting & Aggregation Event 6 Event 7 Event 8 Individual Report of Probability Distribution Matching Game to Recover Public Information Modified Nonlinear Aggregated Function Experimental Economics Program Assumptions • Individuals know their public information • Private & Public Info Independent • Structure of Public Info Arbitrary Experimental Economics Program Matching Game Outcome A B C Reports of Probability Distribution Player 1: q1 q11 q12 q13 Player 2: q2 q21 q22 q23 Player 3: q3 . . . q31 q32 q33 . . . . . . Choose player (3) by Max (match function) . . . Player 1’s Payoff: (match function)*(C1+C2*Log(q33)) Match function: f(q1,q2)=(1-0.5*sum(abs(q1i-q2i)))^2 Experimental Economics Program Matching Game • Any match function f(q1,q2) with property – Max when q1=q2 • Multiple Equilibria • Payoff increases as entropy decreases • Hopefully, individuals report public information Experimental Economics Program Aggregation Function with Public Information Correction Bayes Law with a) Behavioral Correction b) Public Info Correction 1 2 N p1s p 2s p Ns ... q1s q 2s q Ns Ps | I 1 2 N p1s p 2s p Ns ... s q1s q 2 s q Ns Normalizing constant for individual risks i=r(V i /i)c Holding value/Risk - measure relative risk of individuals “market” risk ~sum of prices/winning payoff Experimental Economics Program Public Information Experiments • 5 Experiments • Various Information Structures – All subject received 2 private draws & 2 public draws – All subject received 3 private draws & 1 public draws – All subject received 3 private draws & half of the subjects receive 1 public draws – All subject received 3 private draws & 1 public draws. 2 groups of independent public information. • 9 to 11 participants in each experiments Experimental Economics Program Correcting for Public Information Kullback-Leibler = 0.291 1.2 Omnicient 1 sim aggr IAM 0.8 IAM (true public) 0.6 0.4 0.2 0 A B C D E F G H Public Info Experiment 3, Period 9 11 Players Aggregated I J KL Measures for Public Info Experiments Public Info Correction Perfect Public Info Correction No Info Market Prediction Best Player Nonlinear Aggregation Function 2 draws for all 1.332 (0.595) 0.847 (0.312) 0.932 (0.566) 2.095 (1.196) 0.825 (0.549) 0.279 (0.254) 2 draws for all 2 draws for all 1.420 (0.424) 0.979 (0.573) 0.919 (0.481) 2.911 (2.776) 0.798 (0.532) 0.258 (0.212) 3 3 draws for all 1 draws for all 1.668 (0.554) 1.349 (0.348) 1.033 (0.612) 2.531 (1.920) 0.718 (0.817) 0.366 (0.455) 4 3 draws for all 1 draws for half 1.596 (0.603) 0.851 (0.324) 1.072 (0.604) 0.951 (1.049) 0.798 (0.580) 0.704 (0.691) 1.528 (0.600) 0.798 (0.451) 1.174 (0.652) 0.886 (0.763) 1.015 (0.751) 0.472 (0.397) Private Info Public Info 1 2 draws for all 2 Expt 5 3 draws for all Two groups of public info Experimental Economics Program Summary • IAM with public info correction did better than best person. • IAM with public info correction did better than markets in 4 out of 5 cases. • IAM corrected with true public info did significant better than all other methods. Experimental Economics Program Experimental Economics Program or ab ov e -$ 11 62 .9 m -$ 11 47 .3 m -$ 11 33 .3 m -$ 11 20 .2 m -$ 11 07 .7 m -$ 11 00 .4 m -$ 10 93 .2 m -$ 10 86 .0 m -$ 10 78 .8 m -$ 10 71 .6 m -$ 10 64 .3 m -$ 10 51 .8 m -$ 10 38 .7 m $1 16 2. 9m $1 14 7. 3 $1 13 3. 3 $1 12 0. 2 $1 10 7. 7 $1 10 0. 4 9. 1m -$ 10 24 .7 m -$ 10 0 Actual Value $1053m $1 09 3. 2 $1 08 6. 0 $1 07 8. 8 $1 07 1. 6 $1 06 4. 3 $1 05 1. 8 $1 03 8. 7 $1 02 4. 7 $1 00 9. 1 $0 Implied Probabilities of Revenue Bins, September 2003 35% Official Projection 30% 25% 20% 15% 10% 5% 0% Implied Probabilities of Operating Profit Bins, September 2003 70% Official Projection Actual Value $113m 60% 50% 40% 30% 20% 10% 0% $0 $37.1m $37.1 - $46.1m $46.1m $54.4m $54.4 $62.0m $62.0 $69.3m $69.3 $73.6m $73.6 $77.8m $77.8 $82.0m $82.0 $86.2m $86.2 $90.4m Experimental Economics Program $90.4 - $94.7 - $102.0 - $109.6 - $117.9 - $126.9m $94.7m $102.0m $109.6m $117.9m $126.9m or above Supplementary Experimental Economics Program Previous Research • Academic Studies – Information Aggregation in Markets • Plott, Sunder, Camerer, Forsythe, Lundholm, Weber,… – Pari-mutuel Betting Markets • Plott, Wit & Yang • Real World Applications – Iowa Electronic Markets – Hollywood Stock Exchange – HP Information Markets – Newsfuture – Tradesport.com – … Experimental Economics Program Risk Attitudes 1.000 0.900 0.800 0.700 0.600 Risk Loving 0.500 Risk Neutral 0.400 0.300 Risk Averse 0.200 0.100 0.000 A B C D E F G H Experimental Economics Program I J Dealing with Risks Attitudes: Two-Stage Mechanism Event 1 Stage 1: Information Market Event 2 Event 3 Call Market to Solicit Risk Attitudes Time Event 4 Event 5 Event 6 Event 7 Stage 2: Probability Reporting & Aggregation Individual Report of Probability Distribution Nonlinear Aggregated Function Event 8 Experimental Economics Program Probability Reporting Outcome A B C Reports of Probability Distribution p1 p2 Pays C1+C2*Log(p3) Experimental Economics Program p3 Second Stage: Aggregation Function Bayes Law with Behavioral Correction 1 1s 2 2s N Ns N Ns p p ...p Ps | I 1 2 p1s p2s ...p s Normalizing constant for individual risks i=r(V i / i)c Holding value/Risk - measure relative risk of individuals “market” risk ~sum of prices/winning payoff Experimental Economics Program Private Information Experiments • 5 Experiments • Various Information Conditions • – All subject received 3 draws – Half received 5 draws, half received 1 draw – Half received 3 draws, half received random number of draws 8 to 13 participants in each experiments Experimental Economics Program Next Step • Field Test (Fine and Huberman) … Experimental Economics Program