a new understanding of prediction markets via no-regret learning
TRANSCRIPT
Prediction Markets
• Outcomes i in {1,…,N}• Prices pi for shares that pay off in outcome i• Market scoring rules
Comparison
Market Scoring Rule LearningN outcomes: 1,…,N
N experts: 1,…,N
Prediction by price: Prediction by weights:
Price updating rule for LMSR: Weight updating rule for weighted majority:
tip , tiw ,
Connection-Paving the Road
• Each outcome i can be interpreted as an expert, pricing contract i at $1 and other contracts at $0.
• Let’s assume market run forever before any outcome realizes. When trader comes in and do short-selling, the money paid by the N experts is like a loss.
Connection – Paving the Road
• Define the loss of an expert: at each time t, an trader comes to the market maker, and buys shares on the contract of outcome i.
• Let us just assume that , i.e. only short selling happens.
tir ,0, tir
Connection – Paving the Road
The loss for expert i is:
Choose a s.t.
T
tti
T
ttiTi rrL
1,
1,, )
1(
1
||, ,tirit
Connection-Paving the Road
• As a market maker, your job is to combine the opinions of your experts, and decide the price of each contract.
• Your price should be set properly so that traders don’t want to trade with you at all. Your price for each outcome sums up to 1.
• Still, you lose money when traders come in and sell contracts to you.
Connection – Paving the Road
• Definition of cumulative loss of a market maker (the money market maker paid for all trades):
• -stable cost function: =>
tit
T
t
N
iiTA rqpL ,1
1 1, )(
1
2)0()(
1,
TCqCL TTA
Connection – Paving the Road
• Definition of cumulative loss of a market maker (the money market maker paid for all trades):
• -stable cost function: =>
tit
T
t
N
iiTA rqpL ,1
1 1, )(
1
2)0()(
1,
TCqCL TTA
Actual loss for the market maker
Connection – Paving the Road
• Definition of cumulative loss of a market maker (the money market maker paid for all trades):
• -stable cost function: =>
tit
T
t
N
iiTA rqpL ,1
1 1, )(
1
2)0()(
1,
TCqCL TTA
Actual loss for the market maker
Lower bound
Notation Change
T
tti
T
tti
T
ttiTi lrrL
1,
1,
1,, )
1(
1
T
t
N
itititit
T
t
N
iiTA lwrqpL
1 1,,,1
1 1, )(
1
Connection: Learning to MSR
• This becomes a learning problem. Recall Weighted Majority Updating Rule:
For LMSR cost function: Set the learning rate to be: =>
b/
N
j
bq
bq
titj
ti
e
ew
1
/
/
,,
,
N
j
l
l
ti
ttj
tti
e
ew
1
,,
,
Connection: MSR to Learning
• Recall– We set:– In Theorem 2:
• If LMSR => B= b log N (the proof is waived in the paper (Lemma 5))
• Put all together into Theorem 2 we have:
b/ )/(2 TB
Connection
• Cost Function:– Differentiability, Increasing Monotonicity and
Positive Translation Invariance– Agrawal et al show that:
– This paper also show that the instant price is actually the p in the expression.
• The answer is to set:
• (Theorem 3): The cost function based on the above equation is equivalent to a market scoring rule market using the scoring rule )( psi
• Theorem 3:– Step 1:
– Step 2:• Like HW2, just replace the log scoring rule and cost
function with the equation above and do some KTT condition.
MSR Cost Function
Scoring Rule Convex Function
)( psi
HW2 with LMSR, but not applicable to all scoring rules
)q(
C
MSR Cost Function
Scoring Rule Convex Function
)( psi
HW2 with LMSR, but not applicable to all scoring rules
)q(
C
• Recap B:• Lemma 5: B can be up-bounded by:
• Let us plug this into Theorem 2:
• We have a new bound:
BCqCq TTi )0()(,i
max
Recall FTRL bound:
Discussion
• Continuous price updates versus discrete weight updates
• Direction of implication– Any strictly proper market scoring rule implies
corresponding FTRL algorithm with strictly convex regularizer
– Any FTRL algorithm with differentiable and strictly convex regularizer implies strictly proper scoring rule.
Discussion
• Extensive learning literature may aid progress in prediction markets.
• PermELearn algorithm– Applied to combinatorial markets