Presented By: Ofir Chen Based on: Designing Markets for Prediction by Yilling Chen and David M....

31
Presented By: Ofir Chen Prediction Markets Based on: “Designing Markets for Prediction” by Yilling Chen and David M. Pennock 2010

Transcript of Presented By: Ofir Chen Based on: Designing Markets for Prediction by Yilling Chen and David M....

Page 1: Presented By: Ofir Chen Based on: Designing Markets for Prediction by Yilling Chen and David M. Pennock 2010.

Presented By: Ofir Chen

Prediction Markets

Based on: “Designing Markets for Prediction” by Yilling Chen and David M. Pennock 2010

Page 2: Presented By: Ofir Chen Based on: Designing Markets for Prediction by Yilling Chen and David M. Pennock 2010.

Outline:

- Motivation

- Market Makers- Reminder+ (SR, CF), DPM, Utility function, SCPM

- Incentive compatibility - Agents interaction

- Manipulation

- Expressiveness

- What is Truth- Peer prediction and BTS

Page 3: Presented By: Ofir Chen Based on: Designing Markets for Prediction by Yilling Chen and David M. Pennock 2010.

Motivation

We’d like to predict an event of interest

Ideally, we’d like to make agents say the truth, the whole truth and nothing but the truth – and do it NOW

We’re willing to pay for it…

Market Requirements:- Liquidity - Bounded loss- Discourage manipulation- Extract predictions easily

How can we create such a market??

Page 4: Presented By: Ofir Chen Based on: Designing Markets for Prediction by Yilling Chen and David M. Pennock 2010.

Liquidity:

Liquidity is the ability to trade instantly with no significant movement in the price

How do we encourage agents to talk...

- Simple: the Market Maker (MM) pays them.

- We’ve already seen last time that by subsidizing the market we increase liquidity.

- we’d like to bound that subsidy. we’ll talk about it later…

Page 5: Presented By: Ofir Chen Based on: Designing Markets for Prediction by Yilling Chen and David M. Pennock 2010.

Bergman Divergence (BD):

How do we make them say the truth…

Given a convex function y=f(x) the the BD is:

Nonlinear, non-negative function.

The expected value over , given and :

( , ) ( ) ( ) ( )( )fD p q f p f q f q p q

~

~

[ ( , )] ( , ) ( )

argmax ( [ ( , )])

argmax ( ( , ) ( ))

argmax ( ( , ))

i q f i f

p i q f i

p f

p f

E D e p D q p g q

E D e p

D q p g q

D q p q

That’s a scoring rule for p!!

ie p ~i q

Page 6: Presented By: Ofir Chen Based on: Designing Markets for Prediction by Yilling Chen and David M. Pennock 2010.

Scoring Rule (SR)

With this we can create our first market – Market Scoring Rule (MSR):- Sequential trading.- updating r to r’, requires paying the previous agent - Therefore payoff is - The final r is the market’s prediction.

- Disadvantages: - Not natural, no real contracts are traded.- Participating only once- These limitations may make the market less appealing to potential

agents.

- Solution: Cost Functions

( ) ( , )i f i iS p D e p h

( )iS r( ')iS r

Page 7: Presented By: Ofir Chen Based on: Designing Markets for Prediction by Yilling Chen and David M. Pennock 2010.

Cost function (CF)

Idea: Trade Arrow-Debreu (AD) contracts (instead of probabilities) .AD contract pays $1 if the event happens, and $0 otherwise

Notations and Market definition:

- is a vector indicating the total number of shares of each type ever sold.

- is the amount of shares of type “i”.

- When changing (by buying/selling): Pay

- Price of share i: ,

( ) ( )new oldC q C q

q

q

iq

( )( )i

i

C qp q

dq

1

( ) 1m

ii

p q

Page 8: Presented By: Ofir Chen Based on: Designing Markets for Prediction by Yilling Chen and David M. Pennock 2010.

Cost function (cont.)

Desired properties of a CF:

• Differentiability (to calculate prices)

• Monotonically increasing in

• Positive translation invariant

q

( 1) ( )C q k C q k

Page 9: Presented By: Ofir Chen Based on: Designing Markets for Prediction by Yilling Chen and David M. Pennock 2010.

Cost function Market from MSR (Chen, Vaughan’10)

There’s a one to one mapping between CFM and MSR: Such that and , Agent who change p to p’ in an MSR receives same payoff as changing q to q’ in a CFM.

Agents will profit the same changing q in an Cost Function based Market (CFM) had they changed p in an MSR iff the following holds:

Corollary, there’s a mapping from CF to SR, not presented here.

, ', , 'p p q q ( )C q p ( ') 'C q p

1 1

( ) sup ( )m

m m

i i i ip i i

C q p q p s p

Page 10: Presented By: Ofir Chen Based on: Designing Markets for Prediction by Yilling Chen and David M. Pennock 2010.

DPM – Dynamic Parimutuel Market

- Parimutuel: Winning agents split the total pool of money at the end.

- Dynamic: Prices vary before outcome is determined (same as CFM)

- Main difference: contracts are not Arrow-Debreu.

Each contract i pays off:

The more “winners” the smaller the profit. Is the final q.

- MM has to initially buy contracts to avoid 0 division in price function.

- is the market’s prediction

( )( )

ff

i fi

C qo q

q

p

2( )

( )i i

i

ij

q qp q

C q q

fq

Page 11: Presented By: Ofir Chen Based on: Designing Markets for Prediction by Yilling Chen and David M. Pennock 2010.

Utility function Markets- Utility: utility of an outcome is the total satisfaction received by it.

- Dynamic, AD contracts, probability price market, like CFM.

- MM sets a subjective probability for all events

- MM has a money value vector upon possible outcomes

- MM has a utility function u(m)

- The instantaneous price is defined as the infinitesimal change in the MM utility:

- MM’s expected utility: remains constant (Chen, Pennock ‘07)

1n

( )

( )i i

ij j

u mp

u m

m

( )j jj

u m

Page 12: Presented By: Ofir Chen Based on: Designing Markets for Prediction by Yilling Chen and David M. Pennock 2010.

SCPM: Sequential Convex Parimutuel Mechanism

(Not detailed)

- Agents state their wanted state vector, quantity, and max-price

- the MM decides how many AD contracts to sell to maximize its profit by solving a convex optimization problem.

- Prices are determined using VCG mechanism.

- Prices reflect the market’s prediction

Page 13: Presented By: Ofir Chen Based on: Designing Markets for Prediction by Yilling Chen and David M. Pennock 2010.

Bounded loss:

Subsidies are limited – MM would like to bound its losses.

- MSR:

- CFM:

- DPM: initial market subsidy

- Utility Market: bounded if m is bounded (from below) or u(m) is bounded (from above)

- SCPM: bounded

0( ) ( )i i iS e S r

sup (sup ( ( ) (0)))q i iq C q C

Page 14: Presented By: Ofir Chen Based on: Designing Markets for Prediction by Yilling Chen and David M. Pennock 2010.

So far…

In all the markets we’ve seen, telling the truth should potentially maximize traders’ profit.

But what if…

- Agents can talk to/signal each other?

- Agents manipulate the market?

We’d like to refine our models to incorporate those real-life scenarios.

Page 15: Presented By: Ofir Chen Based on: Designing Markets for Prediction by Yilling Chen and David M. Pennock 2010.

Incentive Compatibility – terms

- BNE – Bayesian Nash Equilibrium.we’ll say that a market is in BNE when all agents already maximized their profits, and any further action from any agent will damage his profit.Most importantly: In a BNE rational traders stop trading.

- PBE – Perfect Bayesian Equilibrium we’ll say that a market is in PBE if through every step, all agents acted to maximize their (expected) utility, and eventually reached an equilibrium.

- Dominant strategy – A strategy is dominant if, regardless of what any other players do, the strategy earns a player a larger payoff than any other. Hence, a strategy is dominant if it is always better than any other strategy.

- Equilibrium Strategy – a strategy that leads to an equilibrium.

Page 16: Presented By: Ofir Chen Based on: Designing Markets for Prediction by Yilling Chen and David M. Pennock 2010.

Incentive Compatibility

How do we encourage agents to say the truth now and nothing but the truth

- We’d like agents to reveal their information truthfully and immediately. Push the market to a truthful equilibrium as fast as possible.

- Rewarding truth-tellers is first step: agents don’t waste time calculating strategies before placing their bids.

- Picking the right type of market is another step.

- Problems:- No-trade theorem(‘82): “Rational traders won’t trade in an all-

rational Continuous-Double-Auction (CDA) market.”- Gradual information leakage may be more beneficial when traders

can participate more than once (Chakraborty and Yilmaz ‘04)- Agents may benefit from manipulations/interactions in the market.

Page 17: Presented By: Ofir Chen Based on: Designing Markets for Prediction by Yilling Chen and David M. Pennock 2010.

Incentive Compatibility – agents interaction

- Signaling through trades may lead agents to lie (“bluffing”) to profit by correcting their bluffs later.

- In reality, it’s hard to avoid agents interactions… Limiting agents to participate only once may partially helpbut keep in mind the problems in the sequential model (MSR).

- In markets that allow any interaction between agents, truth telling is not an equilibrium strategy (Chen ’09)

- Today, researches focus on extracting predictions from a BNEs, even if they are not the truth telling BNE.

Page 18: Presented By: Ofir Chen Based on: Designing Markets for Prediction by Yilling Chen and David M. Pennock 2010.

Incentive Compatibility example model (Chen 2009)

- Market: LMSR (Logarithmic MSR)- Event w with 2 outcomes- n players, each gets si correlated to the event w- Distribution of si and w is common knowledge

- Players play sequentially (1) or when they decide (2).- si|w’s are independent (3) or si’s are independent unconditionally (4).

Analysis shows:- Information is better aggregated when players play sequentially. - If si|w’s are independent, truth telling is the only PBE, Agents tell the

truth as soon as possible. - If si’s are independent unconditionally, the BNE is unknown. Truth-telling

is not even a good strategy, and a BNE might not exist.

Page 19: Presented By: Ofir Chen Based on: Designing Markets for Prediction by Yilling Chen and David M. Pennock 2010.

Manipulation

- An agent can manipulate the market in several ways:

- Take action to change event’s outcome.

- Send misleading signals inside the market.

- Send signals from outside the market.

Page 20: Presented By: Ofir Chen Based on: Designing Markets for Prediction by Yilling Chen and David M. Pennock 2010.

Manipulation - Changing event’s outcome

- Consider a company with n employees that uses a PM to predict its product delivery date.

- An employee can affect the outcome by acting from within the company.

- Note that the company has a desired outcome.

- Shi, Conitzer and Guo (‘09) showed the following:- Allowing one time participation in an MSR market will encourages

the agents to play truthfully, and prevent sending misleading signals between agents.

- The MM can incentivize the agents to not manipulate the outcome by paying times more than in a normal MSR.( )n

Page 21: Presented By: Ofir Chen Based on: Designing Markets for Prediction by Yilling Chen and David M. Pennock 2010.

Manipulation – correlated markets

Consider 2 correlated markets:

- Alice trades in Market A

- Bob makes his trading decision in Market B

- Alice can now trade in market B and potentially benefit from her decision in market A, even if the latter was not truthful.

- Let’s see an example…

Page 22: Presented By: Ofir Chen Based on: Designing Markets for Prediction by Yilling Chen and David M. Pennock 2010.

Manipulation – correlated markets - example

Market A: LMSR, b = 0.1 Market B: LMSR b=1

0.1 log0.5 0.1 log0.4 0.4 log0.9

0.1 log0.5 0.1 log0.9 0.9 log0.9

i i i i

pays A gets from A pays Bob gets fromB

i i i i

a a a log a

a a a log a

MM seeds both markets with initial prob. 0.5

0.5 0.5

Alice changes prob A to 0.4

- Alice believes event w happens with probability 0.9 - Bob is not sure… he’s looking for easy profit (like most of us).

0.4

0.4• Bob follows her

and changes prob B to 0.4

• Alice changes prob B to 0.9

0.9

Page 23: Presented By: Ofir Chen Based on: Designing Markets for Prediction by Yilling Chen and David M. Pennock 2010.

Expressiveness

How do we encourage them to say the truth (now), the whole truth and nothing but the truth …

Motivation: We’d like agents to put as much data as possible in the market.

But How?

Combinatorial bids – bids on more than one outcome. Improves expressiveness!

- Example – horse race:- Horse A will finish before horse B. - Horse A won’t win and horse B won’t win. - The entire permutation of horses.

Page 24: Presented By: Ofir Chen Based on: Designing Markets for Prediction by Yilling Chen and David M. Pennock 2010.

Expressiveness (Cont.)

We’ll examine the market’s 2 computational challenges:

- Pricing: setting the price of a share such that it’s coherent with events probabilities.

- The Auctioneer Problem: Given a set of bids in a combinatorial auction, allocate items to bidders—including the possibility that the auctioneer retains some items—such that the auctioneer’s revenue is maximized.

Page 25: Presented By: Ofir Chen Based on: Designing Markets for Prediction by Yilling Chen and David M. Pennock 2010.

Expressiveness – known results

- Permutation betting: horse racing both auctioneer problem and pricing are hard. auctioneer problem under specific settings can be possible.

- Boolean betting: vector of {0,1}s both auctioneer problem and pricing are hard.

- Tournament betting: sport teams in a playoff tree, leaves are teams Pricing “team A advances to round k” is possible. the auctioneer problem is still hard

- Taxonomy betting: summing tree, leaves are base elements LMSR pricing is possible auctioneer problem and general pricing are hard.

Page 26: Presented By: Ofir Chen Based on: Designing Markets for Prediction by Yilling Chen and David M. Pennock 2010.

Expressiveness (cont.)

- Problems:- Events are obviously correlated, but it’s hard to price them as such. - Even if we could price events properly, analyzing the results is hard

- Recall that polynomial in the number of outcomes is actually exponential number of base events.

- In real life:- Under some settings and when number of all possible outcomes is

bounded and low, it is feasible to allow combinatorial bids. - In practice, it’s not commonly used.

Page 27: Presented By: Ofir Chen Based on: Designing Markets for Prediction by Yilling Chen and David M. Pennock 2010.

But what is truth??

Problem:- Truth may be subjective or non-verifiable:

- Rating the quality of a movie- Determine extinction year of the human race.

Solution:- Peer prediction: determine a relative truth.- Idea (Miller, Resnick, Zeckhauser ‘05) : evaluate Agent’s reports against

the reports of its peers.

Page 28: Presented By: Ofir Chen Based on: Designing Markets for Prediction by Yilling Chen and David M. Pennock 2010.

Peer prediction - (Miller, Resnick, Zeckhauser ‘05)

Consider the following setting:- Each agent gets a signal si on event w. distributions of w and si|w are

common knowledge, but w is not verifiable. - Agent i reports si’. - MM randomly picks a reference agent j and calculates- Agent i will be rewarded according to .

- At the case mentioned, truth telling will lead to a BNE.- Unfortunately, it’s not the only BNE… - Requires a mass of truth-tellers

- Further research shows that there are ways to make truth telling a unique equilibrium under this setting (Jurca and Faltings ‘07).

* ( ' | )i i js P s s*is

Page 29: Presented By: Ofir Chen Based on: Designing Markets for Prediction by Yilling Chen and David M. Pennock 2010.

BTS: Bayesian Truth Serum (Prelec ‘04)

Consider the following setting:- A simple poll – each agent states her opinion - In addition – each agent is asked to estimate the final distribution over

possible answers denoted by S.

- Agents’ score:- Opinion score: the more common it is the higher the score is.

- Poll estimation score: the denominator is the statistical distance between S and P.

- Truthful reporting is a BNE with these settings!- When allowing to reveal partial poll results, this is not the only BNE….- But even then, the gap between the updated poll (affected by ) and

the Agent’s true belief regarding the poll’s outcome (S) is reduced, allowing to extract true prediction from the polls outcome.

1

S P

ip

ip

Page 30: Presented By: Ofir Chen Based on: Designing Markets for Prediction by Yilling Chen and David M. Pennock 2010.

Summary

- We saw prediction markets of different kinds

- We understood some of the setbacks when those markets are used in reality, including some interesting ideas on how to overcome those

- You might have noticed most of the quotation brought here are from last decade, many new results, fast development.

- In reality some those markets can outperform regular polls and surveys.

Page 31: Presented By: Ofir Chen Based on: Designing Markets for Prediction by Yilling Chen and David M. Pennock 2010.

Questions