2. “Prediction Markets” Market
Search Spot Speculative Matching
Future Currency Bet Stock Bond
Insurance Hedging Decision Gambling
Context Direct
3. To Evaluate Institutions
Institution A
When They Use Compare Quality
Similar Inputs Of Outputs
Institution B
4. Collective Forecasting
Forecasts
On Requested
Topics
User
Contributions User Scores
Engagement
5. Issues
Input: Contributions Output: Forecasts, Scores
What info can express? What questions can ask?
How account for costs? How account for value?
Who let in where? Use or validate system?
Enough Incentives Should adjust outputs?
T-shirts enough? Who let see outputs?
Zero-sum scoring?
Sabotage & manipulation
Limit Costs Legal, P.R. risks?
Awkward Interface
Wait for offer accept
Retribution
6. Collective Forecasting Questions
Consensus
What exactly is What exactly are
my influence? my incentives?
My Forecast My Score
How exactly do
I express my Truth
opinion?
7. Editing Interface Is Transparent
If my edit increases
I directly
the consensus chance
Consensus
change the
of true state, I win.
consensus
If decreases, I lose.
My Edits My Score
Truth
8. Issues
Input: Contributions Output: Forecasts, Scores
What info can express? What questions can ask?
How account for costs? How account for value?
Who let in where? Use or validate system?
Enough Incentives Should adjust outputs?
T-shirts enough? Who let see outputs?
Zero-sum scoring?
Sabotage & manipulation
Limit Costs Legal, P.R. risks?
Awkward Interface
Wait for offer accept
Retribution
9. Factors Might Influence Sales
E[Sales|Factor]
Economy recovers fast? P[Factor]
Competitors introduce new version?
We do big promotion?
We lower prices? They lower prices?
We add distribution channel?
We add feature F? They add feature F?
Our defect rate very low?
10. Combo Betting
Win Place Show
Not Not Not
Win
Place
Show
All outcomes Yoopick Facebook Application
11. Sport Finals Tickets
UEFA Austria Croatia Czech Germany Poland Portugal Switzerl. Turkey
EURO
2008
France
Greece v.
Greece Ticket if Greece in Finals
Croatia
Italy
Netherl.
Romania
Russia
Actual
Spain
Game
Sweden
13. PAM Scenario Trade
Return to Focus ?
Update
Payoffs: If SAum3 03105-125& Ave. pay
Select New Price IQcs4 03 <85 IQcs4 03 >85 <85 65%
Max Up 95.13% +$34.74 -$85.18 -$19.72
Buy
Saudi Arabian 10% Up 68.72% +$2.74 -$3.28 -$1.07
Economic Health
You Pick 65 % +1.43 -2.04 +0.34
125
15 30
70 65 No Trade 62.47% $0.00 $0.00 $0.00
100 40 35
94
100
10% Dn 56.79% -$2.61 +$2.74 -$1.12
35 25
Sell
30 35 Exit Issue 48.54% -$15.34 +$26.02 -$6.31
10 10
75
Max Dn 22.98% -$120.74 +$96.61 -$22.22
1 2 3 412
03 03 03 03 04 04
?
Return to Form Execute a Trade
If US military involvement in Saudi Arabia in 3rd Quarter 2003 is not
between 105 and 125, this trade is null and void. Otherwise, if
Iraq civil stability in 4th Quarter 2003 is below 85, then I will receive
$1.43, but if it is not below 85, I will pay $2.04.
Abort trade if price has changed? Execute
14. Imagine A Dashboard Ave. Worth:
$12,459
Us Them A Them B
Base Price $240 $187 $320
Ship Date May ’09 Mar ’09 July ’09
Features Autozoop 38% 69% 15%
Fizzywoo 59% 8% 43%
Unit Sales Total 120K 148K 97K
Base model 82K 65K 88K
Via internet 43K 12K 73K
Promotion Magazine $30K $50K $3K
Circulars $45 $34K $39K
15. Ask For Detail Ave.Worth:
$12,459
Us Them A Them B
Base Price $240 $187 $320
Ship Date May ’09 Mar ’09 July ’09
Features Autozoop 38% 69% 15%
Fizzywoo Ship Date
59% 8% 43%
Them B
Unit Sales Total 120K 148K 97K
2009 2010
Base model 82K 65K 88K
Via internetM A43K J J A S O N D
JF M 12K 73K
Promotion Magazine $30K $50K $3K
Circulars $45 $34K $39K
16. Make An Edit Ave. Worth:
$12,459
Us Them A Them B
Base Price $240 $187 $320
Ship Date May ’09 Mar ’09 July ’09
Features Autozoop 40% 69% 15%
Fizzywoo 59% 8% 43%
If We Have Autozoop,
Unit Sales Total 120K 148K you gain $53
97K
Base model 82K 65K if We Don’t88K It
But Have
You lose $78. OK?
Via internet 43K 12K 73K
Promotion Magazine $30K $50K $3K
Circulars $45 $34K $39K
17. Make an Assumption Scenario: Ave. Worth:
15% $10,724
Us Them A Them B
Base Price $240 $187 $253
Assume Mar
Ship Date Apr ’09 Mar ’09
Features Autozoop 38% 69% 4%
Fizzywoo 59% 8% 13%
Unit Sales Total 120K 148K 107K
Base model 82K 65K 94K
Via internet 43K 12K 84K
Promotion Magazine $30K $50K $17K
Circulars $45 $34K $49K
18. Add 2nd Assumption Scenario: Ave. Worth:
2.3% $10,982
Us Them A Them B
Base Price $240 $187 $253
Assume Mar
Ship Date Apr ’09 Mar ’09
Features Autozoop 38% 69% 4%
Fizzywoo 59% 8% 13%
Unit Sales Total 185K 148K 107K
Base model 97K 65K 94K
Via internet 78K 12K 84K
Assume $40K $50K
Promotion Magazine $17K
Circulars $45 $34K $49K
19. Edit As Before Scenario: Ave. Worth:
2.3% $10,724
Us Them A Them B
Base Price $240 $187 $253
Assume Mar
Ship Date Apr ’09 Mar ’09
Features Autozoop 42% 69% 4%
Fizzywoo 59% 8% 13%
If we have Autozoop,
Unit Sales Total 185K 148K you gain $53
107K
Base model 97K 65K if we don’t94K it
But have
You lose $78. OK?
Via internet 78K 12K 84K
Assume $40K $50K
Promotion Magazine $17K
Circulars $45 $34K $49K
20. Combo Market Maker Best of 5 Mechs
3 subjects, 7 prices, 5 minutes 6 subjects, 256 prices, 5 minutes
8 Variables = 256 States
3 Variables = 8 States
1.600
0.300
0.250 1.400
0.200
KL Distance
KL Distance
1.200
0.150
1.000
0.100
0.800
0.050
0.000 0.600
al
al
on
lue
lue
er
ol
on
er
ol
id u
id u
Po
Po
ak
ak
cti
cti
Va
Va
div
div
tM
Au
tM
Au
ion
ion
d
In
d
In
ine
ke
ine
ke
le
pin
le
pin
ub
ub
ar
ar
mb
mb
gO
gO
M
M
Do
Do
Co
Co
Lo
Lo
21. MSR Info vs. Time – 7 Prices
1
1
% Info Agg. =
0.5
1- KL(prices,group)
KL(uniform,group)
0 0
0 5 10 15
-0.5
0 5 10 15
Minutes
-1 -1
22. MSR Info vs. Time – 255 prices
11
% Info Agg. =
0.5
1- KL(prices,group)
KL(uniform,group)
0 0
0 5 10 15
-0.5
0 5 10 15
Minutes
-1 -1
23. Issues
Input: Contributions Output: Forecasts, Scores
What info can express? What questions can ask?
How account for costs? How account for value?
Who let in where? Use or validate system?
Enough Incentives Should adjust outputs?
T-shirts enough? Sabotage & manipulation
Zero-sum scoring?
Can keep results secret?
Limit Costs Legal, P.R. risks?
Awkward Interface
Wait for offer accept
Retribution
24. A Simple Implementation
f1>1 f2<1
States
Prices
+ q1 $1 if A&B
+
- - q2 $1 if B
User Assets
B
A
LISP: http://hanson.gmu.edu/mktscore-prototype.html
26. Environments: Goals, Training
(Actually: X Z Y)
Want in Environment: Case A B C
1 1 - 1
Many variables, few directly related 2 1 - 0
3 1 - 0
Few people, each not see all variables 4 1 - 0
5 1 - 0
Can compute rational group estimates 6 1 - 1
7 1 - 1
Explainable, fast, neutral 8 1 - 0
Training Environment:
9 1 - 0
10 0 - 0
3 binary variables X,Y,Z, 23 = 8 combos Sum: 9 - 3
P(X=0) = .3, P(X=Y) = .2, P(Z=1)= .5 Same A B C
A -- -- 4
3 people, see 10 cases of: AB, BC, AC B -- -- --
Random map XYZ to ABC C -- -- --
27. Experiment Environment
(Really: W V X S U Z Y T)
8 binary vars: STUVWXYZ Case A B C D E F G H
1 0 1 0 1 - - - -
28 = 256 combinations 2 1 0 0 1 - - - -
3 0 0 1 1 - - - -
20% = P(S=0) = P(S=T) 4 1 0 1 1 - - - -
5 0 1 1 1 - - - -
= P(T=U) = P(U=V) 6 1 0 0 1 - - - -
7 0 1 1 1 - - - -
=… = P(X=Y) = 8 1 0 0 1 - - - -
9 1 0 0 1 - - - -
P(Y=Z) 10 1 0 0 1 - - - -
6 people, each see 10 Sum 6 3 4 10 - - - -
cases: ABCD, EFGH, ABEF, Same A B CD E F G H
CDGH, ACEG, BDFH A -- 1 26 -- -- -- --
B -- -- 73 -- -- -- --
C -- -- -- 4 -- -- -- --
random map STUVWXYZ D -- -- -- -- -- -- -- --
to ABCDEFGH
…
28. Ad Agency Decision Markets
$ Revenue if
$1 $1 if Switch Switch
P(S)
E(R | S)
E(R) Compare!
$ Revenue E(R | not S)
$ Revenue if
$1 if not
not Switch
Switch
29. Corporate Applications
E[ Revenue | Switch ad agency? ]
E[ Revenue | Raise price 10%? ]
E[ Project done date | Drop feature? ]
E[ Project done date | Add personnel? ]
E[ Stock price | Fire CEO? ]
E[ Stock price | Acquire firm X? ]
30. Decision Market Requirements
Legal permission Public credibility
Outcome Traders
Measured Enough informed
Aggregate-enough Decision-insiders
Linear-enough Enough incentives
Conditional-enough Anonymity
Decision Prices
Distinct options Intermediate-enough
Important enough Can show enough
Enough influence
Editor's Notes
<number>
First want to be like a standard economics experiment, whose standards have evolved over the decades to allow them to best be comparable to theory and to each other. Explainable = we take people “off the street” and have to explain everything to them in a few minutes before they start an experiment. Fast = want it to be over within a few minutes, so we can run many periods to get enough data to make significant comparisonsNeutral = emotionally neutral, is standard econ practice, since emotional salience creates lots of variation in outcomes that is hard to model and control for. Expect real situation to have many related variables, but most relations are not direct – experts must look at data to guess if correlation see is just coincidence, or real. Most experts specialize in certain kinds of data, do not study all dataFew people = want to explore thin market limit, where there are more markets than people. We want to be able to compare results to an ideal benchmark of people sharing all info and inferring all they can from it. Of course people rarely get close to this benchmark, and we don’t expect our markets to. Since take people “off the street”, so need a “training” situation to introduce them to the basic concepts.The “test” situation is more challenging, and we expect a better indication of how these mechanisms will work in the field test. 3 binary variables gives 8 combinations of variable values. We randomly pick X first, then pick Y based on X, then pick Z.Since only one of the three pairs of variables is related, only 1/3 of pairs are related. Three people for 8 markets, are in the thin markets regime.They see the same 10 cases drawn as described, but each person not see one variable. And they see the cases in terms of labels A,B,C, not knowhing which is which re X,Y,Z. See sample data to right. People have to look at the correlations and frequencies of the variables they see to try to guess which are X,Y,Z.The one that comes up 1 the most often is most likely X, and the pair that are the least similar are most likely X,Y.Here this subject happens to see the two variables that are actually related, A and C. <number>
The test situation has 8 binary variables, and so 256 combinations of variable values. With only six people trading in 256 markets, this is way into the thin market problem. Here we create a chain of related variables. Since there are 28 variable pairs, and 7 pairs directly related, only