You are on page 1of 40

Algorithmica manuscript

Gaming Prediction Markets: Equilibrium Strategies
with a Market Maker

Yiling Chen
1
, Stanko Dimitrov
2
, Rahul Sami
3
, Daniel M. Reeves
4
, David M.
Pennock
4
, Robin D. Hanson
5
, Lance Fortnow
6
, Rica Gonen
7
1
Harvard University, School of Engineering and Applied Sciences, Cambridge, MA
2
University of Michigan, Dept. of Industrial and Operations Engineering, Ann Arbor, MI
3
University of Michigan, School of Information, Ann Arbor, MI
4
Yahoo! Research, New York, NY
5
George Mason University, Department of Economics, Fairfax, VA
6
Northwestern University, EECS Department, Evanston, IL
7
Yahoo! Research, Santa Clara, CA
Received: 2 June 2008 / Accepted: 29 April 2009
Abstract We study the equilibrium behavior of informed traders interacting with
market scoring rule (MSR) market makers. One attractive feature of MSR is that it
is myopically incentive compatible: it is optimal for traders to report their true be-
liefs about the likelihood of an event outcome provided that they ignore the impact
of their reports on the profit they might garner from future trades. In this paper, we
analyze non-myopic strategies and examine what information structures lead to
truthful betting by traders. Specifically, we analyze the behavior of risk-neutral
traders with incomplete information playing in a dynamic game. We consider
finite-stage and infinite-stage game models. For each model, we study the log-
arithmic market scoring rule (LMSR) with two different information structures:
conditionally independent signals and (unconditionally) independent signals. In
the finite-stage model, when signals of traders are independent conditional on the
state of the world, truthful betting is a Perfect Bayesian Equilibrium (PBE). More-
over, it is the unique Weak Perfect Bayesian Equilibrium (WPBE) of the game.
In contrast, when signals of traders are unconditionally independent, truthful bet-
ting is not a WPBE. In the infinite-stage model with unconditionally independent
signals, there does not exist an equilibrium in which all information is revealed
in a finite amount of time. We propose a simple discounted market scoring rule
that reduces the opportunity for bluffing strategies. We show that in any WPBE for

Preliminary versions of some of the results in this paper were presented in two
conference papers, Chen et al. [10] and Dimitrov and Sami [13].
This is an author-created version. The original publication is available at
www.springerlink.com. DOI 10.1007/s00453-009-9323-2.
2 Y. Chen et al.
the infinite-stage market with discounting, the market price converges to the fully-
revealing price, and the rate of convergence can be bounded in terms of the dis-
counting parameter. When signals are conditionally independent, truthful betting
is the unique WPBE for the infinite-stage market with and without discounting.
Key words Prediction markets, game theory, bluffing, strategic betting
1 Introduction
It has long been observed that, because market prices are influenced by all the
trades taking place, they reflect the combined information of all the traders. The
strongest form of the efficient markets hypothesis [14] posits that information is
incorporated into prices fully and immediately, as soon as it becomes available to
anyone. A prediction market is a financial market specifically designed to take ad-
vantage of this property. For example, to forecast whether a product will launch on
time, a company might ask employees to trade a security that pays $1 if and only
if the product launches by the planned date. Everyone from managers to develop-
ers to administrative assistants with different forms and amounts of information
can bet on the outcome. The resulting price constitutes their collective probability
estimate that the launch will occur on time. Empirically, such prediction markets
outperform experts, group consensus, and polls across a variety of settings [16, 17,
28, 3, 4, 18, 31, 12, 8].
Yet the double-sided auction at the heart of nearly every prediction market is
not incentive compatible. Information holders do not necessarily have incentive to
fully reveal all their information right away, as soon as they obtain it. The extreme
case of this is captured by the so-called no trade theorems [26]: When rational,
risk-neutral agents with common priors interact in an unsubsidized (zero-sum)
market, the agents will not trade at all, even if they have vastly different informa-
tion and posterier beliefs. The informal reason is that any offer by one trader is
a signal to a potential trading partner that results in belief revision discouraging
trade.
The classic market microstructure model of a financial market posits two types
of traders: rational traders and noise traders [24]. The existence of noise traders
turns the game among rational traders into a positive-sum game, thereby resolving
the no-trade paradox. However, even in this setting, the mechanism is not incen-
tive compatible. For example, monopolist information holders will not fully reveal
their information right away: instead, they will leak their information into the mar-
ket gradually over time to obtain a greater profit [7].
Instead of assuming or subsidizing noise traders, a prediction market designer
might choose to directly subsidize the market by employing an automated mar-
ket maker that expects to lose money. Hanson’s market scoring rule market maker
(MSR) is one example [20, 21]. MSR requires a patron to subsidize the market but
guarantees that the patron cannot lose more than a fixed amount set in advance,
regardless of how many shares are exchanged or what outcome eventually occurs.
The greater the subsidy, the greater the effective liquidity of the market. Since
Gaming Prediction Markets: Equilibrium Strategies with a Market Maker 3
traders face a positive-sum game, even rational risk-neutral agents have incentive
to participate. In fact, even a single trader can be induced to reveal information,
something impossible in a standard double auction with no market maker. Hanson
proves that myopic risk-neutral traders have incentive to reveal all their informa-
tion. However, forward-looking traders may not.
Though subsidized market makers improve incentives for information reve-
lation, the mechanisms are still not incentive compatible. Much of the allure of
prediction markets is the promise to gather information from a distributed group
quickly and accurately. However, if traders have demonstrable incentives to ei-
ther hide or falsify information, the accuracy of the resulting forecast may be in
question. One frequent concern about incentives in prediction markets is based on
non-myopic strategies: strategies in which the attacker sacrifices some profit early
in order to mislead other traders, and then later exploit erroneous trades by other
traders, thereby gaining an overall profit.
1.1 Our Results
In this paper, we study strategies under a logarithmic market scoring rule (LMSR)
in a Bayesian extensive-form game setting with incomplete information. We show
that different information structures can lead to radically different strategic prop-
erties by analyzing two natural classes of signal distributions: conditionally inde-
pendent signals and unconditionally independent signals.
For conditionally independent signals, we show that truthful betting is a Per-
fect Bayesian Equilibrium in finite-stage and infinite-stage models. Further, we
show that the truthful betting equilibrium is the unique Weak Perfect Bayesian
Equilibrium in this setting. Thus, bluffing strategies are not a concern.
In contrast, when signals are unconditionally independent, we show that truth-
ful betting is not a Weak Perfect Bayesian Equilibrium in the finite-stage model.
In the infinite-stage model, we show that there is no Weak Perfect Bayesian Equi-
librium that results in full information revelation in a finite number of trades. We
propose and analyze a discounted version of the LMSR that mitigates the strategic
hazard. In the discounted LMSR, we prove that under any WPBE strategy, the rela-
tive entropy between the market price and the optimal full-information price tends
to zero at an exponential rate. The rate of convergence can be bounded in terms of
the discounting parameter and a measure of complementarity in the information
setting.
1.2 Related Work
Theoretical work on price manipulation in financial markets [1, 7, 23] explains the
logic of manipulation and indicates that double auctions are not incentive compat-
ible. This literature has studied manipulation based on releasing false information
(perhaps through trades in other markets), as well as manipulation that only re-
quires strategic manipulation in a single market. The latter form of manipulation
is closely related to our study here. Allen and Gale [1] describe a model in which
4 Y. Chen et al.
a manipulative trader can make a deceptive trade in an early trading stage and then
profit in later stages, even though the other traders are aware of the possibility of
deception and act rationally. They use a stylized model of a multi-stage market; in
contrast, we seek to exactly model a market scoring rule model. Apart from other
advantages of detailed modeling, this allows us to construct simpler examples of
manipulative scenarios: Allen and Gale’s model [1] needs to assume traders with
different risk attitudes to get around no-trade results, which is rendered unneces-
sary by the inherent subsidy with a market scoring rule mechanism. Our model
requires only risk-neutral traders and exactly captures the market scoring rule pre-
diction markets. We refer readers to the paper by Chakraborty and Yilmaz [7] for
references to other research on manipulation in financial markets.
There are some experimental and empirical studies on price manipulation in
prediction markets using double auction mechanisms. The results are mixed, some
giving evidence for the success of price manipulation [19] and some showing the
robustness of prediction markets to price manipulation [6, 22, 29, 30].
Feigenbaum et al. [15] also study prediction markets in which the information
aggregation is sometimes slow, and sometimes fails altogether. In their setting, the
aggregation problems arise from a completely different source: the traders are non-
strategic but extracting individual traders’ information from the market price is dif-
ficult. Here, we study scenarios in which extracting information from prices would
be easy if traders were not strategic; the complexity arises solely from the use of
potentially non-myopic strategies. Nikolova and Sami [27] present an instance in
which myopic strategies are not optimal in an extensive-form game based on the
market, and suggest (but do not analyze) using a form of discounting to reduce
manipulative possibilities in a prediction market.
Axelrod et al. [2] also propose a form of discounting in an experimental
parimutuel market and show that it promotes early trades. Unlike the parimutuel
market, the market scoring rule has an inherent subsidy, so it was not obvious that
discounting would have strategic benefits in our setting as well.
B¨ orgers et al. [5] study when signals are substitutes and complements in a
general setting. Our equilibrium and convergence result suggests that prediction
markets are one domain where this distinction is of practical importance.
This paper is a synthesis and extension of two independent sets of results,
which were presented in preliminary form by Chen et al. [10] and Dimitrov and
Sami [13].
1.3 Structure of the Paper
This article is organized as follows: In section 2, we present a little background
information about market scoring rules. Section 3 details our formal model and
information structures. In Section 4 we investigate how predefined sequence of
play affects players’ expected profits in LMSR. In Section 5 we consider, when
a player can choose to play first or second, what is its equilibrium strategy in
a 2-player LMSR. Results of Sections 4 and 5 serve as building blocks for our
analysis in subsequent sections. Section 6 studies the equilibrium of a 2-player,
Gaming Prediction Markets: Equilibrium Strategies with a Market Maker 5
3-stage game. We generalize this result to arbitrary finite-player finite-stage games
in Section 7. We analyze infinite-stage games in Section 8. Section 9 proposes
a discounted LMSR and analyzes how the mechanism discourages non-truthful
behavior. Finally, we conclude in Section 10.
2 Background
Consider a discrete random variable X that has n mutually exclusive and ex-
haustive outcomes. Subsidizing a market to predict the likelihood of each out-
come, market scoring rules are known to guarantee that the market maker’s loss is
bounded.
2.1 Market Scoring Rules
Hanson [20, 21] shows how a proper scoring rule can be converted into a market
maker mechanism, called market scoring rules (MSR). The market maker uses a
proper scoring rule, S = {s
1
(r), . . . , s
n
(r)}, where r = r
1
, . . . , r
n
is a reported
probability estimate for the random variable X. Conceptually, every trader in the
market may change the current probability estimate to a new estimate of its choice
at any time as long as it agrees to pay the scoring rule payment associated with the
current probability estimate and receive the scoring rule payment associated with
the new estimate. If outcome i is realized, a trader that changes the probability
estimate from r to ˜r pays s
i
(r) and receives s
i
(˜r).
Since a proper scoring rule is incentive compatible for risk-neutral agents, if a
trader can only change the probability estimate once, this modified proper scoring
rule still incentivizes the trader to reveal its true probability estimate. However,
when traders can participate multiple times, they might have incentive to manipu-
late information and mislead other traders.
Because traders change the probability estimate in sequence, MSR can be
thought of as a sequential shared version of the scoring rule. The market maker
pays the last trader and receives payment from the first trader. An MSR market
can be equivalently implemented as a market maker offering n securities, each
corresponding to one outcome and paying $1 if the outcome is realized [20, 9].
Hence, changing the market probability of outcome i to some value r
i
is the same
as buying the security for outcome i until the market price of the security reaches
r
i
. Our analysis in this paper is facilitated by directly dealing with probabilities.
A popular MSR is the logarithmic market scoring rule (LMSR) where the
logarithmic scoring rule
s
i
(r) = b log(r
i
) (b > 0), (1)
is used. A trader’s expected profit in LMSR directly corresponds to the concept
of relative entropy in information theory. If a trader with probability ˜r moves the
market probability from r to ˜r, its expected profit (score) in LMSR is
S(r, ˜r) =
n
¸
i=1
˜ r
i
(s
i
(˜r) −s
i
(r)) = b
n
¸
i=1
˜ r
i
log
˜ r
i
r
i
= bD(˜r||r). (2)
6 Y. Chen et al.
D(p||q) is the relative entropy or Kullback Leibler distance between two proba-
bility mass functions p(x) and q(x) and is defined in [11] as
D(p||q) =
¸
x
p(x) log
p(x)
q(x)
.
D(p||q) is nonnegative and equals zero only when p = q. The maximum amount
a LMSR market maker can lose is b log n. Since b is a scaling parameter, without
loss of generality we assume that b = 1 in the rest of the paper.
2.2 Terminology
Truthful betting (TB) for a player in MSR is the strategy of immediately changing
the market probability to the player’s believed probability. In other words, it is the
strategy of always buying immediately when the price is too low and selling when
the price is too high. “Too low” and “too high” are determined by the player’s in-
formation. The price is too low when the current expected payoff is higher than the
price, and too high when current expected payoff is lower than the price. Truthful
betting fully reveals a player’s payoff-relevant information. Bluffing is the strategy
of betting contrary to one’s information in order to deceive future traders, with
the intent of capitalizing on their resultant misinformed trading. This paper inves-
tigates scenarios where traders with incomplete information have an incentive to
deviate from truthful betting.
1
3 LMSR in a Bayesian Framework
In this part, we introduce our game theoretic model of LMSR market in order to
capture the strategic behavior in LMSR when players have private information.
3.1 General Settings
We consider a single event that is the subject of our predictions. Ω = {Y, N} is
the state space of this event. The true state, ω ∈ Ω, is picked by nature according
to a prior p
0
= p
0
Y
, p
0
N
= Pr(ω = Y ), Pr(ω = N). A market, aiming at
predicting the true state ω, uses a LMSR market maker with initial probability
estimate r
0
= r
0
Y
, r
0
N
.
There are m risk neutral players in the market. Each player i gets a private
signal, c
i
, about the state of the world at the beginning of the market. C
i
is the
signal space of player i with |C
i
| = n
i
. The actual realization of the signal is
observed only by the player receiving the signal. The joint distribution of the true
state and players’ signals, P : Ω×C
1
×· · ·×C
m
→ [0, 1], is common knowledge.
1
With complete information, traders should reveal all information right away in MSR,
because the market degenerates to a race to capitalize on the shared information first.
Gaming Prediction Markets: Equilibrium Strategies with a Market Maker 7
Players trade sequentially, in one or more stages of trading, in the LMSR mar-
ket. Players are risk-neutrual Bayesian agents. If a player i is designated to trade
at some stage k in the sequence, it can condition its beliefs of the likelihood of an
outcome on its private signals, as well as the observed prices after the first k − 1
trades.
Up to this point, we have not made any assumptions about the joint distribution
P. It turns out that the strategic analysis of the market depends critically on inde-
pendence properties of players’ signals. We study two models of independence –
independence conditional on the true state, and unconditional independendence –
that are both natural in different settings. These are described below.
3.1.1 Games with Conditionally Independent (CI) Signals We start with a con-
crete example before formally introducing the model for conditionally independent
signals. Suppose the problem for a prediction market is to predict whether a batch
of product is manufactured with high quality materials or low quality materials. If
the product is manufactured with high quality materials, the probability for a prod-
uct to break in its first month of use is 0.01. If the product is manufactured with
low quality materials, the probability for a product to break in its first month of
use is 0.1. Some consumers who each bought a product have private observations
of whether their products break in the first month of use. The quality of the mate-
rials will be revealed by a test in the future. This is an example where consumers
have conditionally independent signals – conditional on the quality of the materi-
als, consumers’ observations are independent. Conditionally independent signals
are usually appropriate for modeling situations where the true state of the world
has been determined but is unknown, and signal realizations are influenced by the
true state of the world.
Formally, in games with conditionally independent signals, players’ signals are
assumed to be independent conditional on the state of the world, i.e., conditioned
on the eventual outcome of the event. In other words, for any two players i and j,
Pr(c
i
, c
j
|ω) = Pr(c
i
|ω) Pr(c
j
|ω) is always satisfied by P. This class can be inter-
preted as player i’s signal c
i
is independently drawn by nature according to some
conditional probability distribution p
(c
i
|Y )
if the true state is Y , and analogously
p
(c
i
|N)
if the true state is N.
In order to rule out degenerated cases, we further assume that conditionally
independent signals are informative and distinct. An informative signal means that
after observing the signal, a player’s posterior is different from its prior. Intuitively,
when observing a signal does not change a player’s belief, we can simply remove
the signal from the player’s signal space because it does not provide any infor-
mation. Formally, signals are informative if and only if Pr(c
i
= a|ω = Y ) =
Pr(c
i
= a|ω = N), ∀1 ≤ i ≤ m and ∀a ∈ C
i
. Player i has distinct signals
when its posterior probability is different after observing different signals.When
two signals give a player the same posterior, we can combine these two signals
into one because they provide same information. Formally, signals are distinct if
and only if Pr(ω = Y |c
i
= a) = Pr(ω = Y |c
i
= a

), ∀a, a

∈ C
i
, a = a

, and
∀i.
8 Y. Chen et al.
Lemma 1 shows that conditional independence implies unconditional depen-
dence.
Lemma 1 If players have informative signals that are conditionally independent,
their signals are unconditionally dependent.
Proof Suppose the contrary, signals of players i and j, c
i
and c
j
, are uncondition-
ally independent. Then, Pr(c
i
= a, c
j
= b) −Pr(c
i
= a) Pr(c
j
= b) = 0 must be
satisfied for all a ∈ C
i
and b ∈ C
j
. By conditional independence of signals,
Pr(c
i
= a, c
j
= b) −Pr(c
i
= a) Pr(c
j
= b)
=
¸
z
Pr(c
i
= a, c
j
= b|ω = z) Pr(ω = z)

¸
z
Pr(c
i
= a|ω = z) Pr(ω = z)
¸
z
Pr(c
j
= b|ω = z) Pr(ω = z)
=
¸
z
Pr(c
i
= a|ω = z) Pr(c
j
= b|ω = z) Pr(ω = z)

¸
z
Pr(c
i
= a|ω = z) Pr(ω = z)
¸
z
Pr(c
j
= b|ω = z) Pr(ω = z)
= Pr(Y ) Pr(N) (Pr(c
i
= a|Y ) −Pr(c
i
= a|N)) (Pr(c
j
= b|Y ) −Pr(c
j
= b|N)) .
The above expression equals 0 only when Pr(c
i
= a|Y ) − Pr(c
i
= a|N) = 0 or
Pr(c
j
= b|Y ) −Pr(c
j
= b|N) = 0 or both, which contradicts the informativeness
of signals. Hence, c
i
and c
j
are unconditionally dependent.
3.1.2 Games with (Unconditionally) Independent (I) Signals In some other situ-
ations, signal realizations are not caused by the state of the world, but instead, they
might stochastically influence the eventual outcome of the event. For instance,
in a political election prediction market, voters’ private information is their votes,
which can arguably be thought as independent of each other. The election outcome
– which candidate gets the majority of votes – is determined by all votes. This is
an example of a game with unconditionally independent signals. Formally, for any
two players i and j in games with independent signals, Pr(c
i
, c
j
) = Pr(c
i
) Pr(c
j
)
is always satisfied by P.
For games with independent signals, we primarily prove results about the lack
of truthful equilibria. Such results require us to show that there is a strict advan-
tage to deviate to an alternative strategy. In order to rule out degenerate cases in
which the inequalities are not strict, we will often invoke the following general
informativeness condition.
Definition 1 An instance of the prediction market with m players and joint dis-
tribution P satisfies the general informativeness condition if there is no vector of
signals for any m− 1 players that makes the m
th
player’s signals reveal no dis-
tinguishing information about the optimal probability. Formally, for m = 2, the
following property must be true: ∀i, i

∈ C
1
and ∀j, j

∈ C
2
such that i = i

, j =
j

: Pr(Y |c
1
= i, c
2
= j) = Pr(Y |c
1
= i, c
2
= j

) and Pr(Y |c
1
= i, c
2
=
j) = Pr(Y |c
1
= i

, c
2
= j). For m > 2, we must have ∀j, ∀i

= i, Pr(Y |i, j) =
Gaming Prediction Markets: Equilibrium Strategies with a Market Maker 9
Pr(Y |i

, j), where i, i

are two possible signals for any one player, and j is a vector
of signals for the other m−1 players.
By Lemma 1, we know that unconditional independence implies conditional
dependence. We note that the two information structures, conditional indepen-
dence and unconditional independence, that we discuss in this paper are mutually
exclusive, but not exhaustive. We do not consider the case when signals are both
conditionally dependent and unconditionally dependent.
3.2 Equilibrium Concepts
The prediction market model we have described is an extensive-form game be-
tween m players with common prior probabilities but asymmetric information
signals. Specifying a plausible play of the game involves specifying not just the
moves that players make for different information signals, but also the beliefs that
they have at each node of the game tree.
Informally, an assessment A
i
= (σ
i
, µ
i
) for a player i consists of a strategy
σ
i
and a belief system µ
i
. The strategy dictates what move the player will make at
each node in the game tree at which she has to move. We allow for strategies to
be (behaviorally) mixed; indeed, a bluffing equilibrium must involve mixed strate-
gies. To avoid technical measurability issues, we make the mild assumption that a
player’s strategy can randomize over only a finite set of actions at each node. The
belief system component of an assessment specifies what a player believes at each
node of the game tree. In our setting, the only relevant information a trader lacks
is the value of the other trader’s information signal. Thus, the belief at a node con-
sists of probabilities of other players’ signal realizations, contingent on reaching
the node.
An assessment profile (A
1
, ..., A
m
), consisting of an assessment for each
player, is a Weak Perfect Bayesian Equilibrium (WPBE) if and only if, for each
player, the strategies are sequentially rational given their beliefs and their beliefs
at any node that is reached with nonzero probability are consistent with updating
their prior beliefs using Bayes rule, given the strategies. This is a relatively weak
notion of equilibrium for this class of games. Frequently, the refined concepts of
Perfect Bayesian Equilibrium (PBE) or sequential equilibrium, which further re-
quire beliefs at nodes that are off the equilibrium path to be consistent, are used.
In this paper, when giving a specific equilibrium, we use the refined PBE concept.
When proving the nonexistence of truthful betting equilibrium and characterizing
the set of all equilibria, we use the WPBE concept, because the results thus hold
a fortiori for refinements of the WPBE concept. For a formal definition of the
equilibrium concepts, we refer the reader to the book by Mas-Colell et al. [25].
Given the strategy components of a WPBE profile, the belief systems of the
players are completely defined at every node on the equilibrium path (i.e., every
node that is reached with positive probability). In the remainder of this paper, we
will not consider players’ beliefs off the equilibrium path for WPBEs. We will
abuse notation slightly by simply referring to an “equilibrium strategy profile”,
10 Y. Chen et al.
leaving the beliefs implicit, for WPBEs. For PBEs, we will explicitly explain play-
ers’ beliefs in proofs.
4 Profits under Predefined Sequence of Play
In this section, we investigate how sequence of play affects players’ expected prof-
its in LMSR. We answer the question: when two players play myopically accord-
ing to a predefined sequence of play, are the players better off on expectation by
playing first or second?
In particular, we consider two players playing in sequence and define some
notations of strategy and expected profits to facilitate our analysis on how the
profit of the second player depends on the strategy employed by the first player.
Although we are studying 2-stage games here, we use them as a building block for
our results in subsequent sections, so the definitions are more general and the first
player’s strategy may not be truthful. In fact, our analysis apply to the first two
stages of any game.
Definition 2 Given a game and a sequence of play, a first-stage strategy σ
1
for
the first player is a specification, for each possible signal received by that player,
of a mixed strategy over moves.
Two specific first-stage strategies we will study are the truthful strategy (de-
noted σ
T
) and the null strategy (denoted σ
N
). The truthful strategy specifies that,
for each signal c
i
it receives, the first player changes market probability to its pos-
terior belief after seeing the signal. The null strategy specifies that, regardless of
its signal c
i
, the first player leaves the market probability unchanged, the same as
she found it at.
In a 2-stage game in which Alice plays first, followed by Bob, we use notation
π
B

1
) to denote Bob’s expected myopic optimal profit following a first-stage
strategy σ
1
by Alice, assuming that Bob knows the strategy σ
1
and can condition
on it. Likewise, in a game in which Alice moves after Bob, we use π
A

1
) to
denote Alice’s expected profit.
Note that π
B

N
) is equal to the expected myopic profit Bob would have had
if he had moved first, because under the null strategy σ
N
Alice does not move
the market. Also, π
B

T
) is the profit that Bob earns when Alice follows her
myopically optimal strategy.
We begin with the following simple result showing that the truthful strategy is
optimal for a player moving only once. This follows from the myopic optimality
of LMSR, but we include a proof for completeness.
Lemma 2 In a LMSR market, if stage t is player i’s last chance to play and
µ
i
is player i’s belief over actions of previous players, player i’s best response
at stage t is to play truthfully by changing the market probabilities to r
t
=
Pr(Y |c
i
, r
1
, ..., r
t−1
, µ
i
), Pr(N|c
i
, r
1
, ..., r
t−1
, µ
i
), where r
1
, ..., r
t−1
are the
market probability vectors before player i’s action.
Gaming Prediction Markets: Equilibrium Strategies with a Market Maker 11
Proof When a player has its last chance to play in LMSR, it is the same as the
player interacting with a logarithmic scoring rule. Because the logarithmic scor-
ing rule is strictly proper, player i’s expected utility is maximized by truthfully
reporting its posterior probability estimate given the information it has.
Next, we provide a lemma that is very useful for obtaining subsequent theo-
rems. Since the state of the world has only two exclusive and exhaustive outcomes,
Y and N. We use the term position to refer the market probability for outcome Y .
The market probability for outcome N is then uniquely defined because the prob-
abilities for the two outcomes sum to 1.
Lemma 3 Let σ
1
be a first-stage strategy for Alice that minimizes the expression
π
B
(σ) over all possible σ. Then, σ
1
satisfies the following consistency condition:
For any position x that Alice moves to under σ
1
, the probability of the outcome Y
occuring conditioned on σ
1
and the fact that Alice moved to x is exactly equal to
x.
Proof Suppose that σ
1
does not satisfy this condition. We construct a perturbed
first-stage strategy ˆ σ
1
, and show that π
B
(ˆ σ
1
) < π
B

1
), thus contradicting the
minimality of σ
1
. Such a ˆ σ
1
is easily constructed from any given σ
1
, as follows.
We start by setting ˆ σ
1
= σ
1
. Consider any one position x that Alice moves to with
positive probability under strategy σ
1
, and for which the consistency condition is
failed. Let q
x
= Pr(Y |x
A
= x, Alice following σ
1
). Then, whenever σ
1
dictates
that Alice move to x, we set ˆ σ
1
to dictate that Alice move to ˆ x = q
x
instead. We
repeat this perturbation for each x such that x = q
x
, thus resulting in a consistent
strategy ˆ σ
1
.
The actual position that Alice moves to can be thought of as a random variable
on a sample space that includes the randomly-distributed signals Alice receives as
well as the randomization Alice uses to play a mixed strategy. We use x
A
and ˆ x
A
to denote random variables that takes on values x and ˆ x respectively. Similarly,
the event ω is a random variable that takes on two values {Y, N}. Recall that the
moves are actually probability distributions. In order to use summation notation,
we define x
Y
= x, x
N
= 1 − x; and, likewise, ˆ x
Y
= ˆ x, ˆ x
N
= 1 − ˆ x. By
definition of ˆ x, ˆ x
z
= Pr(ω = z|x
A
= ˆ x, Alice following ˆ σ
1
). Note that any x has
a corresponding value of ˆ x; thus, we may write expressions like
¸
x
ˆ x in which
ˆ x is implicitly indexed by x. We use Pr(x, b, z) to represent Pr(x
A
= x, c
B
=
b, ω = z), and other notation is defined similarly.
The intuitive argument we use is as follows: Consider the situation in which
Alice is known to be following strategy σ
1
. The total profit of two consecutive
moves, in a market using the LMSR, is exactly the payoff of moving from the
starting point to the end point of the second move. Hence, for any given position
x that Alice leaves the market price at, we decompose Bob’s response into two
virtual steps. In the first virtual step, Bob moves the market to the corresponding
ˆ x. In the second step, Bob moves the market from ˆ x to his posterior probability
Pr(z|x, b). Bob’s actual profit is the sum of his profit from these two steps. Note
that the second step alone yields Bob at least π
B
(ˆ σ
1
) in expectation, because in
both cases the move begins at ˆ x, and because Bob can infer ˆ x from x. We further
argue that the first step yields a strictly positive expected profit, because for any
12 Y. Chen et al.
position x, the posterior probability of outcome Y is ˆ x
Y
. The first step thus moves
the market from a less accurate value to the most accurate possible value given
information x; this always yields a positive expected profit in the LMSR.
Formalizing this argument, we compare the payoffs π
B

1
) and π
B
(ˆ σ
1
) as
follows.
π
B

1
) −π
B
(ˆ σ
1
) =
¸
x,b,z
Pr(x, b, z) [log Pr(z|x, b) −log x
z
]

¸
x,b,z
Pr(x, b, z) [log Pr(z|ˆ x, b) −log ˆ x
z
]
=
¸
x,b,z
Pr(x, b, z) [log Pr(z|x, b) −log Pr(z|ˆ x, b)]
+
¸
x,b,z
Pr(x, b, z) [log ˆ x
z
−log x
z
]
=
¸
x,b
Pr(x, b)
¸
z
Pr(z|x, b) [log Pr(z|x, b) −log Pr(z|ˆ x, b)]
+
¸
x
Pr(x)
¸
z
Pr(z|x) [log ˆ x
z
−log x
z
]
=
¸
x,b
Pr(x, b)D(p
(ω|x
A
,c
B
)
||p
(ω|ˆ x
A
,c
B
)
)
+
¸
x
Pr(x)
¸
z
ˆ x
z
[log ˆ x
z
−log x
z
]
=
¸
x,b
Pr(x, b)D(p
(ω|x
A
,c
B
)
||p
(ω|ˆ x
A
,c
B
)
)
+
¸
x
Pr(x)D(p
(ˆ x
A
)
||p
(x
A
)
)
where p
(ω|x
A
,c
B
)
and p
(ω|ˆ x
A
,c
B
)
are the conditional distributions of ω, and
p
(ˆ x
A
)
and p
(x
A
)
are the probability distributions of ˆ x
A
and x
A
respectively.
D(p
(ω|x
A
,c
B
)
||p
(ω|ˆ x
A
,c
B
)
) and D(p
(ˆ x
A
)
||p
(x
A
)
) are relative entropy, which is
known to be nonnegative and strictly positive when the two distributions are
not the same. We assumed that σ
1
does not meet the consistency condition, and
thus, there is at least one x such that Pr(x) > 0 and ˆ x = x. Thus, we have
D(p
(ˆ x
A
)
||p
(x
A
)
) > 0. Hence, π
B

1
) > π
B
(ˆ σ
1
). This contradicts the minimality
assumption of σ
1
.
4.1 Profits in CI Games
When Alice and Bob have conditionally independent signals and Alice plays first,
we will prove that, for any strategy σ
1
that Alice chooses, if Bob is aware that Alice
is following strategy σ
1
, his expected payoff from following the myopic strategy
(after conditioning his beliefs on Alice’s actual move) is at least as much as he
could expect if Alice had chosen to play truthfully.
Gaming Prediction Markets: Equilibrium Strategies with a Market Maker 13
First, we show that in CI games, observing Alice’s posterior probabilities is
equally informative to Bob as observing Alice’s signal directly.
Lemma 4 When players have conditionally independent signals, if player i knows
player j’s posterior probabilities Pr(Y |c
j
), Pr(N|c
j
), player i can infer the pos-
terior probabilities conditional on both signals. More specifically,
Pr(ω|c
i
, c
j
) =
Pr(c
i
|ω) Pr(ω|c
j
)
Pr(c
i
|Y ) Pr(Y |c
j
) + Pr(c
i
|N) Pr(N|c
j
)
,
where ω ∈ {Y, N}.
Proof Using Bayes rule, we have
Pr(ω|c
i
, c
j
) =
Pr(ω, c
i
|c
j
)
Pr(c
i
|c
j
)
=
Pr(c
i
|c
j
, ω) Pr(ω|c
j
)
Pr(c
i
|c
j
, Y ) Pr(Y |c
j
) + Pr(c
i
|c
j
, N) Pr(N|c
j
)
=
Pr(c
i
|ω) Pr(ω|c
j
)
Pr(c
i
|Y ) Pr(Y |c
j
) + Pr(c
i
|N) Pr(N|c
j
)
.
The third equality comes from the CI condition.
We can now prove our result on Bob’s myopic profit in the CI setting.
Theorem 1 For any CI game G with Alice playing first and Bob playing second,
and any σ
1
, π
B

1
) ≥ π
B

T
). When signals are informative and distinct, the
equality holds if and only if σ
1
= σ
T
.
Proof We use an information-theoretic argument to prove this result. We can view
Alice’s signal c
A
, Alice’s actual move x
A
(for outcome Y ) , Bob’s signal c
B
, the
event ω as random variables. We then show that Bob’s expected payoff can be
characterized as the entropy of c
B
conditioned on observing x
A
, plus another term
that does not depend on Alice. The more information that x
A
reveals about c
B
,
the lower the conditional entropy H(c
B
|x
A
), and thus, the lower Bob’s expected
profit. However, note that x
A
is a (perhaps randomized) function of Alice’s signal
c
A
alone. A fundamental result from information theory states that x
A
can reveal
no more information about c
B
than c
A
can, and thus, Bob’s profit is minimized
when x
A
reveals as much information as c
A
. By Lemma 4, the truthful strategy
for Alice reveals all information that is in c
A
, and thus, attains the minimum.
We now formalize this argument, retaining the notation above. By Lemma 3,
we can assume that σ
1
is consistent, without loss of generality. Let us analyze
Bob’s expected payoff when Alice follows this strategy σ
1
. The unit of our analysis
is a particular realization of a combination c
A
= a, x
A
= x, c
B
= b, ω = z. Bob
will move to a position y(x, σ
1
, b) = Pr(Y |x
A
= x, Alice following σ
1
, c
B
= b).
For conciseness, we drop σ
1
from the notation, and write y(x, b).
As before use x
z
and y
z
to denote the probability of ω = z inferred from
positions x and y respectively. By definition of the LMSR, we have:
π
B

1
) =
¸
x,b,z
Pr(x, b, z) [log y
z
(x, b) −log x
z
]
14 Y. Chen et al.
Next, we note that y
z
(x, b) = Pr(ω = z|x, b). Further, it is immediate that
¸
b
Pr(b|x, z) = 1. Expanding, we have
π
B

1
) =
¸
x,b
Pr(x, b)
¸
z
Pr(z|x, b) log y
z
(x, b)

¸
x
Pr(x)
¸
z
Pr(z|x)
¸
b
Pr(b|x, z) log x
z
=
¸
x,b
Pr(x, b)
¸
¸
z
Pr(z|x, b) log Pr(z|x, b)
¸

¸
x
Pr(x)
¸
¸
z
Pr(z|x) log Pr(z|x)
¸
= −H(ω|x
A
, c
B
) +H(ω|x
A
)
where we have identified the terms with the standard definition of conditional en-
tropy of random variables [11, pg. 17].
Using the relation H(X, Y ) = H(X) +H(Y |X), we can write
π
B

1
) = −H(ω, x
A
, c
B
) +H(x
A
, c
B
) +H(ω, x
A
) −H(x
A
)
= [H(x
A
, c
B
) −H(x
A
)] −[H(ω, x
A
, c
B
) −H(ω, x
A
)]
= H(c
B
|x
A
) −H(c
B
|ω, x
A
)
= H(c
B
|x
A
) −H(c
B
|ω)
The last transformation H(c
B
|ω, x
A
) = H(c
B
|ω) follows because c
B
is condi-
tionally independent of x
A
conditioned on ω, and thus, knowledge of x
A
does not
alter its conditional distribution or its conditional entropy.
The second term is clearly independent of σ
1
. For the first term, we note that
H(c
B
|c
A
) ≤ H(c
B
|x
A
) because x
A
is a function of c
A
[11, pg. 35]. More specif-
ically,
H(c
B
|x
A
) = −
¸
x
Pr(x)
¸
¸
b
Pr(b|x) log Pr(b|x)
¸
= −
¸
x
Pr(x)
¸
¸
b

¸
a
Pr(a|x) Pr(b|a)

log

¸
a
Pr(a|x) Pr(b|a)
¸
≥ −
¸
x
Pr(x)
¸
b
¸
a
[Pr(a|x) Pr(b|a) log Pr(b|a)]
= −
¸
x
Pr(x)
¸
a
¸
Pr(a|x)
¸
b
[Pr(b|a) log Pr(b|a)]
¸
= −
¸
a
Pr(a)
¸
¸
b
Pr(b|a) log Pr(b|a)
¸
= H(c
B
|c
A
).
Gaming Prediction Markets: Equilibrium Strategies with a Market Maker 15
The inequality follows from the strict convexity of the function xlog x. The
equality holds only when the distribution of c
B
|a is the same as the distribu-
tion of c
B
|a

for all a, a

∈ C
A
, or for any x there exists some a such that
Pr(c
A
= a|x
A
= x) = 1. If c
B
|a and c
B
|a

have the same distribution for all
a and a

, c
A
and c
B
are unconditionally independent which contradicts with the
CI condition by Lemma 1. Thus, the equality holds only when σ
1
is a strategy such
that for any x there exists some a that satisfies Pr(c
A
= a|x
A
= x) = 1.
The truthful strategy σ
T
satisfies the above condition. Given any signal a that
Alice might see, let x
a
be the position that Alice would move to if she followed the
truthful strategy, and let x
T
A
denote the corresponding random variable. Then x
a
=
Pr(Y |c
A
= a). Because signals are distinct, the value of x
a
uniquely corresponds
to signal a. Hence, the condition is satisfied with Pr(c
A
= a|x
T
A
= x
a
) = 1.
All that remains to be shown is that there is no other strategy σ
1
= σ
T
that
satisfies the condition. For some strategy σ
1
, let X
A
denote the value space of x
A
.
Suppose the contrary, for any x ∈ X
A
there exists a
x
such that Pr(c
A
= a
x
|x
A
=
x) = 1. We first show that when x = x

and x, x

∈ X
A
, it must be that a
x
= a
x
.
According to Lemma 3, Pr(Y |x
A
= x, Alice follows σ
1
) = x, which gives
¸
a

Pr(Y |c
A
= a

) Pr(c
A
= a

|x
A
= x) = x =⇒ Pr(Y |c
A
= a
x
) = x.
This means that Pr(Y |c
A
= a
x
) = Pr(Y |c
A
= a
x
) when x = x

. Hence,
a
x
= a
x
because signals are distinct. Next, we show that when c
A
= a, it must
be the case that x
A
= Pr(Y |c
A
= a) with probability 1 according to strategy
σ
1
. If with some positive probability x
A
= x

= Pr(Y |c
A
= a), then, because
Pr(c
A
= a
x
|x
A
= x

) = 1, c
A
= a
x
= a, and contradiction arises. Thus, for all
a ∈ C
A
, when c
A
= a, Alice moves to Pr(Y |c
A
= a) with probability 1 under
σ
1
. Hence, strategy σ
1
is the truthful strategy σ
T
.
The following corollary immediately follows fromTheorem1 and answers that
question we proposed at the beginning of the section for CI games.
Corollary 1 In CI games, when two players play myopically according to a pre-
defined sequence of play, it is better off for a player to play first than second.
Proof We compare Bob’s expected profit in Alice-Bob and Bob-Alice cases when
both players play myopically. Bob’s expected profit in Alice-Bob case equals
π
B

T
), while his expected profit in Bob-Alice cases equals π
B

N
). According
to Theorem 1, π
B

N
) > π
B

T
). Hence, Bob is better off when the Bob-Alice
sequence is used.
4.2 Profits in I Games
When Alice and Bob have unconditionally independent signals and Alice plays
first, we will prove that, for any strategy σ
1
that Alice chooses, if Bob is aware
that Alice is following strategy σ
1
, his expected payoff from following the myopic
strategy (after conditioning his beliefs on Alices actual move) is at least as much
as he could expect if Alice had chosen to play the null strategy σ
N
.
16 Y. Chen et al.
Theorem 2 For any I game G with Alice playing first and Bob playing second and
any first-stage strategy σ
1
, if the market starts with the prior probabilities for the
two outcomes, i.e. r
0
= Pr(ω = Y ), Pr(ω = N), we have π
B

1
) ≥ π
B

N
).
The equality holds if and only if σ
1
= σ
N
.
Proof Assume that there is a strategy σ
1
for Alice that minimizes π
B

1
) over all
possible first-stage strategies. According to Lemma 3, we can assume that σ
1
is
consistent, without loss of generality. We will show that σ
1
must involve Alice not
moving the market probability at all.
Let x
A
be the random variable of Alice’s move for outcome Y under strategy
σ
1
. We first argue that x
A
is deterministic, i.e. it can only have a single value.
Suppose the contrary, then σ
1
would have support over a set of points: at least
two points E, F, and perhaps a set of other points R. In this case, we show that
we can construct a strategy σ

1
such that π
B

1
) < π
B

1
) by “mixing” point
E and another point H. Define u
E
i
and u
F
i
as the probability (under σ
1
) that
Alice has signal c
A
= a
i
and sets x
A
= E and x
A
= F respectively. Let p
E
be
the probability that Alice plays E and similarly p
F
be the probability that Alice
plays F. Without loss of generality let p
F
< p
E
. Define α
i
= Pr(c
A
= a
i
|x
A
=
E) =
u
E
i
p
E
and β
i
= Pr(c
A
= a
i
|x
A
= F) =
u
F
i
p
F
. With these definitions we
can define the myopic move of Bob given that market is at E and c
B
= b
j
as
r
α
j
=
¸
i
α
i
Pr(Y |c
A
= a
i
, c
B
= b
j
). Similarly r
β
j
is defined as the myopic
move of Bob given the market is at F. We use r
α
j
and r
β
j
to denote the probability
vectors for two outcomes r
α
j
, 1 −r
α
j
and r
β
j
, 1 −r
β
j
respectively.
Now, let H = (E + F)/2 be the midpoint of E and F, and consider a new
strategy σ

1
over points E, H, and the same set of remaining points R. Under σ

1
,
the probability that Alice has signal c
A
= a
i
and sets x
A
= E is
p
E
−p
F
p
E
u
E
i
, and
the probability that Alice has signal c
A
= a
i
and sets x
A
= H equals
p
F
p
E
u
E
i
+u
F
i
.
Hence, Alice mixes over E and H with probability p
E
−p
F
and 2p
F
respectively.
As before we can define γ
i
= Pr(c
A
= a
i
|H) =
p
F
p
E
u
E
i
+u
F
i
2p
F
=
1
2
u
E
i
p
E
+
1
2
u
F
i
p
F
=
α
i

i
2
. We can now define the myopic move of Bob given the market is at H
and c
B
= b
j
as r
γ
j
=
¸
i
γ
i
Pr(Y |c
A
= a
i
, c
B
= b
j
). Similarly, we use r
γ
j
to
represent the vector for two outcomes.
We now characterize π
B

1
) as follows, writing p
E
as (p
E
− p
F
) + p
F
to
facilitate comparison with π
B

1
).
π
B

1
) = (p
E
−p
F
)[
¸
j
Pr(c
B
= b
j
)D(r
α
j
||
¸
j
Pr(c
B
= b
j
)r
α
j
)]
+p
F
[
¸
j
Pr(c
B
= b
j
)D(r
α
j
||
¸
j
Pr(c
B
= b
j
)r
α
j
)]
+p
F
[
¸
j
Pr(c
B
= b
j
)D(r
β
j
||
¸
j
Pr(c
B
= b
j
)r
β
j
)]
+ remaining profit over R
Gaming Prediction Markets: Equilibrium Strategies with a Market Maker 17
We also characterize π
B

1
) as:
π
B

1
) = (p
E
−p
F
)[
¸
j
Pr(c
B
= b
j
)D(r
α
j
||
¸
j
Pr(c
B
= b
j
)r
α
j
)
+2p
F
[
¸
j
Pr(c
B
= b
j
)D(r
γ
j
||
¸
j
Pr(c
B
= b
j
)r
γ
j
)
+ remaining profit over R
From the definitions of the myopic moves given the market states, note that r
γ
j
=
r
α
j
+r
β
j
2
. This means that π
B

1
) can be bounded as :
π
B

1
) = (p
E
−p
F
)[
¸
j
Pr(c
B
= b
j
)D(r
α
j
||
¸
j
Pr(c
B
= b
j
)r
α
j
)
+2p
F
[
¸
j
Pr(c
B
= b
j
)D(r
γ
j
||
¸
j
Pr(c
B
= b
j
)r
γ
j
)
+ remaining profit over R
= (p
E
−p
F
)[
¸
j
Pr(c
B
= b
j
)D(r
α
j
||
¸
j
Pr(c
B
= b
j
)r
α
j
)
+2p
F
[
¸
j
Pr(c
B
= b
j
)D(
r
α
j
+r
β
j
2
||
¸
j
Pr(c
B
= b
j
)
r
α
j
+r
β
j
2
)
+ remaining profit over R
< (p
E
−p
F
)[
¸
j
Pr(c
B
= b
j
)D(r
α
j
||
¸
j
Pr(c
B
= b
j
)r
α
j
)]
+p
F
[
¸
j
Pr(c
B
= b
j
)D(r
α
j
||
¸
j
Pr(c
B
= b
j
)r
α
j
)]
+p
F
[
¸
j
Pr(c
B
= b
j
)D(r
β
j
||
¸
j
Pr(c
B
= b
j
)r
β
j
)]
+ remaining profit over R
= π
B

1
)
The last inequality follows from the strict convexity of relative entropy under the
general informativeness condition.
Therefore, for any strategy σ
1
with two or more points in its support, there
always exists a strategy σ

1
such that π
B

1
) < π
B

1
). This means that for any
strategy of Alice that minimized π
B

1
), the strategy must have only one point
in its support. Thus, the strategy does not reveal any information to Bob. Suppose
that the point in the support, x, is such that x = Pr(ω = Y ). This again contradicts
the fact that σ
1
minimizes π
B

1
), as Bob will always make a positive payoff in
expectation if he moves fromx to Pr(ω = Y ). Thus, he would have a larger payoff
overall if Alice left the market at x instead of Pr(ω = Y ). Therefore the strategy
that minimizes the expected payoff of Bob is for Alice to report Pr(ω = Y ).
However, because the market starts with Pr(ω = Y ), it is equivalent to her not
trading at all in the first stage. Therefore we have shown that π
B

1
) ≥ π
B

N
).
Only when σ
1
= σ
N
the equality holds.
The following corollary immediately follows from Theorem 2 and answers the
question we proposed at the beginning of the section for I games.
18 Y. Chen et al.
Corollary 2 In I games that start with the prior probabilities for the two out-
comes, i.e. r
0
= Pr(ω = Y ), Pr(ω = N), when two players play myopically
according to a predefined sequence of play, it is better off for a player to play
second than first.
Proof We compare Bob’s expected profit in Alice-Bob and Bob-Alice cases when
both players play myopically. Bob’s expected profit in Alice-Bob case equals
π
B

T
), while his expected profit in Bob-Alice cases equals π
B

N
). According
to Theorem 2, π
B

T
) ≥ π
B

N
). Hence, Bob is better off when the Alice-Bob
sequence is used.
5 Sequence Selection Game
In this section, we examine the following strategic question: Suppose that each
player will get to report only once. If a player has a choice between playing first
and playing second after observing its signal, which one should it choose?
To answer the question, we define a simple 2-player sequence selection game.
Suppose that Alice and Bob are the only players in the market. Alice gets a signal
c
A
. Similarly, Bob gets a signal c
B
. After getting their signals, Alice and Bob play
the sequence selection game as follows. In the first stage, Alice chooses who—
herself or Bob—plays first. The selected player then changes the market probabil-
ities as they see fit in the second stage. In the third stage, the other player gets the
chance to change the market probabilities. Then, the market closes and the true
state is revealed.
In the rest of this section, we provide equilibria of the sequence selection game
under CI setting and I setting respectively.
5.1 Sequence Selection Equilibrium of CI Games
Consider the sequence selection game, and assume that Alice and Bob have condi-
tionally independent signals. The following theorem gives a PBE of the sequence
selection game when the CI condition is satisfied.
Theorem 3 When Alice and Bob have conditionally independent signals in LMSR,
at a PBE of the sequence selection game
– Alice selects herself to be the first player in the first stage;
– Alice changes the market probability to Pr(Y |c
A
), Pr(N|c
A
) in the second
stage;
– Bob changes the market probability to Pr(Y |c
A
, c
B
), Pr(N|c
A
, c
B
) in the
third stage.
Proof Since we use the PBE concept, we first describe Bob’s belief µ
B
for the off-
equilibrium paths. We denote x = x
1
, ..., x
n
A
as the vector of Alice’s possible
posteriors for the outcome Y . That is, x
i
= Pr(Y |c
A
= a
i
), where a
i
is the i-th
element of C
A
. Without loss of generality, assume that x
i
< x
j
iff i < j. Off
the equilibrium path, when Alice selects Bob to be the first player and the initial
market probability is r
0
, Bob’s belief µ
B
gives
Gaming Prediction Markets: Equilibrium Strategies with a Market Maker 19
– If x
1
≤ r
0
Y
≤ x
n
A
, Pr(Y |r
0
, µ
B
) = r
0
Y
.
– If 0 ≤ r
0
Y
< x
1
, Pr(Y |r
0
, µ
B
) = x
1
.
– If x
n
A
< r
0
Y
≤ 1, Pr(Y |r
0
, µ
B
) = x
n
A
.
Bob’s belief µ
B
is consistent with Alice’s off-equlibrium mixed strategy:
– If c
A
= a
i
and x
i
= r
0
Y
, select Bob as the first player.
– If c
A
= a
i
or c
A
= a
i+1
and x
i
< r
0
Y
< x
i+1
, mix between selecting herself
as the first player and selecting Bob as the first player. When c
A
= a
i
, with
probability α select Bob as the first player and (1 − α) select herself as the
first player. When c
A
= a
i+1
, with probability β select Bob as the first player
and (1 −β) select herself as the first player. The mixing probabilities α and β
satisfy that Pr(Y |select Bob as the first player) = r
0
Y
.
– If c
A
= a
1
and 0 ≤ r
0
Y
< x
1
, select Bob as the first player.
– If c
A
= a
n
A
and x
n
A
< r
0
Y
≤ 1, select Bob as the first player.
– For all other cases, select herself as the first player.
Each player has only one chance to change the market probabilities. Hence, by
Lemma 2, both of them will truthfully reveal all information that they have given
their beliefs when it’s their turn to play. If Alice is the first to play, she will change
the market probabilities to Pr(Y |c
A
), Pr(N|c
A
) in the second stage. Bob, be-
lieving that prices in the second stage are Alice’s posteriors, can calculate his pos-
teriors based on both Alice’s and his own signals. Bob will further changes the
market probabilities to Pr(Y |c
A
, c
B
), Pr(N|c
A
, c
B
) in the third stage. On the
contrary, if Bob is selected as the first player, he will change the market probabili-
ties to Pr(Y |r
0
, c
B
, µ
B
), Pr(N|r
0
, c
B
, µ
B
) in the second stage. Let ˆ a
r
0 denote
a fictitious signal realization of Alice that satisfies Pr(Y |ˆ a
r
0), Pr(N|ˆ a
r
0) = r
0
.
ˆ a
r
0 does not necessarily belong to Alice’s signal space C
A
. According to Lemma
4, Pr(Y |r
0
, c
B
, µ
B
), Pr(N|r
0
, c
B
, µ
B
) equals one of the following
1. Pr(Y |ˆ a
r
0, c
B
), Pr(N|ˆ a
r
0, c
B
) when x
1
≤ r
0
Y
≤ x
n
A
.
2. Pr(Y |a
n
A
, c
B
), Pr(N|a
n
A
, c
B
) when x
n
A
< r
0
Y
≤ 1.
3. Pr(Y |a
1
, c
B
), Pr(N|a
1
, c
B
) when 0 ≤ r
0
Y
< x
1
.
To make her sequence selection in the first stage, Alice essentially compares
her expected utilities conditional on her own signal in the Alice-Bob and Bob-
Alice subgames.
Without loss of generality, suppose Alice has signal a
k
. Let U
I
A
denote Alice’s
expected profit conditioned on her signal when the Alice-Bob subgame is picked.
U
II
A
denotes Alice’s expected profit conditioned on her signal when the Bob-Alice
subgame is picked. Then, for case 1,
U
I
A
=
1
Pr(c
A
= a
k
)
¸
l,z
Pr(a
k
, b
l
, z) log
Pr(ω = z|c
A
= a
k
)
Pr(ω = z|ˆ a
r
0)
, and (3)
U
II
A
=
1
Pr(c
A
= a
k
)
¸
l,z
Pr(a
k
, b
l
, z) log
Pr(ω = z|c
A
= a
k
, c
B
= b
l
)
Pr(ω = z|ˆ a
r
0, c
B
= b
l
)
, (4)
where b
l
represents the l-th signal element in C
B
, and Pr(a
k
, b
l
, z) represents the
joint probability of c
A
= a
k
, c
B
= b
l
, and ω = z. For conciseness, we use Pr(a
k
)
20 Y. Chen et al.
to represent Pr(c
A
= a
k
) and similarly for other terms. The difference in expected
profits for Alice in the two subgames is:
U
I
A
−U
II
A
=
1
Pr(a
k
)
¸
l,z
Pr(a
k
, b
l
, z) log
Pr(z|a
k
) Pr(z|ˆ a
r
0, b
l
)
Pr(z|ˆ a
r
0) Pr(z|a
k
, b
l
)
(5)
=
1
Pr(a
k
)
¸
l,z
Pr(a
k
, b
l
, z) log
Pr(a
k
, b
l
) Pr(ˆ a
r
0)
Pr(ˆ a
r
0, b
l
) Pr(a
k
)
=
¸
l
Pr(b
l
|a
k
) log
Pr(b
l
|a
k
)
Pr(b
l
|ˆ a
r
0)
= D

p
(c
B
|a
k
)
||p
(c
B
|ˆ a
r
0)

, (6)
where p
(c
B
|a
k
)
is the probability distribution of Bob’s signal conditional on sig-
nal a
k
and p
(c
B
|ˆ a
r
0)
is the probability distribution of Bob’s signal conditional
on the fictitious signal ˆ a
r
0. The second equality comes from Bayes rule and
the conditional independence of signals. D(p||q) ≥ 0. The equality holds only
when distributions p and q are the same. We thus have U
I
A
− U
II
A
≥ 0. When
Pr(Y |a
k
), Pr(N|a
k
) = r
0
, p
(c
B
|a
k
)
is different from p
(c
B
|ˆ a
r
0)
and hence
U
I
A
−U
II
A
is strictly greater than 0.
Since the market only has two exclusive outcomes, cases 2 and 3 are symmet-
ric. We only need to prove one of them. Without loss of generality, consider case
2 where x
n
A
< r
0
Y
≤ 1, Alice’s expected profit conditioned on her signal for the
Bob-Alice subgame becomes
¯
U
II
A
=
1
Pr(c
A
= a
k
)
¸
l,z
Pr(a
k
, b
l
, z) log
Pr(ω = z|c
A
= a
k
, c
B
= b
l
)
Pr(ω = z|c
A
= a
n
A
, c
B
= b
l
)
, (7)
while U
I
A
remains the same as in (3). We obtain
U
I
A

¯
U
II
A
=
1
Pr(a
k
)
¸
l,z
Pr(a
k
, b
l
, z) log
Pr(z|a
k
)
Pr(z|a
n
A
)
+
¸
z
Pr(z|a
k
) log
Pr(z|a
n
A
)
r
0
z

¯
U
II
A
= D

p
(c
B
|a
k
)
||p
(c
B
|a
n
A
)

+
¸
z
Pr(z|a
k
) log
Pr(z|a
n
A
)
r
0
z
.
The second term in the above expression is positive because r
0
Y
> Pr(Y |a
n
A
) ≥
Pr(Y |a
k
). Hence, U
I
A

¯
U
II
A
> 0. Alice does not want to deviate from selecting
herself as the first player.
Corollary 1 in Section 4 indicates that if myopic players follow predefined
sequence of play then on expectation a player is better off by playing first than
second in CI games. Theorem 3 further strengthens this result and states that if a
strategic player have a choice, at a PBE the player will always choose to play first
in CI games no matter what signal it gets.
Gaming Prediction Markets: Equilibrium Strategies with a Market Maker 21
5.2 Sequence Selection Equilibrium of I games
In this part, we consider the alternative model in which players have independent
signals. The following theorem gives a PBE of the sequence selection game when
the I condition is satisfied.
Theorem 4 If the market starts with the prior probabilities for the two outcomes,
i.e. r
0
= Pr(ω = Y ), Pr(ω = N), and Alice and Bob have independent signals,
at a PBE of the sequence selection game
– Alice in the first stage selects Bob to be the first player;
– Bob changes the market probability to Pr(Y |c
B
), Pr(N|c
B
) in the second
stage;
– Alice changes the market probability to Pr(Y |c
A
, c
B
), Pr(N|c
A
, c
B
) in the
third stage.
Proof If Alice always selects Bob as the first player, this action does not reveal
any information to Bob. Knowing Alice’s strategy, according to Lemma 2, Bob’s
best responds in the second stage is to truthfully report the posterior probability
conditioned on his own signal and Alice will report truthfully conditioned on both
signals in the third stage. Thus, we need to prove that Alice does not want to deviate
from selecting herself as the first player. Without loss of generality, assume Alice
has signal a
k
.
If Alice selects herself as the first player, her best response in the second stage
is to report truthfully. Hence, the expected profit conditioned on her signal in this
case, denoted as U
I
A
, is
U
I
A
=
¸
z
Pr(ω = z|c
a
= a
k
) log
Pr(ω = z|c
a
= a
k
)
Pr(ω = z)
= D(p
(ω|a
k
)
||p
(ω)
) (8)
If Alice selects Bob as the first player, her expected profit conditioned on her sig-
nal, denoted as U
II
A
is
U
II
A
=
¸
l,z
Pr(c
b
= b
l
|a
k
) Pr(ω = z|a
k
, b
l
) log
Pr(ω = z|a
k
, b
l
)
Pr(ω = z|b
l
)
=
¸
l
Pr(c
b
= b
l
|a
k
)D(p
(ω|a
k
,b
l
)
||p
(ω|b
l
)
) (9)
Because signals are independent, we have Pr(b
l
|a
k
) = Pr(b
l
). Hence,
Pr(z|a
k
) =
¸
l
Pr(b
l
|a
k
) Pr(z|a
k
, b
l
) =
¸
l
Pr(b
l
) Pr(z|a
k
, b
l
). By Bayes rule,
Pr(z) =
¸
l
Pr(b
l
) Pr(z|b
l
). The difference in expected profits for Alice is:
U
I
A
−U
II
A
= D(p
(ω|a
k
)
||p
(ω)
) −
¸
l
Pr(b
l
)D(p
(ω|a
k
,b
l
)
||p
(ω|b
l
)
)
= D(
¸
l
Pr(b
l
)p
(ω|a
k
,b
l
)
||
¸
l
Pr(b
l
)p
(ω|b
l
)
) −
¸
l
Pr(b
l
)D(p
(ω|a
k
,b
l
)
||p
(ω|b
l
)
)
≤ 0.
22 Y. Chen et al.
The inequality follows from the convexity of relative entropy [11]. Thus, Alice
does not want to deviate from selecting Bob as the first player.
The strategies and beliefs off the equilibrium path are straightforward. If Alice
selects herself to be the first players. In the second stage, her best response is to
report her true posterior. Bob believes that Alice reports truthfully.
Corollary 2 in Section 4 indicates that if myopic players follow predefined se-
quence of play then on expectation a player is better off by playing second than
first in I games. Theorem 4 further strengthens this result and states that if a strate-
gic player have a choice, at a PBE the player will always choose to play second in
I games no matter what signal it gets.
6 The Alice-Bob-Alice Game
We now consider a 3-stage Alice-Bob-Alice game, where Alice plays in the first
and third stages and Bob plays in the second stage. Alice may change the market
probabilities however she wants in the first stage. Observing Alice’s action, Bob
may change the probabilities in the second stage. Alice can take another action in
the third stage. Then, the market closes and the true state is revealed.
Let r
t
= r
t
Y
, r
t
N
be the market probabilities in stage t. Lemma 5 character-
izes the equilibrium strategy of Alice in the third stage.
Lemma 5 In a 3-stage Alice-Bob-Alice game in LMSR, at a WPBE Alice changes
the market probabilities to r
3
= r
3
Y
, r
3
N
= Pr(Y |a
k
, b
l
), Pr(N|a
k
, b
l
in the
third stage, when Alice has signal a
k
and Bob has signal b
l
.
Proof This is proved by applying Lemma 2 to both Bob and Alice. At a WPBE,
beliefs are consistent with strategies. Alice and Bob act as if they knoweach other’s
strategy. Since Bob only gets one chance to play, according to Lemma 2 Bob plays
truthfully and fully reveals his information. Because the third stage is Alice’s last
chance to change the probabilities, Alice behaves truthfully and fully reveals her
information, including her own signal and Bob’s signal inferred from Bob’s action
in the second stage.
6.1 Truthful Equilibrium of CI Games
We study the PBE of the game when Alice and Bob have conditionally independent
signals. Built on our equilibrium result for the sequence selection game, Theorem
5 describes a PBE of the Alice-Bob-Alice game.
Theorem 5 When Alice and Bob have conditionally independent signals in LMSR,
at a PBE of the 3-stage Alice-Bob-Alice game
– Alice changes the market probability to r
1
= Pr(Y |c
A
), Pr(N|c
A
) in the
first stage;
– Bob in the second stage changes the market probability to r
2
=
Pr(Y |c
A
, c
B
), Pr(N|c
A
, c
B
);
Gaming Prediction Markets: Equilibrium Strategies with a Market Maker 23
– Alice does nothing in the third stage.
Proof Bob’s belief µ
B
is similar to that of the sequence selection game as de-
scribed in the proof for Theorem 3. We denote x = x
1
, ..., x
n
A
as the vector
of Alice’s possible posteriors for the outcome Y . That is, x
i
= Pr(Y |c
A
= a
i
),
where a
i
is the i-th element of C
A
. Without loss of generality, assume that x
i
< x
j
iff i < j. When Alice changes market probability to r
1
in the first stage, Bob’s
belief µ
B
gives
– If x
1
≤ r
1
Y
≤ x
n
A
, Pr(Y |r
1
, µ
B
) = r
1
Y
.
– If 0 ≤ r
1
Y
< x
1
, Pr(Y |r
1
, µ
B
) = x
1
.
– If x
n
A
< r
1
Y
≤ 1, Pr(Y |r
1
, µ
B
) = x
n
A
.
When observing a r
1
Y
= y = x
i
for all 1 ≤ i ≤ n
A
, Bob’s belief µ
B
is consistent
with Alice’s off-equlibrium mixed strategy:
– If c
A
= a
i
or c
A
= a
i+1
and x
i
< y < x
i+1
, mix between truthful betting
and reporting r
1
Y
= y. When c
A
= a
i
, with probability α change market
probability to r
1
Y
= y and (1−α) play truthful betting. When c
A
= a
i+1
, with
probability β change market probability to r
1
Y
= y and (1 − β) play truthful
betting. The mixing probabilities α and β satisfy that Pr(Y |r
1
Y
= y) = y.
– If c
A
= a
1
and 0 ≤ y < x
1
, change market probability to r
1
Y
= y.
– If c
A
= a
n
A
and x
n
A
< y ≤ 1, change market probability to r
1
Y
= y.
– For all other cases, report truthfully.
By Lemma 2, Bob does not want to deviate from truthful betting in the second
stage given that Alice truthfully reports her posteriors in the first stage.
We show that Alice does not want to deviate by changing market probabilities
to r
1
= Pr(Y |c
A
), Pr(N|c
A
). Without loss of generality, assume that Alice has
signal c
A
= a
k
. Consider the two cases:
1. When Alice does not deviate: Alice changes market probabilities to her true
posteriors Pr(Y |a
k
), Pr(N|a
k
) in the first stage; Bob changes the probabil-
ities to Pr(Y |a
k
, c
B
), Pr(N|a
k
, c
B
) in the second stage; Alice does nothing
in the third stage.
2. When Alice deviates: Alice changes market probabilities to r
1
that is dif-
ferent from Pr(Y |a
k
), Pr(N|a
k
); Bob changes market probabilities to
Pr(Y |r
1
, c
B
, µ
B
), Pr(N|r
1
, c
B
, µ
B
) in the second stage, and Alice plays
a best response according to Lemma 5 by changing market probabilities to
Pr(Y |a
k
, c
B
), Pr(N|a
k
, c
B
) in the third stage.
We compare Alice’s expected profits conditional on her signal in these two
cases with the aid of the sequence selection game. The total payoff of two consec-
utive moves, in a market using the LMSR, is exactly the payoff of moving from the
starting point to the end point of the second move. Hence, the expected profit that
Alice gets from case 1 is the same as what she gets from the following sequence
of actions: (a) Alice changes market probabilities to r
1
that are different from her
posteriors in the first stage; (b) A sequence selection game starts with initial mar-
ket probabilities r
1
; Alice selects herself to be the first player; (c) Alice changes
24 Y. Chen et al.
market probabilities to Pr(Y |a
k
), Pr(N|a
k
); (d) Bob changes market probabili-
ties to Pr(Y |a
k
, c
B
), Pr(N|a
k
, c
B
). Similarly, the expected profit that Alice gets
from case 2 is the same as what she gets from the following sequence of actions:
(a’) Alice changes market probabilities to r
1
that are different from her poste-
riors in the first stage; (b’) A sequence selection game starts with initial market
probabilities r
1
; Alice selects Bob to be the first player; (c’)Bob changes market
probabilities to Pr(Y |r
1
, c
B
, µ
B
), Pr(N|r
1
, c
B
, µ
B
); (d’) Alice changes mar-
ket probabilities to Pr(Y |a
k
, c
B
), Pr(N|a
k
, c
B
). Alice’s expected profit from
(a) is the same as that from (a’). But according to Theorem 3, Alice’s expected
profit from (b), (c), and (d) is greater than or equal to that from (b’), (c’), and (d’).
Hence, Alice does not want to deviate from truthful betting.
A natural question to ask, after we know that Alice being truthful is a PBE, is
whether there are other equilibria with Alice bluffing in the first stage. Theorem 6
says that there are not when players have informative and distinct signals.
Theorem 6 When Alice and Bob have conditionally independent signals that are
informative and distinct, the truthful betting PBE is the unique WPBE of the 3-
stage Alice-Bob-Alice game.
Proof According to Lemma 5, at a WPBE the market probability in the third stage
always reveals information of both Alice and Bob. Thus, the total expected profit
of Alice and Bob from the game is
π
AB
=
¸
k,l,z
Pr(c
A
= a
k
, c
B
= b
l
, ω = z) log
Pr(ω = z|c
A
= a
k
, c
B
= b
l
)
r
0
z
,
which is fixed given r
0
and the joint probability distribution P. Let σ be Al-
ice’s strategy at a WPBE. Then, Alice’s expected profit by following σ is π
A
σ
=
π
AB
− π
B
(σ), where π
B
(σ) is Bob’s expected myopic optimal profit follow-
ing a first round strategy σ as defined in Section 4. According to Theorem 1,
π
B
(σ) ≥ π
B

T
) and the equality holds only when σ = σ
T
. Thus, for any WPBE
strategy σ of Alice, at the equilibrium π
A
σ
≤ π
AB
−π
B

T
), with equality holds
only when σ = σ
T
. However, any strategy σ such that π
A
σ
< π
AB
−π
B

T
) can
not be a WPBE strategy, because Alice can simply deviate to the truthful strategy
and get higher expected profit. Hence, truthful betting is the only WPBE of the
game.
6.2 Nonexistence of Truthful Equilibrium in I Games
We study the equilibrium of the game when Alice and Bob have independent sig-
nals and the general informativeness condition holds. In particular, we show that
the truthful strategy in this setting is not a WPBE. Alice has an incentive to deviate
from the truthful strategy on her first trade in the market.
The truthful strategy for Alice is the same as that in Section 6.1: change market
probability to r
1
= Pr(Y |c
A
= a
k
), Pr(N|c
A
= a
k
) in the first stage. We define
Gaming Prediction Markets: Equilibrium Strategies with a Market Maker 25
the bluffing strategy for Alice as: change market probability to ˜r
1
= Pr(Y |c
A
=
˜ a
k
), Pr(N|c
A
= ˜ a
k
), where ˜ a
k
= a
k
is another signal in C
A
.
Since the total payoff of two consecutive moves in LMSR is exactly the payoff
of moving from the starting point to the end point of the second move, we may
now consider a bluffing strategy by Alice during the first trade as a move from
r
0
to r
1
immediately followed by a move from r
1
to ˜r
1
. Due to the overlap in
the expected score in the truthful strategy and the bluffing strategy for Alice, we
eliminate the r
0
to r
1
move in our comparison analysis and treat the myopic move
as if it had zero expected profit and analyze the bluffing strategy as a move from
r
1
to ˜r
1
.
Lemma 6 Assume that Bob expects Alice to play truthfully, and reacts accord-
ingly. If Alice observes her signal, c
A
= a
k
, and bluffs, her expected change in
profit over following the truthful strategy, is
∆ =
¸
l
Pr(c
B
= b
l
|c
A
= a
k
)D(p
(ω|a
k
,b
l
)
||p
(ω|˜ a
k
,b
l
)
) −D(p
(ω|a
k
)
||p
(ω|˜ a
k
)
),
where p
(ω|·)
denotes the probability distribution of ω conditional on signal real-
izations.
Proof For conciseness, we use Pr(a
k
) to represent Pr(c
A
= a
k
) and similarly for
other terms. Alice’s first trade is a move from r
1
to ˜r
1
, by the definition of LMSR
her expected score from the first move is:
Pr(Y |a
k
) log
Pr(Y |˜ a
k
)
Pr(Y |a
k
)
+ Pr(N|a
k
) log
Pr(N|˜ a
k
)
Pr(N|a
k
)
= −D(p
(ω|a
k
)
||p
(ω|˜ a
k
)
).
As Bob has a probability of Pr(b
l
|a
k
) of seeing b
l
as his signal, Alice
will move the market from r
2
= Pr(Y |˜ a
k
, b
l
), Pr(N|˜ a
k
, b
l
) to r
3
=
Pr(Y |a
k
, b
l
), Pr(N|a
k
, b
l
) with probability Pr(b
l
|a
k
). Therefore, the expected
score of Alice’s second trade in the market is:
X
l
Pr(b
l
|a
k
)
»
Pr(Y |a
k
, b
l
) log
Pr(Y |a
k
, b
l
)
Pr(Y |˜ a
k
, b
l
)
+ Pr(N|a
k
, b
l
)
Pr(N|a
k
, b
l
)
Pr(N|˜ a
k
, b
l
)

=
X
l
Pr(b
l
|a
k
)D(p
(ω|a
k
,b
l
)
||p
(ω|˜ a
k
,b
l
)
)
Thus, Alice’s total expected payoff conditional on her signal is the sum of the
above two expressions, which gives ∆ in our theorem.
Theorem 7 If players have independent signals that satisfy the general informa-
tiveness condition, truthful betting is not a WPBE for the Alice-Bob-Alice game.
Proof Lemma 6 gives the expected profit change of Alice if she were to deviate
from truthful betting. Below we show that the expected profit change of deviating
from the truthful strategy is greater than zero, thus Alice has an incentive to deviate
from the truthful strategy. We have
¸
l
Pr(b
l
|a
k
)D(p
(ω|a
k
,b
l
)
||p
(ω|˜ a
k
,b
l
)
)
≥ D(
¸
l
Pr(b
l
|a
k
)p
(ω|a
k
,b
l
)
||
¸
l
Pr(b
l
|a
k
)p
(ω|˜ a
k
,b
l
)
)
= D(p
(ω|a
k
)
||p
(ω|˜ a
k
)
)
26 Y. Chen et al.
The inequality follows from the convexity of relative entropy [11]. We note that by
the general informativeness condition the inequality is strict. The equality follows
from the independence of the two signals, Pr(b
l
|a
k
) = Pr(b
l
), and the law of total
probability. Therefore, ∆ > 0. Truthful betting is not a WPBE strategy.
6.2.1 Explicit Bluffing Equilibria in a Restricted Game We now introduce a sim-
ple model of independent signals and show that bluffing can be an equilibrium.
In our model, Alice and Bob each see an independent coin flip and then partici-
pate in a LMSR prediction market with outcomes corresponding to whether or not
both coins came up heads. Thus, ω ∈ {HH, (HT|TH|TT)}. We again consider an
Alice-Bob-Alice game structure.
Theorem 8 Consider the Alice-Bob-Alice LMSR coin-flipping game, where the
probability of heads is p. Now restrict Alice’s first stage strategies to either play
truthful betting (TB) or as if her coin is heads (
ˆ
H). A WPBE in this game has Alice
play TB with probability t = 1 +
p
(1−(1−p)
−1/p
)(1−p)
, and otherwise play
ˆ
H.
Proof To show that playing TB with probability t is an equilibrium, we first com-
pute Bob’s best response to such a strategy and then show that Bob’s strategy
makes Alice indifferent between her pure strategies. Bob’s best response is, if he
has heads, to set the probability of HH to the probability that Alice has heads given
that she bets as if she does (we denote Alice betting as if she had heads as
ˆ
H):
Pr(HH |
ˆ
HH) =
Pr(
ˆ
HH | HH) Pr(HH)
Pr(
ˆ
HH | HH) Pr(HH) + Pr(
ˆ
HH | TH) Pr(TH)
=
1 · p
2
1 · p
2
+ (1 −t)(1 −p)p
= 1 −(1 −p)
1
p
.
(10)
If Bob has tails he sets the probability of HH to zero. Assuming such a strategy
for Bob, we can compute Alice’s expected profit for playing TB. This is done
by computing, for each outcome in {HH,HT,TH,TT}, her profit for moving the
probability from p
0
to p or 0, plus her profit for moving the probability to 0 or 1
from where Bob moves it. We have
p
2

log
p
p
0
+ log
1
x

+p(1 −p) log
1−p
1−p
0
+ (1 −p)p log
1−0
1−p
0
+ (1 −p)
2
log
1−0
1−p
0
(11)
where x is Bob’s posterior probability when he has heads and Alice appears to
have heads. Similarly, Alice’s expected profit for always pretending to have heads
in the first stage is:
p
2

log
p
p
0
+ log
1
x

+p(1 −p) log
1−p
1−p
0
+ (1 −p)p

log
1−p
1−p
0
+ log
1−0
1−x

+ (1 −p)
2
log
1−p
1−p
0
.
(12)
Gaming Prediction Markets: Equilibrium Strategies with a Market Maker 27
Since (11) and (12) are equal when x is set according to (10), Alice is indifferent
between truthfulness and bluffing when Bob expects her to play TB with probabil-
ity t. It is therefore in equilibrium for Alice to play TB with probability t, that is,
Alice should, with 1 − t probability, pretend to have seen heads regardless of her
actual information.
7 Finite-Player Finite-Stage Game
We extend our results for the Alice-Bob-Alice game to games with a finite number
of players and finite stages in LMSR. Each player can change the market proba-
bilities multiple times and all changes happen in sequence.
7.1 Truthful Equilibrium of CI games
We first consider the CI setting where players have conditionally independent sig-
nals. The following theorem characterizes the PBE of the game.
Theorem 9 In the finite-player, finite-stage game with LMSR, if players have con-
ditionally independent signals, at a PBE of the game all players report their poste-
rior probabilities in their first round of play. If signals are informative and distinct,
the truthful betting PBE is the unique WPBE of the game.
Proof Given that every player believes that all players before it act truthfully,
we prove the theorem recursively using backward induction. If it’s player i’s last
chance to play, it will truthfully report its posterior probabilities by Lemma 2. If it’s
player i’s second to last chance to play, there are other players standing in between
its second to last chance to play and its last chance to play. We can combine the
signals of those players standing in between as one signal and treat those players
as one composite player. Because signals are conditionally independent, the signal
of the composite player is conditionally independent of the signal of player i. The
game becomes an Alice-Bob-Alice game for player i and at a PBE player i reports
truthfully at its second to last chance to play according to Theorem 5. Inferring
recursively, it is a PBE at which any player reports truthfully at its first chance to
play.
If signals are informative and distinct, Theorem 6 states that the truthful betting
PBE is the unique WPBE of the Alice-Bob-Alice game. This implies that truthful
betting PBE is also a unique WPBE for the finite-player finite-stage game. Suppose
the contrary, there is a bluffing WPBE. Select the last player who bluffs in the
sequence. Without loss of generality, suppose the player is Alice and she bluffs
at some stage t. Then, Alice must play truthfully at some stage t

> t + 1, and
does not play in stages between t and t

. Combining signals of players between
t and t

as one composite signal. The game from stage t to stage t

becomes an
Alice-Bob-Alice game. Alice bluffing at stage t contradicts with the uniqueness of
the WPBE of the Alice-Bob-Alice game.
28 Y. Chen et al.
7.2 No Truthful Equilibrium in I Games
For the I setting where players have independent signals, the result for finite-player
finite-stage game directly follow from our result on the Alice-Bob-Alice game.
Theorem 10 In the finite-player, finite-stage game with LMSR, if players have
independent signals that satisfy the general informativeness condition, and there
is at least one player who trades more than once, truthful betting is not a WPBE
strategy.
Proof Consider the last two trades of a player who plays more than once, for
example Alice. If t and t

> t + 1 are the last two stages for Alice to trade and
all other players play truthfully, we can combine signals of all players who trade
between t and t

as a composite signal possessed by a composite agent. Because
signals are independent, Alice’s signal is independent of the composite signal. The
game from stage t to stage t

can be viewed as an Alice-Bob-Alice game. By
Theorem 7, we know that Alice is better off by bluffing at stage t. Hence, truthful
betting is not a WPBE.
8 Finite-Player Infinite-Stage Games
In this section, we analyze a model in which the number of potential moves can be
infinite. This is motivated by the observation that there is often no pre-defined last
stage for a player. A player may decide to have one more trade at any time before
the market closes.
We extend our results for finite-stage games to infinite-stage games and exam-
ine the equilibrium for CI and I settings respectively.
8.1 Truthful Betting Equilibrium in CI Games
Games with conditionally independent signals remain to have a unique equilib-
rium, truthful betting PBE, as is shown in the following theorem.
Theorem 11 In a potentially infinite-stage game with conditionally independent
signals, at a finite PBE all players truthfully report their posterior probabilities at
their first chance of play. If signals are informative and distinct, the truthful betting
equilibrium is the unique finite WPBE of the game.
Proof If there are mplayers, truthful betting reveals all information after mstages.
Knowing that other players are playing truthfully, a player who plays in stage
i < m does not want to deviate to bluffing in stage i and come to trade again in
stage m + 1, because the game between stages i and m + 1 can be viewed as an
Alice-Bob-Alice game. According to Theorem 5 player i would prefer to report
truthfully.
If there exists a finite equilibriumthat reveals all information at stage t, suppose
the last time for a player to bluff is in stage i, then it must report truthfully in some
Gaming Prediction Markets: Equilibrium Strategies with a Market Maker 29
stage j and i + 1 < j ≤ t. The game between stages i and j can be viewed
as an Alice-Bob-Alice game. If signals are informative and distinct, according to
Theorem 6, truthful betting equilibrium is the only WPBE of the Alice-Bob-Alice
game. Contradiction arises. Truthful betting is the only finite equilibrium.
8.2 Nonexistence of Finite Equilibrium in I Games
Theorem 12 In an infinite-stage game with independent signals that satisfy the
general informativeness condition, there is no WPBE that reveals all information
after finite number of trades.
Proof For contradiction, consider any WPBE strategy profile σ that always results
in the optimal market prediction after some number t of stages in a potentially
infinite-stage game. Consider the last two players, Alice and Bob without loss of
generality. They play in stages t −1 and t respectively. Because the WPBE reveals
all information, both Alice and Bob reports truthfully at stages t − 1 and t. Then,
according to Theorem 7, Alice would want to deviate to bluff in stage t − 1 and
play again in stage t +1. Thus, σ can not be a WPBE strategy profile. There is not
finite equilibrium.
Intuitively if a player has some private information that may be revealed by
its signal, it has an incentive to not fully divulge this information when it may
play multiple times in the market. The general informativeness condition guaran-
tees that no matter what the signal distribution is, a player will always have some
private information revealed by its signal.
9 Discounted LMSR and Entropy Reduction
For infinite-stage games, when players have independent signals, the market can
not fully aggregate information of all players with finite trades. We propose a dis-
counted logarithmic market scoring rule and show that the I games can converge
to full information aggregation exponentially.
We first introduce the discounted LMSR mechanism in Section 9.1. Then, we
characterize the infinite-stage games as continuation games in Section 9.2 and in-
troduce the concepts of complementarity and substitution of signals in Section
9.3. We present a natural metric to quantify degree of information aggregation in
Section 9.4. Facilitated with results in Sections 9.2, 9.3, and 9.4, we present our
convergence analysis for discounted LMSR in Section 9.5.
9.1 Discounted LMSR
One way to address the incentives that traders with independent signals have to
bluff in a market using the logarithmic market scoring rule is to reduce future
payoffs using a discount parameter, perhaps resulting in an incentive to play the
truthful strategy. Based on this intuition, we propose the discounted LMSR market.
30 Y. Chen et al.
Let δ ∈ (0, 1) be a discount parameter. The δ-discounted market scoring
rule is a market microstructure in which traders update the predicted probability
of the event under consideration just as they would in the regular market scoring
rule. However, the value (positive or negative) of trade is discounted over time.
2
For simplicity, we assume a strict alternating sequence of trades between Alice
and Bob. Suppose a trader moves the prediction for outcome Y from p to q in
the i-th stage of the market, and the outcome Y is later observed to happen. The
trader would then be given a payoff of δ
i−1
(log q − log p). On the other hand, if
the outcome Y did not happen, and the trader moves the prediction from p to q in
the i-th stage the trader would earn a payoff of δ
i−1
(log(1 −q) −log(1 −p)). The
regular market scoring rule corresponds to δ = 1.
Clearly, the myopic strategic properties of the market scoring rule are retained
in the discounted form. We note that for CI games, truthful betting is still the
unique WPBE in the discounted LMSR. Intuitively, for an Alice-Bob-Alice CI
game in the discounted LMSR, Alice does not want to deviate from the truthful
betting strategy as in the non-discounted LMSR because for any deviating strat-
egy her expected profit in the first stage is the same in both the non-discounted
LMSR and the discounted LMSR but her expected profit in the third stage in the
discounted LMSR is strictly less than that in the non-discounted LMSR, mak-
ing deviation even more undesirable. As truthful betting is the unique WPBE in
an Alice-Bob-Alice CI game in the discounted LMSR, the same arguments in Sec-
tions 7.1 and 8.1 can show that truthful betting is the unique WPBE for finite-stage
and infinite-stage CI games with finite players in the discounted LMSR. Thus,
from now on, we only focus on I games in the discounted LMSR.
We will show that the discounted LMSR can have better non-myopic strategic
properties for I games. For simplicity, we assume that there are only two play-
ers, Alice and Bob. They alternately play in the market, with Alice being the first
player. Even though that the discounted LMSRmay admit equilibria in which play-
ers bluff with some probability, we will show that, in any WPBE strategy profile
σ, the market price will converge towards the optimal probability for the particular
realized set of information signals. In other words, although complete aggregation
of information may not happen in two rounds, it does surely happen in the long
run at any WPBE.
9.2 Continuation Games
It is useful to think about an infinite-stage market game G as a first-stage move
followed by a subsequent game H that perhaps depends on the first-stage move.
This decomposition is possible in equilibrium analysis, because the actual first-
stage strategy σ
1
is common knowledge in an equilibrium profile. Then, the beliefs
of the players after observing the first move x are consistent with private signals
2
It is possible that traders intrinsically discount future profits; for example, they may be
uncertain about how long their information will remain private. We do not, however, assume
any intrinsic discounting by traders; rather, we are proposing a mechanism that explicitly
discounts gains and losses in later rounds.
Gaming Prediction Markets: Equilibrium Strategies with a Market Maker 31
and a common prior p

that depends on the initial prior p
0
, the first-stage strategy
σ
1
and the observed move x.
Definition 3 We say that H is a continuation game of G if there is a first-stage
strategy σ
1
and a move x such that x is played with nonzero probability under σ,
the prior p

in H is consistent with p
0
, σ
1
, and x, and the sequence of play in H
is the same as the subsequence of play in G after the first move in G. We denote
this G → H.
We define the relation

→ as the reflexive transitive closure of →, i.e.,
we write G

→ H and say G reduces to H if there is a sequence G =
H
0
, H
1
, H
2
, . . . , H
k
= H such that H
i
→ H
i+1
for all i. We define G

→ G
to always be true.
Observe that if G

→ H, then, the full-information optimal probabilities
Pr(Y |c
A
= a, c
B
= b) for any given pair of signal realizations (a, b) are the
same in G and H. Only the prior over different signals changes after a first-round
strategy, not the full-information posterior probabilities.
Lemma 7 shows that continuation games preserve the independence property
of the original game.
Lemma 7 Suppose G

→ H. Then,
– If G satisfies the conditional independence (CI) property, so does H.
– If G satisfies the independence (I) property, so does H.
Proof We need to show that if H
i
→ H
i+1
, H
i+1
satisfies the same independence
property as H
i
.
Let σ be a WPBE strategy for game H
i
. Without loss of generality, suppose
that x is one of the points that the first player in H
i
plays with positive probability
according to σ. Let P
i
be the joint probability distribution for H
i
. We use P
i
{·}
to represent probability of some outcome under P
i
. Let P
i
{c
1
= k|x, σ} be the
probability that player 1 has signal k conditional on that it plays x. Then, knowing
player 1 follows σ, player 2 updates its beliefs after seeing x, the continuation
game H
i+1
has joint probability distribution P
i+1
such that
P
i+1
{c
1
= k, c
2
= l, ω = z} = P
i
{c
1
= k, c
2
= l, ω = z|x, σ}
= P
i
{c
1
= k|x, σ}P
i
{c
2
= l, ω = z|c
1
= k}
=
P
i
{c
1
= k|x, σ}
P
i
{c
1
= k|σ}
P
i
{c
1
= k, c
2
= l, ω = z}
where k, l, and z represent signal realization of player 1, signal realization of
player 2, and realized outcome respectively.
32 Y. Chen et al.
We first show that if P
i
of game H
i
satisfies CI condition, P
i+1
of game H
i+1
also satisfies CI condition.
P
i+1
{c
1
= k, c
2
= l|ω = z} =
P
i+1
{c
1
= k, c
2
= l, ω = z}
P
i+1
{ω = z}
=
P
i
{c
1
=k|x,σ}
P
i
{c
1
=k|σ}
P
i
{c
1
= k, c
2
= l, ω = z}
P
i+1
{ω = z}
=
P
i
{c
1
=k|x,σ}
P
i
{c
1
=k|σ}
P
i
{c
1
= k, c
2
= l|ω = z}
P
i+1
{ω = z}/P
i
{ω = z}
=
P
i
{c
1
=k|x,σ}
P
i
{c
1
=k|σ}
P
i
{c
1
= k|ω = z}P
i
{c
2
= l|ω = z}
P
i+1
{ω = z}/P
i
{ω = z}
.
P
i+1
{c
1
= k|ω = z} =
P
i
{c
1
=k|x,σ}
P
i
{c
1
=k|σ}
P
i
{c
1
= k, ω = z}
P
i+1
{ω = z}
=
P
i
{c
1
=k|x,σ}
P
i
{c
1
=k|σ}
P
i
{c
1
= k|ω = z}
P
i+1
{ω = z}/P
i
{ω = z}
.
P
i+1
{c
2
= l|ω = z} =
¸
k
P
i
{c
1
=k|x,σ}
P
i
{c
1
=k|σ}
P
i
{c
1
= k, c
2
= l, ω = z}
P
i+1
{ω = z}
=
¸
k
P
i
{c
1
=k|x,σ}
P
i
{c
1
=k|σ}
P
i
{c
1
= k, c
2
= l|ω = z}
P
i+1
{ω = z}/P
i
{ω = z}
=
¸
k
P
i
{c
1
=k|x,σ}
P
i
{c
1
=k|σ}
P
i
{c
1
= k, c
2
= l|ω = z}
¸
k
P
i
{c
1
=k|x,σ}
P
i
{c
1
=k|σ}
P
i
{c
1
= k, ω = z}/P
i
{ω = z}
= P
i
{c
2
= l|ω = z}.
Thus, we also have
P
i+1
{c
1
= k, c
2
= l|ω = z} = P
i+1
{c
1
= k|ω = z}P
i+1
{c
2
= l|ω = z}.
CI condition holds for H
i+1
.
We then show that if P
i
of game H
i
satisfies I condition, P
i+1
of game H
i+1
also satisfies I condition.
P
i+1
{c
1
= k, c
2
= l} =
¸
ω
P
i+1
{c
1
= k, c
2
= l, ω = z}
=
¸
ω
P
i
{c
1
= k|x, σ}
P
i
{c
1
= k|σ}
P
i
{c
1
= k, c
2
= l, ω = z}
=
P
i
{c
1
= k|x, σ}
P
i
{c
1
= k|σ}
P
i
{c
1
= k, c
2
= l}
=
P
i
{c
1
= k|x, σ}
P
i
{c
1
= k|σ}
P
i
{c
1
= k|σ}P
i
{c
2
= l}
= P
i
{c
1
= k|x, σ}P
i
{c
2
= l}
Gaming Prediction Markets: Equilibrium Strategies with a Market Maker 33
The second to the last equality is because of the independence condition of H
i
. We
also have
P
i+1
{c
1
= k} =
¸
l
P
i+1
{c
1
= k, c
2
= l}
=
¸
l
P
i
{c
1
= k|x, σ}
P
i
{c
1
= k|σ}
P
i
{c
1
= k, c
2
= l}
=
¸
l
P
i
{c
1
= k|x, σ}
P
i
{c
1
= k|σ}
P
i
{c
1
= k|σ}P
i
{c
2
= l}
= P
i
{c
1
= k|x, σ}
and
P
i+1
{c
2
= l} =
¸
k
P
i+1
{c
1
= k, c
2
= l}
=
¸
k
P
i
{c
1
= k|x, σ}
P
i
{c
1
= k|σ}
P
i
{c
1
= k, c
2
= l}
=
¸
k
P
i
{c
1
= k|x, σ}
P
i
{c
1
= k|σ}
P
i
{c
1
= k|σ}P
i
{c
2
= l}
= P
i
{c
2
= l}.
Hence, P
i+1
{c
1
= k, c
2
= l} = P
i+1
{c
1
= k}P
i+1
{c
2
= l}. H
i+1
satisfies
independence.
By induction, if G

→ H, then H preserves the same independence property as
G.
9.3 Complementarity or Substitution of Signals
In this section, we introduce the concept of complementarity for signals. Intu-
itively, if signals of Alice and Bob are complements, the expected profit from
knowing both signals is greater than the sum of the expected profits from knowing
only the individual signals. If signals are substitutes, the relation is reversed.
For a game G participated by Alice and Bob, the total profit potential in G is
π
AB
= π
A

N
) +π
B

T
). In other words, π
AB
is the sum of Alice’s and Bob’s
expected profit if they both follow the truthful strategy, and Alice moves first. We
note that under either the CI condition or the I condition, the order of play is not
important, i.e., π
AB
= π
B

N
) +π
A

T
) = H(ω|c
A
, c
B
).
Now we define the complementarity coefficient of a game that captures the
complementarity of players’ signals.
Definition 4 The complementarity coefficient C(G) of a market game G is de-
fined as
C(G) =
π
A

N
) +π
B

N
)
π
AB
,
34 Y. Chen et al.
where the right-hand side profits are implicitly a function of the game G under
consideration.
The complementarity bound
ˆ
C
σ
(G) is the minimum value of C(H) over all
games H that G could reduce to by following the strategy profile σ:
ˆ
C
σ
(G) = min
H:G

→H by σ
C(H)
We note that if C(G) < 1, signals are complements. If C(G) > 1, signals are
substitutes.
Lemma 8 Under the CI model, any game G with informative and distinct signals
satisfies
ˆ
C
σ
(G) > 1 for any strategy profile σ.
Proof From Theorem 1, it follows that π
B

N
) > π
B

T
) for any CI game G
with informative and distinct signals. Thus,
π
A

N
) +π
B

N
) > π
A

N
) +π
B

T
) = π
AB
⇒ C(G) > 1.
By lemma 7, we know that any H such that G

→ H itself satisfies the CI condition
no matter what strategy the reduction follows. Thus, we must have
ˆ
C
σ
(G) > 1 for
any σ.
Lemma 9 Under the I model, any game G that starts with the prior probability
satisfies C(G) < 1, and thus,
ˆ
C
σ
(G) < 1.
Proof From Theorem 2, it follows that π
B

T
) > π
B

N
) for any I game G
starting with the prior probability. Thus,
π
A

N
) +π
B

N
) < π
A

N
) +π
B

T
) = π
AB
⇒ C(G) < 1.
By definition of
ˆ
C, we have
ˆ
C
σ
(G) ≤ C(G), and thus we must have
ˆ
C
σ
(G) < 1.

9.4 Profit Potential and Entropy
We now present a natural metric, the information deficit D
i
, that quantifies the
degree of aggregation in a prediction market after i trades under strategy profile σ:
D
i
is the expectation, over all possible signal realizations and the randomization
of moves as dictated by σ, of the relative entropy between the optimal probability
(given the realization of the signals) and the actual probability after i stages.
Formally, for a strategy profile σ and a number of stages t, an information
node φ consists of a realization of the signals of the two players and a sequence
of t trades in the market. Let p
i
(φ) = Pr(Y |first i trades in φ, σ), and denote
p
i
(φ) = p
i
(φ), 1 − p
i
(φ). Let p

(φ) = Pr(Y |realization of signals in φ), and
denote p

(φ) = p

(φ), 1 − p

(φ). The aggregative effect of the strategy pro-
file σ after i trades is summarized by the collection of information nodes φ that
can be reached, the market probability p
i
(φ) in stage i, and the associated ex-ante
Gaming Prediction Markets: Equilibrium Strategies with a Market Maker 35
probability Pr(φ) of reaching each such information node. Now for i ≤ t, we
define
D
i
(σ) = E
σ
[D(p

(φ)||p
i
(φ))]
=
¸
φ:φ∈σ
Pr(φ)D(p

(φ)||p
i
(φ)) .
When t = 0, the information nodes φ correspond to different realizations of the
signals. Note that D
0
is the same as π
AB
defined in the previous section.
If D
i
= 0, it implies that the market will always have reached its optimal
probability for the realized signals by the i-th stages. If D
i
> 0, it indicates that,
with positive probability, the market has not yet reached the optimal probability.
D
i
is always nonnegative, because the relative entropy is always nonnegative.
We now show that, in addition to measuring the distance from full information
aggregation, D
i
also enables interesting strategic analysis. The key result is that D
i
can be related to the expected payoff of the i-th stage move in the non-discounted
(standard) LMSR:
Lemma 10 Let M
i
denote the expected profit (over all signal nodes φ
i
) of the i-th
trade under profile σ. Then, M
i
= D
i−1
−D
i
Proof
M
i
=
¸
φ:φ∈σ
Pr(φ)
¸
p

(φ) log
p
i
(φ)
p
i−1
(φ)
+ (1 −p

(φ)) log
1 −p
i
(φ)
1 −p
i−1
(φ)

=
¸
φ:φ∈σ
Pr(φ)

D(p

||p
i−1
) −D(p

||p
i
)

= D
i−1
−D
i
The first equality holds by the definition of M
i
and the second by the definition of
relative entropy.
This suggests another interpretation for D
i
: D
i
represents the expected value
of the potential profit left for trades after the i-th trade.
Given the definition of M
i
, we can now define
˜
M
i
as the expected profit of the
i-th trade in the discounted LMSR. We assume discounting after every even trade,
i.e., after every trade by Bob.
˜
M
i
=

δ
k
(D
2k
−D
2k+1
) ∀i = 2k + 1
δ
k
(D
2k+1
−D
2(k+1)
) ∀i = 2k + 2
(13)
Using the definition of
˜
M
i
we write the total profit for Alice in the δ-
discounted LMSR as:
Alice’s payoff =
¸
i:i=2k,k∈Z
+
˜
M
i−1
= D
0
−D
1
+δ(D
2
−D
3
) +. . .
We can similarly rewrite Bob’s expected payoff as:
Bob’s payoff =
¸
i:i=2k,k∈Z
+
˜
M
i
= D
1
−D
2
+δ(D
3
−D
4
) +. . .
36 Y. Chen et al.
We reiterate that the definition of D
i
is not dependent on δ. It is a measure
of the informational distance between the prices after i trades in profile σ and the
optimal prices. Of course, the stability of a given strategy profile σ may change
with δ.
9.5 Asymptotic Convergence in Discounted LMSR for I Games
In this part, we bound the value of D
n
for large n, in any WPBE, for I games
that start with the prior probability. We will express our convergence bound in
terms of complementarity coefficient, C(G). Note that if C(G) = 0, the myopic
strategy may not involve any movement by either player, and thus, we could have
lack of information aggregation even with the myopic strategy. We exclude such
degenerate cases, and assume that C(G) > 0.
Now, fix a discount parameter δ, and a particular instance of the two-person
market game induced by the δ-discounted LMSR. Let σ be any WPBE of this
market game. Without loss of generality we assume that Bob moves second in the
market.
In the first stage, Alice will take some (perhaps mixed) action x dictated by
σ. Under the equilibrium strategy profile σ, Bob will revise his beliefs consistent
with the profile σ and the realized action x. According to Theorem 2, we have
π
B
(σ) ≥ π
B

N
).
Recall that after stage 2, the total expected payoff of both players is at most
δD
2
. We also know that the total expected payoff of Alice in equilibrium is at
least π
A

N
), because if not, a simple deviation to the truthful strategy would
be beneficial. Because π
B
(σ) ≥ π
B

N
), the total expected payoff of Bob in
equilibrium is also at least π
B

N
). This means that the total payoff of the first
two rounds in the market is at least π
A

N
) +π
B

N
) −δD
2
. Therefore we can
bound D
2
as:
D
2
≤ D
0
−[π
A

N
) +π
B

N
) −δD
2
].
Note that D
0
= π
AB
.
This argument generalizes to any even number of rounds, by looking at the
total expected profits within the first 2k moves, and the remaining profit potential
δ
k
D
2k
:
D
2k
≤ D
0
−[π
A

N
) +π
B

N
) −δ
k
D
2k
] (14)
⇐⇒ (1 −δ
k
)D
2k
≤ D
0
−[π
A

N
) +π
B

N
)]
= D
0

1 −
π
A

N
) +π
B

N
)
D
0

⇐⇒ D
2k

D
0
(1 −C(G))
1 −δ
k
(15)
Note that, by definition, C(G) ≥
ˆ
C
σ
(G). Thus, we can rewrite inequality (15)
as:
D
2k

D
0
(1 −
ˆ
C
σ
(G))
1 −δ
k
(16)
Gaming Prediction Markets: Equilibrium Strategies with a Market Maker 37
From inequality (16) we see that the bound on D
2k
depends only on δ, as
D
0
and
ˆ
C
σ
(G) are both constants for any instance of the market and equilibrium
profile σ. Now consider the remainder of the game after 2k rounds, denoted as
H
2k
, then G

→ H
2k
. According to Lemma 7, H
2k
preserves the independence
property of G. Hence, the equilibrium profile σ will also induce an equilibrium
profile on H
2k
. We can repeat the argument to bound D
4k
in terms of D
2k
, etc.
In this way, we rewrite inequality (16) in terms of δ and for a round n = 2km.
We set k such that δ
k/2
=
ˆ
C
σ
(G), i.e. k =
2 log
ˆ
C
σ
(G)
log δ
. Using this value of k and a
value m =
n
2k
we note that:
D
n
≤ D
0

1−
ˆ
C
σ
(G)
1−δ
k

m
= D
0

1−
ˆ
C
σ
(G)
(1−δ
k/2
)(1+δ
k/2
)

m
= D
0
(1 +δ
k/2
)
−m
= D
0
(1 +
ˆ
C
σ
(G))

n
2k
= D
0
(1 +
ˆ
C
σ
(G))

nlog δ
4 log
ˆ
C
σ
(G)
= D
0
δ
n
−log(1+
ˆ
C
σ
(G))
4 log
ˆ
C
σ
(G)
(17)
Note that
−log(1+
ˆ
C
σ
(G))
4 log
ˆ
C
σ
(G)
depends only on the complementarity bound of
the game. According to Lemma 9, for any game G with independent signal,
ˆ
C
σ
(G) < 1. Hence,
−log(1+
ˆ
C
σ
(G))
4 log
ˆ
C
σ
(G)
> 0. Therefore, inequality (17) shows that
the relative entropy of the probabilities with respect to the optimal probabilities
reduces exponentially over time.
To summarize, we have proven the following result:
Theorem 13 Let G be a market game in the I model, with a δ-discounted log-
arithmic market scoring rule. For this game, let σ be any WPBE strategy profile.
Then, for all n > 1, the information deficit D
n
after n rounds of trading is bounded
by
D
n
≤ D
0
δ
n
−log(1+
ˆ
C
σ
(G))
4 log
ˆ
C
σ
(G)

10 Conclusion
We have investigated the strategic behavior of traders in LMSR prediction mar-
kets using dynamic games with incomplete information. Specifically, we examine
different scenarios where traders at equilibrium bet truthfully or bluff.
Two different information structures, conditional independence and uncondi-
tional independence of signals, are considered. Both finite-stage and infinite-stage
games are investigated. We show that traders with conditionally independent sig-
nals may be worse off by bluffing in LMSR. Moreover, for both finite-stage and
infinite-stage games, truthful betting is not only a PBE strategy but also a unique
WPBE strategy for all traders in LMSR.
38 Y. Chen et al.
On the other hand, when the signals of traders are unconditionally independent,
truthful betting is not a WPBE for finite-stage games. Traders have incentives to
strategically mislead other traders with the intent of correcting the errors made
by others in a later stage. We provide a bluffing equilibrium for a restricted two-
player game. For infinite-stage games with unconditionally independent signals,
there does not exist any WPBE that fully reveals information of all players in
finite stages.
We use a simple modification to the market scoring rule, which includes a
form of discounting, to ameliorate this potential problem. This allows us to prove
a bound on the rate at which the error of the market, as measured by the relative
entropy between full information aggregation and the actual market probability,
reduces exponentially over time. The exponent depends on the complementarity
coefficient of the market instance.
The need for discounting shows a connection to classical bargaining settings in
which players bargain over how to divide a surplus they can jointly create. In a pre-
diction market, informed players can extract a profit from the market. Moreover,
players can pool their information together to make sharper predictions than either
could alone and thus extract an even larger profit. They might engage in bluffing
strategies to bargain over how this subsidy is divided. Explicit discounting can
make this bargaining more efficient.
Acknowledgments
Stanko Dimitrov was supported by NSF grants CCF-0728768 and IGERT-
0114368. Rahul Sami was partially supported by NSF grant CCF-0728768.
We are grateful to Paul J. Healy and Mohammad Mahdian for helpful discus-
sions. We thank John Ledyard for pointing us to earlier work on discounting in
parimutuel markets. We thank anonymous reviewers for their valuable comments.
References
1. F. Allen and D. Gale. Stock-price manipulation. The Review of Financial Studies,
(5):503–529, 1992.
2. B. S. Axelrod, B. J. Kulick, C. R. Plott, and K. A. Roust. Design improved parimutuel-
type information aggregation mechanisms: Inaccuracies and the long-shot bias as dise-
quilibrium phenomena. Journal of Economic Behavior and Organization, 69:170–181,
2009.
3. J. E. Berg, R. Forsythe, F. D. Nelson, and T. A. Rietz. Results from a dozen years of
election futures markets research. In C. A. Plott and V. Smith, editors, Handbook of
Experimental Economics Results. North Holland, Amsterdam, 2008.
4. J. E. Berg and T. A. Rietz. Prediction markets as decision support systems. Information
Systems Frontier, 5:79–93, 2003.
5. T. B¨ orgers, A. Hernando-Veciana, and D. Kr¨ ahmer. When are signals complements or
substitutes? Working paper, 2007.
6. C. F. Camerer. Can asset markets be manipulated? A field experiment with race-track
betting. Journal of Political Economy, (106):457–482, 1998.
Gaming Prediction Markets: Equilibrium Strategies with a Market Maker 39
7. A. Chakraborty and B. Yilmaz. Manipulation in market order models. Journal of
Financial Markets, 7(2):187–206, 2004.
8. K.-Y. Chen and C. R. Plott. Information aggregation mechanisms: Concept, design and
implementation for a sales forecasting problem. Working paper No. 1131, California
Institute of Technology, Division of the Humanities and Social Sciences, 2002.
9. Y. Chen and D. M. Pennock. A utility framework for bounded-loss market makers.
In Proceedings of the 23rd Conference on Uncertainty in Artificial Intelligence (UAI
2007), pages 49–56, Vancouver, BC Canada, 2007.
10. Y. Chen, D. M. Reeves, D. M. Pennock, R. D. Hanson, L. Fortnow, and R. Gonen.
Bluffing and strategic reticence in prediction markets. In Lecture Notes in Computer
Science, The Third Workshop on Internet and Network Economics (WINE 2007), pages
70–81, 2007.
11. T. M. Cover and J. A. Thomas. Elements of Information Theory. John Wiley & Sons,
Inc., 1991.
12. S. Debnath, D. M. Pennock, C. L. Giles, and S. Lawrence. Information incorporation
in online in-game sports betting markets. In Proceedings of the Fourth Annual ACM
Conference on Electronic Commerce (EC’03), San Diego, CA, 2003.
13. S. Dimitrov and R. Sami. Non-myopic strategies in prediction markets. In Proceedings
of the Nineth ACMConference on Electronic Commerce (EC’08), NewYork, July 2008.
ACM.
14. E. F. Fama. Efficient capital market: A review of theory and empirical work. Journal
of Finance, 25:383–417, 1970.
15. J. Feigenbaum, L. Fortnow, D. M. Pennock, and R. Sami. Computation in a distributed
information market. Theoretical Computer Science, 343:114–132, 2005. (A prelimi-
nary version appeared in the 2003 ACM Conference on Electronic Commerce).
16. R. Forsythe, F. Nelson, G. R. Neumann, and J. Wright. Anatomy of an experimental
political stock market. American Economic Review, 82(5):1142–1161, 1992.
17. R. Forsythe, T. A. Rietz, and T. W. Ross. Wishes, expectations, and actions: A sur-
vey on price formation in election stock markets. Journal of Economic Behavior and
Organization, 39:83–110, 1999.
18. J. M. Gandar, W. H. Dare, C. R. Brown, and R. A. Zuber. Informed traders and price
variations in the betting market for professional basketball games. Journal of Finance,
LIII(1):385–401, 1999.
19. J. Hansen, C. Schmidt, and M. Strobel. Manipulation in political stock markets - pre-
conditions and evidence. Technical Report, 2001.
20. R. D. Hanson. Combinatorial information market design. Information Systems Fron-
tiers, 5(1):107–119, 2003.
21. R. D. Hanson. Logarithmic market scoring rules for modular combinatorial information
aggregation. Journal of Prediction Markets, 1(1):1–15, 2007.
22. R. D. Hanson, R. Oprea, and D. Porter. Information aggregation and manipulation in
an experimental market. Journal of Economic Behavior and Organization, 60(4):449–
459, 2007.
23. P. Kumar and D. J. Seppi. Futures manipulation with cash settlement. Journal of
Finance, (47):1485–1502, 1992.
24. A. S. Kyle. Continuous auctions and insider trading. Econometrica, 53(6):1315–1336,
1985.
25. A. Mas-Colell, M. D. Whinston, and J. R. Green. Microeconomics Theory. Oxford
University Press, New York, NY, 1995.
26. P. Milgrom and N. L. Stokey. Information, trade and common knowledge. Journal of
Economic Theory, 26(1):17–27, 1982.
40 Y. Chen et al.
27. E. Nikolova and R. Sami. A strategic model for information markets. In Proceedings
of the Seventh Annual ACM Conference on Electronic Commerce (EC ’07), 2007.
28. K. Oliven and T. A. Rietz. Suckers are born, but markets are made: Individual ratio-
nality, arbitrage and market efficiency on an electronic futures market. Management
Science, 50(3):336–351, 2004.
29. P. W. Rhode and K. S. Strumpf. Historical presidential betting markets. Journal of
Economic Perspectives, 18(2):127–142, 2004.
30. P. W. Rhode and K. S. Strumpf. Manipulating political stock markets: A field experi-
ment and a century of observational data. Working Paper, 2007.
31. R. H. Thaler and W. T. Ziemba. Anomalies: Parimutuel betting markets: Racetracks
and lotteries. Journal of Economic Perspectives, 2(2):161–174, 1988.

2

Y. Chen et al.

the infinite-stage market with discounting, the market price converges to the fullyrevealing price, and the rate of convergence can be bounded in terms of the discounting parameter. When signals are conditionally independent, truthful betting is the unique WPBE for the infinite-stage market with and without discounting. Key words Prediction markets, game theory, bluffing, strategic betting 1 Introduction It has long been observed that, because market prices are influenced by all the trades taking place, they reflect the combined information of all the traders. The strongest form of the efficient markets hypothesis [14] posits that information is incorporated into prices fully and immediately, as soon as it becomes available to anyone. A prediction market is a financial market specifically designed to take advantage of this property. For example, to forecast whether a product will launch on time, a company might ask employees to trade a security that pays $1 if and only if the product launches by the planned date. Everyone from managers to developers to administrative assistants with different forms and amounts of information can bet on the outcome. The resulting price constitutes their collective probability estimate that the launch will occur on time. Empirically, such prediction markets outperform experts, group consensus, and polls across a variety of settings [16, 17, 28, 3, 4, 18, 31, 12, 8]. Yet the double-sided auction at the heart of nearly every prediction market is not incentive compatible. Information holders do not necessarily have incentive to fully reveal all their information right away, as soon as they obtain it. The extreme case of this is captured by the so-called no trade theorems [26]: When rational, risk-neutral agents with common priors interact in an unsubsidized (zero-sum) market, the agents will not trade at all, even if they have vastly different information and posterier beliefs. The informal reason is that any offer by one trader is a signal to a potential trading partner that results in belief revision discouraging trade. The classic market microstructure model of a financial market posits two types of traders: rational traders and noise traders [24]. The existence of noise traders turns the game among rational traders into a positive-sum game, thereby resolving the no-trade paradox. However, even in this setting, the mechanism is not incentive compatible. For example, monopolist information holders will not fully reveal their information right away: instead, they will leak their information into the market gradually over time to obtain a greater profit [7]. Instead of assuming or subsidizing noise traders, a prediction market designer might choose to directly subsidize the market by employing an automated market maker that expects to lose money. Hanson’s market scoring rule market maker (MSR) is one example [20, 21]. MSR requires a patron to subsidize the market but guarantees that the patron cannot lose more than a fixed amount set in advance, regardless of how many shares are exchanged or what outcome eventually occurs. The greater the subsidy, the greater the effective liquidity of the market. Since

Gaming Prediction Markets: Equilibrium Strategies with a Market Maker

3

traders face a positive-sum game, even rational risk-neutral agents have incentive to participate. In fact, even a single trader can be induced to reveal information, something impossible in a standard double auction with no market maker. Hanson proves that myopic risk-neutral traders have incentive to reveal all their information. However, forward-looking traders may not. Though subsidized market makers improve incentives for information revelation, the mechanisms are still not incentive compatible. Much of the allure of prediction markets is the promise to gather information from a distributed group quickly and accurately. However, if traders have demonstrable incentives to either hide or falsify information, the accuracy of the resulting forecast may be in question. One frequent concern about incentives in prediction markets is based on non-myopic strategies: strategies in which the attacker sacrifices some profit early in order to mislead other traders, and then later exploit erroneous trades by other traders, thereby gaining an overall profit. 1.1 Our Results In this paper, we study strategies under a logarithmic market scoring rule (LMSR) in a Bayesian extensive-form game setting with incomplete information. We show that different information structures can lead to radically different strategic properties by analyzing two natural classes of signal distributions: conditionally independent signals and unconditionally independent signals. For conditionally independent signals, we show that truthful betting is a Perfect Bayesian Equilibrium in finite-stage and infinite-stage models. Further, we show that the truthful betting equilibrium is the unique Weak Perfect Bayesian Equilibrium in this setting. Thus, bluffing strategies are not a concern. In contrast, when signals are unconditionally independent, we show that truthful betting is not a Weak Perfect Bayesian Equilibrium in the finite-stage model. In the infinite-stage model, we show that there is no Weak Perfect Bayesian Equilibrium that results in full information revelation in a finite number of trades. We propose and analyze a discounted version of the LMSR that mitigates the strategic hazard. In the discounted LMSR, we prove that under any WPBE strategy, the relative entropy between the market price and the optimal full-information price tends to zero at an exponential rate. The rate of convergence can be bounded in terms of the discounting parameter and a measure of complementarity in the information setting. 1.2 Related Work Theoretical work on price manipulation in financial markets [1, 7, 23] explains the logic of manipulation and indicates that double auctions are not incentive compatible. This literature has studied manipulation based on releasing false information (perhaps through trades in other markets), as well as manipulation that only requires strategic manipulation in a single market. The latter form of manipulation is closely related to our study here. Allen and Gale [1] describe a model in which

4

Y. Chen et al.

a manipulative trader can make a deceptive trade in an early trading stage and then profit in later stages, even though the other traders are aware of the possibility of deception and act rationally. They use a stylized model of a multi-stage market; in contrast, we seek to exactly model a market scoring rule model. Apart from other advantages of detailed modeling, this allows us to construct simpler examples of manipulative scenarios: Allen and Gale’s model [1] needs to assume traders with different risk attitudes to get around no-trade results, which is rendered unnecessary by the inherent subsidy with a market scoring rule mechanism. Our model requires only risk-neutral traders and exactly captures the market scoring rule prediction markets. We refer readers to the paper by Chakraborty and Yilmaz [7] for references to other research on manipulation in financial markets. There are some experimental and empirical studies on price manipulation in prediction markets using double auction mechanisms. The results are mixed, some giving evidence for the success of price manipulation [19] and some showing the robustness of prediction markets to price manipulation [6, 22, 29, 30]. Feigenbaum et al. [15] also study prediction markets in which the information aggregation is sometimes slow, and sometimes fails altogether. In their setting, the aggregation problems arise from a completely different source: the traders are nonstrategic but extracting individual traders’ information from the market price is difficult. Here, we study scenarios in which extracting information from prices would be easy if traders were not strategic; the complexity arises solely from the use of potentially non-myopic strategies. Nikolova and Sami [27] present an instance in which myopic strategies are not optimal in an extensive-form game based on the market, and suggest (but do not analyze) using a form of discounting to reduce manipulative possibilities in a prediction market. Axelrod et al. [2] also propose a form of discounting in an experimental parimutuel market and show that it promotes early trades. Unlike the parimutuel market, the market scoring rule has an inherent subsidy, so it was not obvious that discounting would have strategic benefits in our setting as well. B¨ rgers et al. [5] study when signals are substitutes and complements in a o general setting. Our equilibrium and convergence result suggests that prediction markets are one domain where this distinction is of practical importance. This paper is a synthesis and extension of two independent sets of results, which were presented in preliminary form by Chen et al. [10] and Dimitrov and Sami [13]. 1.3 Structure of the Paper This article is organized as follows: In section 2, we present a little background information about market scoring rules. Section 3 details our formal model and information structures. In Section 4 we investigate how predefined sequence of play affects players’ expected profits in LMSR. In Section 5 we consider, when a player can choose to play first or second, what is its equilibrium strategy in a 2-player LMSR. Results of Sections 4 and 5 serve as building blocks for our analysis in subsequent sections. Section 6 studies the equilibrium of a 2-player,

market scoring rules are known to guarantee that the market maker’s loss is bounded. Section 9 proposes a discounted LMSR and analyzes how the mechanism discourages non-truthful behavior. An MSR market can be equivalently implemented as a market maker offering n securities. .Gaming Prediction Markets: Equilibrium Strategies with a Market Maker 5 3-stage game. . if a trader can only change the probability estimate once. rn is a reported probability estimate for the random variable X. (1) is used. The market maker uses a proper scoring rule. Finally. this modified proper scoring rule still incentivizes the trader to reveal its true probability estimate. 2. If a trader with probability ˜ moves the r market probability from r to ˜. we conclude in Section 10. they might have incentive to manipulate information and mislead other traders. Because traders change the probability estimate in sequence. Subsidizing a market to predict the likelihood of each outcome. . changing the market probability of outcome i to some value ri is the same as buying the security for outcome i until the market price of the security reaches ri . every trader in the market may change the current probability estimate to a new estimate of its choice at any time as long as it agrees to pay the scoring rule payment associated with the current probability estimate and receive the scoring rule payment associated with the new estimate. However. A trader’s expected profit in LMSR directly corresponds to the concept of relative entropy in information theory. where r = r1 . S = {s1 (r). . If outcome i is realized. 2 Background Consider a discrete random variable X that has n mutually exclusive and exhaustive outcomes. Our analysis in this paper is facilitated by directly dealing with probabilities. Conceptually. called market scoring rules (MSR). A popular MSR is the logarithmic market scoring rule (LMSR) where the logarithmic scoring rule si (r) = b log(ri ) (b > 0). We generalize this result to arbitrary finite-player finite-stage games in Section 7. 21] shows how a proper scoring rule can be converted into a market maker mechanism. sn (r)}. . r r Since a proper scoring rule is incentive compatible for risk-neutral agents. ˜) = r i=1 ri (si (˜) − si (r)) = b ˜ r ri log ˜ i=1 ri ˜ = bD(˜||r). We analyze infinite-stage games in Section 8. Hence. 9]. . when traders can participate multiple times. . . The market maker pays the last trader and receives payment from the first trader. each corresponding to one outcome and paying $1 if the outcome is realized [20. its expected profit (score) in LMSR is r n n S(r. MSR can be thought of as a sequential shared version of the scoring rule. a trader that changes the probability estimate from r to ˜ pays si (r) and receives si (˜). r ri (2) .1 Market Scoring Rules Hanson [20.

P : Ω×C1 ×· · ·×Cm → [0. without loss of generality we assume that b = 1 in the rest of the paper.6 Y. with the intent of capitalizing on their resultant misinformed trading. Bluffing is the strategy of betting contrary to one’s information in order to deceive future traders. it is the strategy of always buying immediately when the price is too low and selling when the price is too high. about the state of the world at the beginning of the market. 3. N } is the state space of this event. ω ∈ Ω.1 3 LMSR in a Bayesian Framework In this part. 1]. A market. Pr(ω = N ) . The actual realization of the signal is observed only by the player receiving the signal. D(p||q) is the relative entropy or Kullback Leibler distance between two probability mass functions p(x) and q(x) and is defined in [11] as D(p||q) = x p(x) log p(x) . is common knowledge. p0 = Pr(ω = Y ). Ci is the signal space of player i with |Ci | = ni . because the market degenerates to a race to capitalize on the shared information first. we introduce our game theoretic model of LMSR market in order to capture the strategic behavior in LMSR when players have private information. is picked by nature according to a prior p0 = p0 .2 Terminology Truthful betting (TB) for a player in MSR is the strategy of immediately changing the market probability to the player’s believed probability. Since b is a scaling parameter. q(x) D(p||q) is nonnegative and equals zero only when p = q. rN . and too high when current expected payoff is lower than the price. ci . There are m risk neutral players in the market. Chen et al. Ω = {Y. This paper investigates scenarios where traders with incomplete information have an incentive to deviate from truthful betting. “Too low” and “too high” are determined by the player’s information. 1 With complete information. The joint distribution of the true state and players’ signals. . Truthful betting fully reveals a player’s payoff-relevant information.1 General Settings We consider a single event that is the subject of our predictions. traders should reveal all information right away in MSR. uses a LMSR market maker with initial probability 0 0 estimate r0 = rY . 2. In other words. aiming at Y N predicting the true state ω. Each player i gets a private signal. The maximum amount a LMSR market maker can lose is b log n. The price is too low when the current expected payoff is higher than the price. The true state.

we further assume that conditionally independent signals are informative and distinct. Formally. players’ signals are assumed to be independent conditional on the state of the world. Suppose the problem for a prediction market is to predict whether a batch of product is manufactured with high quality materials or low quality materials. If the product is manufactured with low quality materials. Some consumers who each bought a product have private observations of whether their products break in the first month of use.1 Games with Conditionally Independent (CI) Signals We start with a concrete example before formally introducing the model for conditionally independent signals. as well as the observed prices after the first k − 1 trades. In order to rule out degenerated cases. We study two models of independence – independence conditional on the true state. in the LMSR market. and ∀i. If a player i is designated to trade at some stage k in the sequence.When two signals give a player the same posterior. 3. ∀1 ≤ i ≤ m and ∀a ∈ Ci .e. when observing a signal does not change a player’s belief. Player i has distinct signals when its posterior probability is different after observing different signals.. If the product is manufactured with high quality materials. It turns out that the strategic analysis of the market depends critically on independence properties of players’ signals. and analogously p(ci |N ) if the true state is N . in games with conditionally independent signals. cj |ω) = Pr(ci |ω) Pr(cj |ω) is always satisfied by P. Intuitively.Gaming Prediction Markets: Equilibrium Strategies with a Market Maker 7 Players trade sequentially. These are described below. . signals are distinct if and only if Pr(ω = Y |ci = a) = Pr(ω = Y |ci = a ). An informative signal means that after observing the signal. we have not made any assumptions about the joint distribution P. Formally. a = a . This is an example where consumers have conditionally independent signals – conditional on the quality of the materials. Conditionally independent signals are usually appropriate for modeling situations where the true state of the world has been determined but is unknown. signals are informative if and only if Pr(ci = a|ω = Y ) = Pr(ci = a|ω = N ). Formally. the probability for a product to break in its first month of use is 0. a ∈ Ci . for any two players i and j. consumers’ observations are independent.1. and signal realizations are influenced by the true state of the world. This class can be interpreted as player i’s signal ci is independently drawn by nature according to some conditional probability distribution p(ci |Y ) if the true state is Y . in one or more stages of trading. and unconditional independendence – that are both natural in different settings. conditioned on the eventual outcome of the event. it can condition its beliefs of the likelihood of an outcome on its private signals. we can combine these two signals into one because they provide same information. The quality of the materials will be revealed by a test in the future. Pr(ci . i.1. we can simply remove the signal from the player’s signal space because it does not provide any information. Up to this point.01. a player’s posterior is different from its prior. ∀a. In other words. Players are risk-neutrual Bayesian agents. the probability for a product to break in its first month of use is 0.

which can arguably be thought as independent of each other. ∀i = i. For m > 2. which contradicts the informativeness of signals. Then. Lemma 1 If players have informative signals that are conditionally independent. signals of players i and j. Pr(ci = a. they might stochastically influence the eventual outcome of the event. j) = . j = j : Pr(Y |c1 = i. the following property must be true: ∀i. This is an example of a game with unconditionally independent signals. signal realizations are not caused by the state of the world. we primarily prove results about the lack of truthful equilibria. we must have ∀j. Chen et al. but instead. we will often invoke the following general informativeness condition. i ∈ C1 and ∀j. Formally. in a political election prediction market. for m = 2.2 Games with (Unconditionally) Independent (I) Signals In some other situations. Such results require us to show that there is a strict advantage to deviate to an alternative strategy. j ∈ C2 such that i = i . In order to rule out degenerate cases in which the inequalities are not strict.1. The above expression equals 0 only when Pr(ci = a|Y ) − Pr(ci = a|N ) = 0 or Pr(cj = b|Y ) − Pr(cj = b|N ) = 0 or both. voters’ private information is their votes. Pr(ci . Pr(ci = a. c2 = j) = Pr(Y |c1 = i . cj ) = Pr(ci ) Pr(cj ) is always satisfied by P. Formally. Hence. Definition 1 An instance of the prediction market with m players and joint distribution P satisfies the general informativeness condition if there is no vector of signals for any m − 1 players that makes the mth player’s signals reveal no distinguishing information about the optimal probability. Pr(Y |i. For instance. are unconditionally independent. cj = b) − Pr(ci = a) Pr(cj = b) = 0 must be satisfied for all a ∈ Ci and b ∈ Cj . cj = b) − Pr(ci = a) Pr(cj = b) = Pr(ci = a.8 Y. 3. their signals are unconditionally dependent. For games with independent signals. c2 = j). By conditional independence of signals. c2 = j) = Pr(Y |c1 = i. ci and cj . cj = b|ω = z) Pr(ω = z) Pr(ci = a|ω = z) Pr(ω = z) z z z − = z Pr(cj = b|ω = z) Pr(ω = z) Pr(ci = a|ω = z) Pr(cj = b|ω = z) Pr(ω = z) − Pr(ci = a|ω = z) Pr(ω = z) z z Pr(cj = b|ω = z) Pr(ω = z) = Pr(Y ) Pr(N ) (Pr(ci = a|Y ) − Pr(ci = a|N )) (Pr(cj = b|Y ) − Pr(cj = b|N )) . for any two players i and j in games with independent signals. Lemma 1 shows that conditional independence implies unconditional dependence. c2 = j ) and Pr(Y |c1 = i. ci and cj are unconditionally dependent. Proof Suppose the contrary. The election outcome – which candidate gets the majority of votes – is determined by all votes.

We do not consider the case when signals are both conditionally dependent and unconditionally dependent. In this paper. [25]. is a Weak Perfect Bayesian Equilibrium (WPBE) if and only if. for each player. j). we use the WPBE concept. when giving a specific equilibrium. conditional independence and unconditional independence. When proving the nonexistence of truthful betting equilibrium and characterizing the set of all equilibria. Informally. Frequently. An assessment profile (A1 . and j is a vector of signals for the other m − 1 players. By Lemma 1. where i. that we discuss in this paper are mutually exclusive. which further require beliefs at nodes that are off the equilibrium path to be consistent. consisting of an assessment for each player. are used.. The belief system component of an assessment specifies what a player believes at each node of the game tree. but not exhaustive. every node that is reached with positive probability).. Specifying a plausible play of the game involves specifying not just the moves that players make for different information signals. We allow for strategies to be (behaviorally) mixed. the refined concepts of Perfect Bayesian Equilibrium (PBE) or sequential equilibrium. i are two possible signals for any one player. 3.Gaming Prediction Markets: Equilibrium Strategies with a Market Maker 9 Pr(Y |i . . The strategy dictates what move the player will make at each node in the game tree at which she has to move. the strategies are sequentially rational given their beliefs and their beliefs at any node that is reached with nonzero probability are consistent with updating their prior beliefs using Bayes rule. the belief systems of the players are completely defined at every node on the equilibrium path (i. contingent on reaching the node.. We will abuse notation slightly by simply referring to an “equilibrium strategy profile”. the belief at a node consists of probabilities of other players’ signal realizations. For a formal definition of the equilibrium concepts. because the results thus hold a fortiori for refinements of the WPBE concept. but also the beliefs that they have at each node of the game tree. µi ) for a player i consists of a strategy σi and a belief system µi . This is a relatively weak notion of equilibrium for this class of games. we will not consider players’ beliefs off the equilibrium path for WPBEs. Am ). . given the strategies. we use the refined PBE concept. indeed. we know that unconditional independence implies conditional dependence. Given the strategy components of a WPBE profile.2 Equilibrium Concepts The prediction market model we have described is an extensive-form game between m players with common prior probabilities but asymmetric information signals. we make the mild assumption that a player’s strategy can randomize over only a finite set of actions at each node. we refer the reader to the book by Mas-Colell et al.e. Thus.. an assessment Ai = (σi . the only relevant information a trader lacks is the value of the other trader’s information signal. a bluffing equilibrium must involve mixed strategies. In the remainder of this paper. In our setting. We note that the two information structures. To avoid technical measurability issues.

Two specific first-stage strategies we will study are the truthful strategy (denoted σ T ) and the null strategy (denoted σ N ). We answer the question: when two players play myopically according to a predefined sequence of play.. are the players better off on expectation by playing first or second? In particular.10 Y. a first-stage strategy σ1 for the first player is a specification. For PBEs.. 4 Profits under Predefined Sequence of Play In this section.. . where r1 . but we include a proof for completeness. µi ). we consider two players playing in sequence and define some notations of strategy and expected profits to facilitate our analysis on how the profit of the second player depends on the strategy employed by the first player. if stage t is player i’s last chance to play and µi is player i’s belief over actions of previous players. Although we are studying 2-stage games here. of a mixed strategy over moves. we investigate how sequence of play affects players’ expected profits in LMSR. The null strategy specifies that. so the definitions are more general and the first player’s strategy may not be truthful. µi ) .. we will explicitly explain players’ beliefs in proofs. r1 . we use them as a building block for our results in subsequent sections. . player i’s best response at stage t is to play truthfully by changing the market probabilities to rt = Pr(Y |ci . In a 2-stage game in which Alice plays first. for each possible signal received by that player. we use notation π B (σ1 ) to denote Bob’s expected myopic optimal profit following a first-stage strategy σ1 by Alice. regardless of its signal ci . Note that π B (σ N ) is equal to the expected myopic profit Bob would have had if he had moved first. assuming that Bob knows the strategy σ1 and can condition on it.. rt−1 . in a game in which Alice moves after Bob. . the first player leaves the market probability unchanged. Pr(N |ci . rt−1 are the market probability vectors before player i’s action. rt−1 . the same as she found it at. Definition 2 Given a game and a sequence of play. Lemma 2 In a LMSR market. We begin with the following simple result showing that the truthful strategy is optimal for a player moving only once. π B (σ T ) is the profit that Bob earns when Alice follows her myopically optimal strategy. The truthful strategy specifies that. Also. for each signal ci it receives.. In fact. our analysis apply to the first two stages of any game.. r1 . followed by Bob. because under the null strategy σN Alice does not move the market. for WPBEs. This follows from the myopic optimality of LMSR. the first player changes market probability to its posterior belief after seeing the signal. Chen et al... we use π A (σ1 ) to denote Alice’s expected profit. Likewise. leaving the beliefs implicit. .

Recall that the moves are actually probability distributions. The total profit of two consecutive moves.Gaming Prediction Markets: Equilibrium Strategies with a Market Maker 11 Proof When a player has its last chance to play in LMSR. Since the state of the world has only two exclusive and exhaustive outcomes. and. we set σ1 to dictate that Alice move to x = qx instead. and because Bob can infer x from x. The intuitive argument we use is as follows: Consider the situation in which Alice is known to be following strategy σ1 . likewise. Because the logarithmic scoring rule is strictly proper. in a market using the LMSR. and other notation is defined similarly. ω = z). ˆ We start by setting σ1 = σ1 . By ˆ ˆ ˆ ˆ definition of x. as follows. we may write expressions like x x in which ˆ ˆ x is implicitly indexed by x. Bob moves the market from x to his posterior probability ˆ ˆ P r(z|x. Alice following σ1 ). the probability of the outcome Y occuring conditioned on σ1 and the fact that Alice moved to x is exactly equal to x. thus resulting in a consistent strategy σ1 . Let qx = P r(Y |xA = x. Note that any x has ˆ ˆ ˆ ˆ a corresponding value of x. We use the term position to refer the market probability for outcome Y . thus contradicting the ˆ σ minimality of σ1 . In the second step. player i’s expected utility is maximized by truthfully reporting its posterior probability estimate given the information it has. Next. ˆ The actual position that Alice moves to can be thought of as a random variable on a sample space that includes the randomly-distributed signals Alice receives as well as the randomization Alice uses to play a mixed strategy. cB = ˆ b. We use Pr(x. whenever σ1 dictates that Alice move to x. We construct a perturbed first-stage strategy σ1 . is exactly the payoff of moving from the starting point to the end point of the second move. The market probability for outcome N is then uniquely defined because the probabilities for the two outcomes sum to 1. N }. thus. Then. it is the same as the player interacting with a logarithmic scoring rule. xz = Pr(ω = z|xA = x. Note that the second step alone yields Bob at least π B (ˆ1 ) in expectation. ˆ the event ω is a random variable that takes on two values {Y. xN = 1 − x. Bob’s actual profit is the sum of his profit from these two steps. because in σ both cases the move begins at x. σ1 satisfies the following consistency condition: For any position x that Alice moves to under σ1 . Hence. Bob moves the market to the corresponding x. for any given position x that Alice leaves the market price at. In order to use summation notation. We ˆ ˆ repeat this perturbation for each x such that x = qx . b). Then. Such a σ1 is easily constructed from any given σ1 . and show that π B (ˆ1 ) < π B (σ1 ). we define xY = x. and for which the consistency condition is failed. we decompose Bob’s response into two virtual steps. Proof Suppose that σ1 does not satisfy this condition. Lemma 3 Let σ1 be a first-stage strategy for Alice that minimizes the expression π B (σ) over all possible σ. Y and N . we provide a lemma that is very useful for obtaining subsequent theorems. z) to represent Pr(xA = x. We further ˆ ˆ argue that the first step yields a strictly positive expected profit. because for any . We use xA and xA ˆ to denote random variables that takes on values x and x respectively. Alice following σ1 ). Similarly. b. xY = x. xN = 1 − x. Consider any one position x that Alice moves to with ˆ positive probability under strategy σ1 . In the first virtual step.

b)D(p(ω|xA . π B (σ1 ) − π B (ˆ1 ) = σ x.cB ) are the conditional distributions of ω. we will prove that. z) [log xz − log xz ] ˆ Pr(z|x.cB ) ||p(ω|ˆA . Hence. this always yields a positive expected profit in the LMSR. Thus.b Pr(x. position x. b)] x Pr(z|x) [log xz − log xz ] ˆ + x. there is at least one x such that Pr(x) > 0 and x = x. z) [log Pr(z|x.b. which is x x known to be nonnegative and strictly positive when the two distributions are not the same. 4.cB ) and p(ω|ˆA . b)D(p(ω|xA .b. for any strategy σ1 that Alice chooses. if Bob is aware that Alice is following strategy σ1 .b. b. b) − log Pr(z|ˆ. and thus.b Pr(x. b) − log xz ] Pr(x.z Pr(x. We assumed that σ1 does not meet the consistency condition.cB ) ) and D(p(ˆA ) ||p(xA ) ) are relative entropy.1 Profits in CI Games When Alice and Bob have conditionally independent signals and Alice plays first. b) − log Pr(z|ˆ. b)] x Pr(x. b. z) [log Pr(z|ˆ. z) [log Pr(z|x. b. his expected payoff from following the myopic strategy (after conditioning his beliefs on Alice’s actual move) is at least as much as he could expect if Alice had chosen to play truthfully.cB ) ||p(ω|ˆA .12 Y. This contradicts the minimality σ x assumption of σ1 . Formalizing this argument. we have ˆ D(p(ˆA ) ||p(xA ) ) > 0. The first step thus moves ˆ the market from a less accurate value to the most accurate possible value given information x. Chen et al. and x p(ˆA ) and p(xA ) are the probability distributions of xA and xA respectively.cB ) ) x Pr(x)D(p(ˆA ) ||p(xA ) ) x + x where p(ω|xA .cB ) ) x Pr(x) x z + = x.cB ) ||p(ω|ˆA . b.b. b) z + x Pr(x) z = x. ˆ x D(p(ω|xA . .b xz [log xz − log xz ] ˆ ˆ Pr(x. the posterior probability of outcome Y is xY . π B (σ1 ) > π B (ˆ1 ).z = x. b) − log xz ] x ˆ − = x. we compare the payoffs π B (σ1 ) and π B (ˆ1 ) as σ follows.z Pr(x.z x. b) [log Pr(z|x.

Let us analyze Bob’s expected payoff when Alice follows this strategy σ1 . A fundamental result from information theory states that xA can reveal no more information about cB than cA can. Theorem 1 For any CI game G with Alice playing first and Bob playing second. we can assume that σ1 is consistent. z) [log yz (x. b). Pr(ci |Y ) Pr(Y |cj ) + Pr(ci |N ) Pr(N |cj ) Proof Using Bayes rule. Bob will move to a position y(x. Alice’s actual move xA (for outcome Y ) . More specifically. the equality holds if and only if σ1 = σ T . Alice following σ1 .b. ω = z. we have: π B (σ1 ) = x. However. Y ) Pr(Y |cj ) + Pr(ci |cj . and thus. Pr(ω|ci . cj ) = where ω ∈ {Y. without loss of generality. σ1 . We can view Alice’s signal cA .z Pr(x. we drop σ1 from the notation. Proof We use an information-theoretic argument to prove this result. By definition of the LMSR. b. observing Alice’s posterior probabilities is equally informative to Bob as observing Alice’s signal directly. Bob’s profit is minimized when xA reveals as much information as cA . When signals are informative and distinct. b) − log xz ] . note that xA is a (perhaps randomized) function of Alice’s signal cA alone. N ) Pr(N |cj ) Pr(ci |ω) Pr(ω|cj ) = . if player i knows player j’s posterior probabilities Pr(Y |cj ). By Lemma 4. the lower Bob’s expected profit. player i can infer the posterior probabilities conditional on both signals. For conciseness. plus another term that does not depend on Alice. The more information that xA reveals about cB . cj ) = = Pr(ω. N }. Pr(ci |ω) Pr(ω|cj ) . attains the minimum. ω) Pr(ω|cj ) Pr(ci |cj . By Lemma 3.Gaming Prediction Markets: Equilibrium Strategies with a Market Maker 13 First. cB = b). xA = x. cB = b. and thus. We now formalize this argument. retaining the notation above. and thus. Pr(ci |cj . b) = Pr(Y |xA = x. Pr(N |cj ) . Lemma 4 When players have conditionally independent signals. the truthful strategy for Alice reveals all information that is in cA . and any σ1 . ci |cj ) Pr(ci |cj ) The third equality comes from the CI condition. π B (σ1 ) ≥ π B (σ T ). we show that in CI games. and write y(x. The unit of our analysis is a particular realization of a combination cA = a. the event ω as random variables. we have Pr(ω|ci . We then show that Bob’s expected payoff can be characterized as the entropy of cB conditioned on observing xA . Pr(ci |Y ) Pr(Y |cj ) + Pr(ci |N ) Pr(N |cj ) We can now prove our result on Bob’s myopic profit in the CI setting. Bob’s signal cB . the lower the conditional entropy H(cB |xA ). As before use xz and yz to denote the probability of ω = z inferred from positions x and y respectively.

For the first term. we can write π B (σ1 ) = −H(ω. The second term is clearly independent of σ1 . Using the relation H(X. it is immediate that b Pr(b|x. and thus. cB ) + H(ω. 35]. b) log Pr(z|x. b) z Pr(z|x. cB ) − H(ω. Further. xA . xA ) = H(cB |xA ) − H(cB |ω) = [H(xA .b Pr(x. Expanding.14 Y. 17]. we have π B (σ1 ) = x. Chen et al. More specifically. b) z Pr(z|x. b) = Pr(ω = z|x. we note that H(cB |cA ) ≤ H(cB |xA ) because xA is a function of cA [11. b) log yz (x. Next. z) log xz Pr(x. Y ) = H(X) + H(Y |X). b) Pr(z|x) z b − = x. b) Pr(z|x) log Pr(z|x) z − Pr(x) x = −H(ω|xA . we note that yz (x. cB ) + H(xA . xA ) − H(xA ) = H(cB |xA ) − H(cB |ω. cB ) − H(xA )] − [H(ω. xA )] The last transformation H(cB |ω. xA . H(cB |xA ) = − =− ≥− =− =− Pr(x) x b Pr(b|x) log Pr(b|x) Pr(a|x) Pr(b|a) log b a a Pr(x) x Pr(a|x) Pr(b|a) Pr(x) x b a [Pr(a|x) Pr(b|a) log Pr(b|a)] Pr(a|x) a b Pr(x) x [Pr(b|a) log Pr(b|a)] Pr(a) a b Pr(b|a) log Pr(b|a) = H(cB |cA ). pg. z) = 1. cB ) + H(ω|xA ) where we have identified the terms with the standard definition of conditional entropy of random variables [11. b). pg. . xA ) = H(cB |ω) follows because cB is conditionally independent of xA conditioned on ω. knowledge of xA does not alter its conditional distribution or its conditional entropy.b Pr(x) x Pr(b|x.

Hence. Bob is better off when the Bob-Alice sequence is used. Proof We compare Bob’s expected profit in Alice-Bob and Bob-Alice cases when both players play myopically. ax = ax because signals are distinct. the condition is satisfied with Pr(cA = a|xT = xa ) = 1. when cA = a. Alice moves to Pr(Y |cA = a) with probability 1 under σ1 . Corollary 1 In CI games. the equality holds only when σ1 is a strategy such that for any x there exists some a that satisfies Pr(cA = a|xA = x) = 1. This means that Pr(Y |cA = ax ) = Pr(Y |cA = ax ) when x = x . Given any signal a that Alice might see. and let xT denote the corresponding random variable. A All that remains to be shown is that there is no other strategy σ1 = σ T that satisfies the condition. cA and cB are unconditionally independent which contradicts with the CI condition by Lemma 1. it is better off for a player to play first than second. The equality holds only when the distribution of cB |a is the same as the distribution of cB |a for all a. If with some positive probability xA = x = Pr(Y |cA = a). it must be that ax = ax . for all a ∈ CA . Because signals are distinct. or for any x there exists some a such that Pr(cA = a|xA = x) = 1. strategy σ1 is the truthful strategy σ T . for any strategy σ1 that Alice chooses. Thus. According to Lemma 3. Pr(Y |xA = x. x ∈ XA . According to Theorem 1. when two players play myopically according to a predefined sequence of play. we show that when cA = a. For some strategy σ1 . Suppose the contrary. π B (σ N ) > π B (σ T ). Hence. Next. then. we will prove that. The following corollary immediately follows from Theorem 1 and answers that question we proposed at the beginning of the section for CI games. We first show that when x = x and x.2 Profits in I Games When Alice and Bob have unconditionally independent signals and Alice plays first. Bob’s expected profit in Alice-Bob case equals π B (σ T ). while his expected profit in Bob-Alice cases equals π B (σ N ). 4. Then xa = A Pr(Y |cA = a). it must be the case that xA = Pr(Y |cA = a) with probability 1 according to strategy σ1 . Alice follows σ1 ) = x. cA = ax = a. If cB |a and cB |a have the same distribution for all a and a . the value of xa uniquely corresponds to signal a. if Bob is aware that Alice is following strategy σ1 . his expected payoff from following the myopic strategy (after conditioning his beliefs on Alices actual move) is at least as much as he could expect if Alice had chosen to play the null strategy σ N . for any x ∈ XA there exists ax such that Pr(cA = ax |xA = x) = 1. Hence. Thus. which gives a Pr(Y |cA = a ) Pr(cA = a |xA = x) = x =⇒ Pr(Y |cA = ax ) = x. let XA denote the value space of xA . because Pr(cA = ax |xA = x ) = 1.Gaming Prediction Markets: Equilibrium Strategies with a Market Maker 15 The inequality follows from the strict convexity of the function x log x. let xa be the position that Alice would move to if she followed the truthful strategy. Hence. The truthful strategy σ T satisfies the above condition. and contradiction arises. . a ∈ CA .

i. Pr(ω = N ) . without loss of generality. if the market starts with the prior probabilities for the two outcomes. cB = bj ). and pF the probability that Alice has signal cA = ai and sets xA = H equals pE uEi +uFi . cB = bj ). and the same set of remaining points R. π B (σ1 ) = (pE − pF )[ +pF [ j j Pr(cB = bj )D(rα || j j Pr(cB = bj )rα )] j j Pr(cB = bj )D(rα || j Pr(cB = bj )D(rβ || j Pr(cB = bj )rα )] j Pr(cB = bj )rβ )] j j +pF [ j + remaining profit over R . In this case. Alice mixes over E and H with probability pE − pF and 2pF respectively. Chen et al. We will show that σ1 must involve Alice not moving the market probability at all. Define uEi and uFi as the probability (under σ1 ) that Alice has signal cA = ai and sets xA = E and xA = F respectively. it can only have a single value. Let pE be the probability that Alice plays E and similarly pF be the probability that Alice plays F . writing pE as (pE − pF ) + pF to facilitate comparison with π B (σ1 ). Let xA be the random variable of Alice’s move for outcome Y under strategy σ1 . then σ1 would have support over a set of points: at least two points E. With these definitions we E F can define the myopic move of Bob given that market is at E and cB = bj as β α rj = αi Pr(Y |cA = ai . we have π B (σ1 ) ≥ π B (σ N ). Define αi = Pr(cA = ai |xA = u u E) = pEi and βi = Pr(cA = ai |xA = F ) = pFi . and consider a new strategy σ1 over points E. Similarly rj is defined as the myopic move of Bob given the market is at F . 1 − rj respectively. let H = (E + F )/2 be the midpoint of E and F . pF pE uEi +uFi 2pF = 1 uEi 2 pE + 1 u Fi 2 pF = We can now define the myopic move of Bob given the market is at H γ and cB = bj as rj = γi Pr(Y |cA = ai . Proof Assume that there is a strategy σ1 for Alice that minimizes π B (σ1 ) over all possible first-stage strategies. Now. −p the probability that Alice has signal cA = ai and sets xA = E is pEpE F uEi . we use rγ to j i represent the vector for two outcomes.16 Y. H. and perhaps a set of other points R.e. F . We first argue that xA is deterministic. r0 = Pr(ω = Y ). The equality holds if and only if σ1 = σ N . i. Hence. Without loss of generality let pF < pE . We now characterize π B (σ1 ) as follows.e. we show that we can construct a strategy σ1 such that π B (σ1 ) < π B (σ1 ) by “mixing” point E and another point H. 1 − rj and rj . Suppose the contrary. We use rα and rβ to denote the probability j j β β α α vectors for two outcomes rj . i As before we can define γi = Pr(cA = ai |H) = αi +βi 2 . Theorem 2 For any I game G with Alice playing first and Bob playing second and any first-stage strategy σ1 . According to Lemma 3. Under σ1 . Similarly. we can assume that σ1 is consistent.

Therefore the strategy that minimizes the expected payoff of Bob is for Alice to report Pr(ω = Y ). The following corollary immediately follows from Theorem 2 and answers the question we proposed at the beginning of the section for I games. This again contradicts the fact that σ1 minimizes π B (σ1 ). x. as Bob will always make a positive payoff in expectation if he moves from x to Pr(ω = Y ). 2 From the definitions of the myopic moves given the market states. Suppose that the point in the support. .Gaming Prediction Markets: Equilibrium Strategies with a Market Maker 17 We also characterize π B (σ1 ) as: π B (σ1 ) = (pE − pF )[ +2pF [ j j Pr(cB = bj )D(rα || j j Pr(cB = bj )rα ) j j Pr(cB = bj )D(rγ || j Pr(cB = bj )rγ ) j + remaining profit over R rα +rβ j j . Thus. because the market starts with Pr(ω = Y ). the strategy does not reveal any information to Bob. he would have a larger payoff overall if Alice left the market at x instead of Pr(ω = Y ). Thus. is such that x = Pr(ω = Y ). the strategy must have only one point in its support. there always exists a strategy σ1 such that π B (σ1 ) < π B (σ1 ). note that rγ = j This means that π B (σ1 ) can be bounded as : j π B (σ1 ) = (pE − pF )[ +2pF [ j Pr(cB = bj )D(rα || j bj )D(rγ || j j Pr(cB = bj )rα ) j j Pr(cB = Pr(cB = bj )rγ ) j Pr(cB = bj )rα ) j j + remaining profit over R = (pE − pF )[ Pr(cB = bj )D(rα || j j +2pF [ j Pr(cB = bj )D( rα + rβ j j 2 || Pr(cB = bj ) j rα + rβ j j 2 ) + remaining profit over R < (pE − pF )[ Pr(cB = bj )D(rα || j j Pr(cB = bj )rα )] j j +pF [ j Pr(cB = bj )D(rα || j Pr(cB = bj )D(rβ || j Pr(cB = bj )rα )] j j +pF [ j Pr(cB = bj )rβ )] j j + remaining profit over R = π B (σ1 ) The last inequality follows from the strict convexity of relative entropy under the general informativeness condition. This means that for any strategy of Alice that minimized π B (σ1 ). for any strategy σ1 with two or more points in its support. Therefore we have shown that π B (σ1 ) ≥ π B (σ N ). it is equivalent to her not trading at all in the first stage. Therefore. However. Only when σ1 = σ N the equality holds.

Pr(N |cA ) in the second stage. π B (σ T ) ≥ π B (σ N ). Hence. Bob’s belief µB gives . cB ) in the third stage. we examine the following strategic question: Suppose that each player will get to report only once. That is. Pr(N |cA .1 Sequence Selection Equilibrium of CI Games Consider the sequence selection game. Bob’s expected profit in Alice-Bob case equals π B (σ T ). we define a simple 2-player sequence selection game. Theorem 3 When Alice and Bob have conditionally independent signals in LMSR. Proof We compare Bob’s expected profit in Alice-Bob and Bob-Alice cases when both players play myopically. where ai is the i-th element of CA . Chen et al. After getting their signals. and assume that Alice and Bob have conditionally independent signals. when two players play myopically according to a predefined sequence of play. r0 = Pr(ω = Y ). i. while his expected profit in Bob-Alice cases equals π B (σ N ). . We denote x = x1 . we provide equilibria of the sequence selection game under CI setting and I setting respectively.e. xnA as the vector of Alice’s possible posteriors for the outcome Y . Similarly.. – Alice changes the market probability to Pr(Y |cA ). Proof Since we use the PBE concept. which one should it choose? To answer the question. Bob gets a signal cB . 5 Sequence Selection Game In this section. If a player has a choice between playing first and playing second after observing its signal. Pr(ω = N ) .. The following theorem gives a PBE of the sequence selection game when the CI condition is satisfied. xi = Pr(Y |cA = ai ). – Bob changes the market probability to Pr(Y |cA . assume that xi < xj iff i < j. Then. it is better off for a player to play second than first. In the third stage. Off the equilibrium path. Alice gets a signal cA . According to Theorem 2. The selected player then changes the market probabilities as they see fit in the second stage. the market closes and the true state is revealed. Without loss of generality. 5.. In the rest of this section. Suppose that Alice and Bob are the only players in the market. when Alice selects Bob to be the first player and the initial market probability is r0 . Alice chooses who— herself or Bob—plays first. we first describe Bob’s belief µB for the offequilibrium paths. Alice and Bob play the sequence selection game as follows. the other player gets the chance to change the market probabilities. cB ). Bob is better off when the Alice-Bob sequence is used. Corollary 2 In I games that start with the prior probabilities for the two outcomes. In the first stage. at a PBE of the sequence selection game – Alice selects herself to be the first player in the first stage.18 Y.

select herself as the first player. cB . mix between selecting herself as the first player and selecting Bob as the first player. To make her sequence selection in the first stage. Alice essentially compares her expected utilities conditional on her own signal in the Alice-Bob and BobAlice subgames. 0 – If cA = ai or cA = ai+1 and xi < rY < xi+1 . µB ) equals one of the following 0 1. 0 3. µB ) = x1 . and ω = z. cB ). Pr(N |a1 . When cA = ai . cB ) when xnA < rY ≤ 1. Then. Pr(Y |r . Pr(N |ˆr0 . Pr(N |ˆr0 ) = r0 . µB ). we use Pr(ak ) . for case 1. z) log l. Pr(N |cA ) in the second stage. Pr(N |cA . Pr(Y |anA . Pr(N |r0 . Bob will further changes the market probabilities to Pr(Y |cA . Pr(Y |r0 . 19 Bob’s belief µB is consistent with Alice’s off-equlibrium mixed strategy: 0 – If cA = ai and xi = rY .z Pr(ω = z|cA = ak ) . µB ). z) represents the joint probability of cA = ak . Pr(Y |r0 . µB ) in the second stage. with probability α select Bob as the first player and (1 − α) select herself as the first player. cB ) when x1 ≤ rY ≤ xnA . bl . µB ) = xnA . if Bob is selected as the first player. 0 0 – If 0 ≤ rY < x1 . select Bob as the first player. I UA = 1 Pr(cA = ak ) Pr(ak . bl . cB . believing that prices in the second stage are Alice’s posteriors. cB ) in the third stage. a a ar0 does not necessarily belong to Alice’s signal space CA . select Bob as the first player. The mixing probabilities α and β 0 satisfy that Pr(Y |select Bob as the first player) = rY . by Lemma 2. a a 0 2. – For all other cases. Pr(Y |r0 . cB = bl ) a where bl represents the l-th signal element in CB . cB ). Pr(N |r0 . and Pr(ak . can calculate his posteriors based on both Alice’s and his own signals. On the contrary. Pr(N |anA . with probability β select Bob as the first player and (1 − β) select herself as the first player. both of them will truthfully reveal all information that they have given their beliefs when it’s their turn to play. she will change the market probabilities to Pr(Y |cA ). When cA = ai+1 . cB .Gaming Prediction Markets: Equilibrium Strategies with a Market Maker 0 0 – If x1 ≤ rY ≤ xnA . cB = bl . and Pr(ω = z|ˆr0 ) a (3) (4) II UA = 1 Pr(cA = ak ) Pr(ak . Bob. cB = bl ) . Pr(Y |ˆr0 . Let UA denote Alice’s expected profit conditioned on her signal when the Alice-Bob subgame is picked. select Bob as the first player.z Pr(ω = z|cA = ak . Pr(Y |a1 . suppose Alice has signal ak . 0 – If xnA < rY ≤ 1. Pr(ω = z|ˆr0 . According to Lemma ˆ 4. he will change the market probabilities to Pr(Y |r0 . 0 – If cA = a1 and 0 ≤ rY < x1 . Let ar0 denote ˆ a fictitious signal realization of Alice that satisfies Pr(Y |ˆr0 ). 0 – If cA = anA and xnA < rY ≤ 1. Hence. Each player has only one chance to change the market probabilities. I Without loss of generality. cB . If Alice is the first to play. µB ) = rY . cB ). bl . For conciseness. II UA denotes Alice’s expected profit conditioned on her signal when the Bob-Alice subgame is picked. cB ) when 0 ≤ rY < x1 . z) log l. cB ).

We obtain I II UA − UA = 1 Pr(ak ) Pr(ak . D(p||q) ≥ 0. p(cB |ak ) is different from p(cB |ˆr0 ) and hence a I II UA − UA is strictly greater than 0. z) log l. Chen et al. Alice’s expected profit conditioned on her signal for the Bob-Alice subgame becomes II UA = 1 Pr(cA = ak ) Pr(ak . We only need to prove one of them. cB = bl ) . . Pr(N |ak ) = r .z Pr(ω = z|cA = ak . a where p(cB |ak ) is the probability distribution of Bob’s signal conditional on signal ak and p(cB |ˆr0 ) is the probability distribution of Bob’s signal conditional a on the fictitious signal ar0 . Alice does not want to deviate from selecting herself as the first player. Without loss of generality. We thus have UA − UA ≥ 0. z) log l. bl . z) log l. cB = bl ) I while UA remains the same as in (3). at a PBE the player will always choose to play first in CI games no matter what signal it gets. consider case 0 2 where xnA < rY ≤ 1. z) log l. Theorem 3 further strengthens this result and states that if a strategic player have a choice. cases 2 and 3 are symmetric. The equality holds only II I when distributions p and q are the same. bl ) Pr(ˆr0 ) a Pr(ˆr0 . When 0 Pr(Y |ak ). The difference in expected profits for Alice in the two subgames is: I II UA − UA = 1 Pr(ak ) 1 Pr(ak ) Pr(ak .z Pr(z|ak ) + Pr(z|anA ) Pr(z|ak ) log z Pr(z|anA ) II − UA 0 rz = D p(cB |ak ) ||p(cB |anA ) + Pr(z|ak ) log z Pr(z|anA ) . 0 The second term in the above expression is positive because rY > Pr(Y |anA ) ≥ Corollary 1 in Section 4 indicates that if myopic players follow predefined sequence of play then on expectation a player is better off by playing first than second in CI games. bl ) a Pr(z|ˆr0 ) Pr(z|ak . UA − UA > 0.z Pr(z|ak ) Pr(z|ˆr0 . (7) Pr(ω = z|cA = anA . bl . to represent Pr(cA = ak ) and similarly for other terms. Hence. Since the market only has two exclusive outcomes. bl . bl . bl ) Pr(ak ) a (5) = = Pr(ak . The second equality comes from Bayes rule and ˆ the conditional independence of signals.20 Y. 0 rz I II Pr(Y |ak ).z l Pr(bl |ak ) log Pr(bl |ak ) Pr(bl |ˆr0 ) a (6) = D p(cB |ak ) ||p(cB |ˆr0 ) . bl ) a Pr(ak .

bl ) = l Pr(bl ) Pr(z|ak . Theorem 4 If the market starts with the prior probabilities for the two outcomes. assume Alice has signal ak . The following theorem gives a PBE of the sequence selection game when the I condition is satisfied. according to Lemma 2.Gaming Prediction Markets: Equilibrium Strategies with a Market Maker 21 5. Pr(z) = l Pr(bl ) Pr(z|bl ). her expected profit conditioned on her sigII nal. Thus. Pr(ω = N ) . Without loss of generality. cB ) in the third stage.z Pr(cb = bl |ak ) Pr(ω = z|ak . bl ). bl ) Pr(ω = z|bl ) (9) = l Pr(cb = bl |ak )D(p(ω|ak . denoted as UA . we need to prove that Alice does not want to deviate from selecting herself as the first player. i. cB ). The difference in expected profits for Alice is: I II UA − UA = D(p(ω|ak ) ||p(ω) ) − l Pr(bl )D(p(ω|ak . – Bob changes the market probability to Pr(Y |cB ). her best response in the second stage is to report truthfully. By Bayes rule.e. bl ) log Pr(ω = z|ak .bl ) ||p(ω|bl ) ) Because signals are independent. Bob’s best responds in the second stage is to truthfully report the posterior probability conditioned on his own signal and Alice will report truthfully conditioned on both signals in the third stage. this action does not reveal any information to Bob. Hence. Pr(z|ak ) = l Pr(bl |ak ) Pr(z|ak . the expected profit conditioned on her signal in this I case. – Alice changes the market probability to Pr(Y |cA . Knowing Alice’s strategy. If Alice selects herself as the first player.2 Sequence Selection Equilibrium of I games In this part. Pr(N |cB ) in the second stage. is I UA = z Pr(ω = z|ca = ak ) log Pr(ω = z|ca = ak ) = D(p(ω|ak ) ||p(ω) ) (8) Pr(ω = z) If Alice selects Bob as the first player. at a PBE of the sequence selection game – Alice in the first stage selects Bob to be the first player. we consider the alternative model in which players have independent signals.bl ) ||p(ω|bl ) ) l = D( l Pr(bl )p(ω|ak .bl ) ||p(ω|bl ) ) ≤ 0. denoted as UA is II UA = l. r0 = Pr(ω = Y ). and Alice and Bob have independent signals. we have Pr(bl |ak ) = Pr(bl ). Hence. . Pr(N |cA .bl ) || l Pr(bl )p(ω|bl ) ) − Pr(bl )D(p(ω|ak . Proof If Alice always selects Bob as the first player.

Pr(N |cA ) in the first stage. Bob believes that Alice reports truthfully. – Bob in the second stage changes the market probability to r2 = Pr(Y |cA . cB ) . rN be the market probabilities in stage t. Alice can take another action in the third stage. Pr(N |ak . Lemma 5 characterizes the equilibrium strategy of Alice in the third stage. Built on our equilibrium result for the sequence selection game. Then. Alice behaves truthfully and fully reveals her information. according to Lemma 2 Bob plays truthfully and fully reveals his information.22 Y. Alice does not want to deviate from selecting Bob as the first player. The inequality follows from the convexity of relative entropy [11]. her best response is to report her true posterior. at a PBE the player will always choose to play second in I games no matter what signal it gets. t t Let rt = rY . If Alice selects herself to be the first players. when Alice has signal ak and Bob has signal bl . Since Bob only gets one chance to play. Chen et al. bl in the third stage. Proof This is proved by applying Lemma 2 to both Bob and Alice. Observing Alice’s action. 6 The Alice-Bob-Alice Game We now consider a 3-stage Alice-Bob-Alice game. at a PBE of the 3-stage Alice-Bob-Alice game – Alice changes the market probability to r1 = Pr(Y |cA ). cB ). . Theorem 4 further strengthens this result and states that if a strategic player have a choice. Because the third stage is Alice’s last chance to change the probabilities. Theorem 5 When Alice and Bob have conditionally independent signals in LMSR. At a WPBE. including her own signal and Bob’s signal inferred from Bob’s action in the second stage. Pr(N |cA . Lemma 5 In a 3-stage Alice-Bob-Alice game in LMSR. the market closes and the true state is revealed. at a WPBE Alice changes 3 3 the market probabilities to r3 = rY . Alice and Bob act as if they know each other’s strategy. Alice may change the market probabilities however she wants in the first stage. where Alice plays in the first and third stages and Bob plays in the second stage. Theorem 5 describes a PBE of the Alice-Bob-Alice game. Corollary 2 in Section 4 indicates that if myopic players follow predefined sequence of play then on expectation a player is better off by playing second than first in I games. bl ). The strategies and beliefs off the equilibrium path are straightforward. beliefs are consistent with strategies. Thus. Bob may change the probabilities in the second stage. rN = Pr(Y |ak . In the second stage. 6.1 Truthful Equilibrium of CI Games We study the PBE of the game when Alice and Bob have conditionally independent signals.

Pr(N |ak .Gaming Prediction Markets: Equilibrium Strategies with a Market Maker 23 – Alice does nothing in the third stage. When cA = ai . When cA = ai+1 . µB ). the expected profit that Alice gets from case 1 is the same as what she gets from the following sequence of actions: (a) Alice changes market probabilities to r1 that are different from her posteriors in the first stage. Pr(Y |r . Bob’s belief µB gives 1 1 – If x1 ≤ rY ≤ xnA . where ai is the i-th element of CA . When Alice does not deviate: Alice changes market probabilities to her true posteriors Pr(Y |ak ). and Alice plays a best response according to Lemma 5 by changing market probabilities to Pr(Y |ak . cB . µB ) = x1 . cB . xi = Pr(Y |cA = ai ). Alice does nothing in the third stage. 1 – If cA = a1 and 0 ≤ y < x1 . When Alice changes market probability to r1 in the first stage. Pr(N |ak ) in the first stage. xnA as the vector of Alice’s possible posteriors for the outcome Y . with probability α change market 1 probability to rY = y and (1 − α) play truthful betting. We denote x = x1 . Pr(N |r1 . change market probability to rY = y. cB ). Without loss of generality. By Lemma 2. is exactly the payoff of moving from the starting point to the end point of the second move. cB ) in the second stage. Pr(Y |r1 . Consider the two cases: 1. 1 – If cA = anA and xnA < y ≤ 1. 1 1 – If 0 ≤ rY < x1 . report truthfully. µB ) in the second stage. change market probability to rY = y. We show that Alice does not want to deviate by changing market probabilities to r1 = Pr(Y |cA ). – For all other cases. Bob does not want to deviate from truthful betting in the second stage given that Alice truthfully reports her posteriors in the first stage. The mixing probabilities α and β satisfy that Pr(Y |rY = y) = y. µB ) = xnA . Bob changes market probabilities to Pr(Y |r1 . assume that xi < xj iff i < j. 1 – If xnA < rY ≤ 1. Pr(N |ak ) . That is. The total payoff of two consecutive moves. Pr(Y |r1 . cB ). (b) A sequence selection game starts with initial market probabilities r1 .. 1 When observing a rY = y = xi for all 1 ≤ i ≤ nA . with 1 probability β change market probability to rY = y and (1 − β) play truthful 1 betting. Bob’s belief µB is consistent with Alice’s off-equlibrium mixed strategy: – If cA = ai or cA = ai+1 and xi < y < xi+1 ... Without loss of generality. Pr(N |ak . in a market using the LMSR. (c) Alice changes . Alice selects herself to be the first player. mix between truthful betting 1 and reporting rY = y. When Alice deviates: Alice changes market probabilities to r1 that is different from Pr(Y |ak ). Pr(N |cA ) . cB ) in the third stage. 2. assume that Alice has signal cA = ak . Bob changes the probabilities to Pr(Y |ak . Proof Bob’s belief µB is similar to that of the sequence selection game as described in the proof for Theorem 3. Hence. We compare Alice’s expected profits conditional on her signal in these two cases with the aid of the sequence selection game. µB ) = rY . .

Alice does not want to deviate from truthful betting. We define . Pr(N |r1 . cB ) . (d) Bob changes market probabilities to Pr(Y |ak . According to Theorem 1. market probabilities to Pr(Y |ak ).1: change market probability to r1 = Pr(Y |cA = ak ).2 Nonexistence of Truthful Equilibrium in I Games We study the equilibrium of the game when Alice and Bob have independent signals and the general informativeness condition holds. where π (σ) is Bob’s expected myopic optimal profit following a first round strategy σ as defined in Section 4.24 Y. cB . A natural question to ask. Similarly. (c’)Bob changes market probabilities to Pr(Y |r1 . 6. and (d) is greater than or equal to that from (b’). However. Alice’s expected profit from (b). Theorem 6 says that there are not when players have informative and distinct signals. 0 rz which is fixed given r0 and the joint probability distribution P. µB ) . truthful betting is the only WPBE of the game. Pr(N |ak ) . π B (σ) ≥ π B (σ T ) and the equality holds only when σ = σ T . the truthful betting PBE is the unique WPBE of the 3stage Alice-Bob-Alice game. Pr(N |ak . and (d’). Then. the total expected profit of Alice and Bob from the game is π AB = k. because Alice can simply deviate to the truthful strategy and get higher expected profit. we show that the truthful strategy in this setting is not a WPBE. The truthful strategy for Alice is the same as that in Section 6. Proof According to Lemma 5. at a WPBE the market probability in the third stage always reveals information of both Alice and Bob. Hence. (b’) A sequence selection game starts with initial market probabilities r1 . after we know that Alice being truthful is a PBE. (d’) Alice changes market probabilities to Pr(Y |ak . Alice has an incentive to deviate from the truthful strategy on her first trade in the market. Pr(N |cA = ak ) in the first stage. cB = bl . Thus. for any WPBE A strategy σ of Alice.z Pr(cA = ak . Alice’s expected profit from (a) is the same as that from (a’). In particular. cB = bl ) . (c). Let σ be AlA ice’s strategy at a WPBE. Theorem 6 When Alice and Bob have conditionally independent signals that are informative and distinct. cB . Pr(N |ak . µB ). Alice selects Bob to be the first player. cB ). the expected profit that Alice gets from case 2 is the same as what she gets from the following sequence of actions: (a’) Alice changes market probabilities to r1 that are different from her posteriors in the first stage. cB ) . Chen et al. any strategy σ such that πσ < π AB − π B (σ T ) can not be a WPBE strategy. with equality holds T A only when σ = σ . is whether there are other equilibria with Alice bluffing in the first stage. Hence. Alice’s expected profit by following σ is πσ = AB B B π − π (σ). cB ). But according to Theorem 3. (c’). at the equilibrium πσ ≤ π AB − π B (σ T ). ω = z) log Pr(ω = z|cA = ak .l. Thus.

Alice’s first trade is a move from r1 to ˜1 . where ak = ak is another signal in CA . ˜ ˜ ˜ Since the total payoff of two consecutive moves in LMSR is exactly the payoff of moving from the starting point to the end point of the second move. a Pr(Y |ak ) Pr(N |ak ) As Bob has a probability of Pr(bl |ak ) of seeing bl as his signal. her expected change in profit over following the truthful strategy. Theorem 7 If players have independent signals that satisfy the general informativeness condition. the expected score of Alice’s second trade in the market is: – » Pr(N |ak . bl ) Pr(Y |˜k . and bluffs. bl ) to r3 = a a Pr(Y |ak . and reacts accordingly. Therefore. Proof For conciseness. bl ) Pr(Y |ak .bl ) ) a = D(p(ω|ak ) ||p(ω|˜k ) ) a . we use Pr(ak ) to represent Pr(cA = ak ) and similarly for other terms. we eliminate the r0 to r1 move in our comparison analysis and treat the myopic move as if it had zero expected profit and analyze the bluffing strategy as a move from r1 to ˜1 . bl ). by the definition of LMSR r her expected score from the first move is: Pr(Y |ak ) log Pr(N |˜k ) a Pr(Y |˜k ) a + Pr(N |ak ) log = −D(p(ω|ak ) ||p(ω|˜k ) ). bl ) Pr(bl |ak ) Pr(Y |ak . If Alice observes her signal. Due to the overlap in r the expected score in the truthful strategy and the bluffing strategy for Alice. We have l Pr(bl |ak )D(p(ω|ak . Alice’s total expected payoff conditional on her signal is the sum of the above two expressions.bl ) ||p(ω|˜k . Below we show that the expected profit change of deviating from the truthful strategy is greater than zero.bl ) || l Pr(bl |ak )p(ω|˜k . is ∆= l Pr(cB = bl |cA = ak )D(p(ω|ak . which gives ∆ in our theorem. Pr(N |cA = ak ) . bl ) a Pr(N |˜k . r Lemma 6 Assume that Bob expects Alice to play truthfully. Pr(N |˜k .bl ) ) a X l Thus.bl ) ) a l ≥ D( Pr(bl |ak )p(ω|ak . bl ) log + Pr(N |ak . Alice will move the market from r2 = Pr(Y |˜k .Gaming Prediction Markets: Equilibrium Strategies with a Market Maker 25 the bluffing strategy for Alice as: change market probability to ˜1 = Pr(Y |cA = r ak ).bl ) ||p(ω|˜k . bl ) a lX = Pr(bl |ak )D(p(ω|ak . truthful betting is not a WPBE for the Alice-Bob-Alice game.bl ) ) − D(p(ω|ak ) ||p(ω|˜k ) ). we may now consider a bluffing strategy by Alice during the first trade as a move from r0 to r1 immediately followed by a move from r1 to ˜1 . bl ) with probability Pr(bl |ak ).bl ) ||p(ω|˜k . thus Alice has an incentive to deviate from the truthful strategy. cA = ak . a a where p(ω|·) denotes the probability distribution of ω conditional on signal realizations. bl ). Proof Lemma 6 gives the expected profit change of Alice if she were to deviate from truthful betting. Pr(N |ak .

we can compute Alice’s expected profit for playing TB.2. Theorem 8 Consider the Alice-Bob-Alice LMSR coin-flipping game. Truthful betting is not a WPBE strategy. if he has heads. 1 (10) If Bob has tails he sets the probability of HH to zero. and otherwise play H. Similarly. Chen et al. (12) . (HT|TH|TT)}. In our model.TH. and the law of total probability. The equality follows from the independence of the two signals. We again consider an Alice-Bob-Alice game structure.1 Explicit Bluffing Equilibria in a Restricted Game We now introduce a simple model of independent signals and show that bluffing can be an equilibrium. Bob’s best response is. Assuming such a strategy for Bob. for each outcome in {HH. her profit for moving the probability from p0 to p or 0. The inequality follows from the convexity of relative entropy [11]. Thus. Alice and Bob each see an independent coin flip and then participate in a LMSR prediction market with outcomes corresponding to whether or not both coins came up heads. 6. ω ∈ {HH. Alice’s expected profit for always pretending to have heads in the first stage is: p2 log p p0 + log 1 x 1−p + p(1 − p) log 1−p0 + (1 − p)p log 1−p 1−p0 + log 1−0 1−x + (1 − p)2 log 1−p 1−p0 . A WPBE in this game has Alice p ˆ play TB with probability t = 1 + 1−(1−p)−1/p (1−p) . plus her profit for moving the probability to 0 or 1 from where Bob moves it. ( ) Proof To show that playing TB with probability t is an equilibrium. Now restrict Alice’s first stage strategies to either play ˆ truthful betting (TB) or as if her coin is heads (H). We note that by the general informativeness condition the inequality is strict. We have p2 log p p0 + log 1 x + (1 − p)2 log 1−p + p(1 − p) log 1−p0 + (1 − p)p log 1−0 1−p0 1−0 1−p0 (11) where x is Bob’s posterior probability when he has heads and Alice appears to have heads. This is done by computing. we first compute Bob’s best response to such a strategy and then show that Bob’s strategy makes Alice indifferent between her pure strategies. ∆ > 0.TT}. to set the probability of HH to the probability that Alice has heads given ˆ that she bets as if she does (we denote Alice betting as if she had heads as H): ˆ Pr(HH | HH) = ˆ Pr(HH | HH) Pr(HH) ˆ ˆ Pr(HH | HH) Pr(HH) + Pr(HH | TH) Pr(TH) 2 1·p = 1 · p2 + (1 − t)(1 − p)p = 1 − (1 − p) p . Pr(bl |ak ) = Pr(bl ).HT. Therefore. where the probability of heads is p.26 Y.

it is a PBE at which any player reports truthfully at its first chance to play. there is a bluffing WPBE. Suppose the contrary. Alice is indifferent between truthfulness and bluffing when Bob expects her to play TB with probability t. We can combine the signals of those players standing in between as one signal and treat those players as one composite player. The game from stage t to stage t becomes an Alice-Bob-Alice game. Proof Given that every player believes that all players before it act truthfully. Inferring recursively.1 Truthful Equilibrium of CI games We first consider the CI setting where players have conditionally independent signals. 7 Finite-Player Finite-Stage Game We extend our results for the Alice-Bob-Alice game to games with a finite number of players and finite stages in LMSR. pretend to have seen heads regardless of her actual information. and does not play in stages between t and t . suppose the player is Alice and she bluffs at some stage t.Gaming Prediction Markets: Equilibrium Strategies with a Market Maker 27 Since (11) and (12) are equal when x is set according to (10). If signals are informative and distinct. Alice bluffing at stage t contradicts with the uniqueness of the WPBE of the Alice-Bob-Alice game. Select the last player who bluffs in the sequence. The following theorem characterizes the PBE of the game. at a PBE of the game all players report their posterior probabilities in their first round of play. the truthful betting PBE is the unique WPBE of the game. 7. . Theorem 6 states that the truthful betting PBE is the unique WPBE of the Alice-Bob-Alice game. it will truthfully report its posterior probabilities by Lemma 2. with 1 − t probability. Because signals are conditionally independent. we prove the theorem recursively using backward induction. finite-stage game with LMSR. Alice must play truthfully at some stage t > t + 1. Then. If signals are informative and distinct. Alice should. there are other players standing in between its second to last chance to play and its last chance to play. Combining signals of players between t and t as one composite signal. Each player can change the market probabilities multiple times and all changes happen in sequence. This implies that truthful betting PBE is also a unique WPBE for the finite-player finite-stage game. If it’s player i’s last chance to play. The game becomes an Alice-Bob-Alice game for player i and at a PBE player i reports truthfully at its second to last chance to play according to Theorem 5. that is. if players have conditionally independent signals. Theorem 9 In the finite-player. If it’s player i’s second to last chance to play. It is therefore in equilibrium for Alice to play TB with probability t. the signal of the composite player is conditionally independent of the signal of player i. Without loss of generality.

Proof If there are m players. for example Alice.2 No Truthful Equilibrium in I Games For the I setting where players have independent signals. Hence. Proof Consider the last two trades of a player who plays more than once. Because signals are independent. the result for finite-player finite-stage game directly follow from our result on the Alice-Bob-Alice game. If signals are informative and distinct. suppose the last time for a player to bluff is in stage i. 8 Finite-Player Infinite-Stage Games In this section. truthful betting is not a WPBE strategy. Chen et al. we know that Alice is better off by bluffing at stage t. Knowing that other players are playing truthfully. as is shown in the following theorem. If t and t > t + 1 are the last two stages for Alice to trade and all other players play truthfully. According to Theorem 5 player i would prefer to report truthfully. 7. at a finite PBE all players truthfully report their posterior probabilities at their first chance of play. The game from stage t to stage t can be viewed as an Alice-Bob-Alice game. finite-stage game with LMSR. if players have independent signals that satisfy the general informativeness condition. we analyze a model in which the number of potential moves can be infinite. A player may decide to have one more trade at any time before the market closes. truthful betting reveals all information after m stages.28 Y. a player who plays in stage i < m does not want to deviate to bluffing in stage i and come to trade again in stage m + 1. and there is at least one player who trades more than once. By Theorem 7. If there exists a finite equilibrium that reveals all information at stage t. truthful betting is not a WPBE. Theorem 10 In the finite-player. because the game between stages i and m + 1 can be viewed as an Alice-Bob-Alice game. 8. This is motivated by the observation that there is often no pre-defined last stage for a player. Alice’s signal is independent of the composite signal. the truthful betting equilibrium is the unique finite WPBE of the game.1 Truthful Betting Equilibrium in CI Games Games with conditionally independent signals remain to have a unique equilibrium. Theorem 11 In a potentially infinite-stage game with conditionally independent signals. We extend our results for finite-stage games to infinite-stage games and examine the equilibrium for CI and I settings respectively. we can combine signals of all players who trade between t and t as a composite signal possessed by a composite agent. truthful betting PBE. then it must report truthfully in some .

9.5. Alice and Bob without loss of generality. There is not finite equilibrium. perhaps resulting in an incentive to play the truthful strategy. They play in stages t − 1 and t respectively.3. Then.2 Nonexistence of Finite Equilibrium in I Games Theorem 12 In an infinite-stage game with independent signals that satisfy the general informativeness condition. Proof For contradiction.2 and introduce the concepts of complementarity and substitution of signals in Section 9. it has an incentive to not fully divulge this information when it may play multiple times in the market. a player will always have some private information revealed by its signal. If signals are informative and distinct. according to Theorem 6. when players have independent signals. We propose a discounted logarithmic market scoring rule and show that the I games can converge to full information aggregation exponentially. 8.2. Then. we propose the discounted LMSR market.1. We first introduce the discounted LMSR mechanism in Section 9.1 Discounted LMSR One way to address the incentives that traders with independent signals have to bluff in a market using the logarithmic market scoring rule is to reduce future payoffs using a discount parameter. Based on this intuition. The general informativeness condition guarantees that no matter what the signal distribution is. Because the WPBE reveals all information. Intuitively if a player has some private information that may be revealed by its signal. according to Theorem 7. both Alice and Bob reports truthfully at stages t − 1 and t. and 9.4.4. there is no WPBE that reveals all information after finite number of trades. Consider the last two players. Truthful betting is the only finite equilibrium.3. Alice would want to deviate to bluff in stage t − 1 and play again in stage t + 1. Facilitated with results in Sections 9. Thus. 9 Discounted LMSR and Entropy Reduction For infinite-stage games. truthful betting equilibrium is the only WPBE of the Alice-Bob-Alice game. σ can not be a WPBE strategy profile. We present a natural metric to quantify degree of information aggregation in Section 9. we characterize the infinite-stage games as continuation games in Section 9. consider any WPBE strategy profile σ that always results in the optimal market prediction after some number t of stages in a potentially infinite-stage game. the market can not fully aggregate information of all players with finite trades. The game between stages i and j can be viewed as an Alice-Bob-Alice game. we present our convergence analysis for discounted LMSR in Section 9. . 9.Gaming Prediction Markets: Equilibrium Strategies with a Market Maker 29 stage j and i + 1 < j ≤ t. Contradiction arises.

1 and 8. In other words.2 For simplicity. Let δ ∈ (0. The δ-discounted market scoring rule is a market microstructure in which traders update the predicted probability of the event under consideration just as they would in the regular market scoring rule. for an Alice-Bob-Alice CI game in the discounted LMSR. we assume that there are only two players. However. 1) be a discount parameter. however. The trader would then be given a payoff of δ i−1 (log q − log p).1 can show that truthful betting is the unique WPBE for finite-stage and infinite-stage CI games with finite players in the discounted LMSR. for example. we only focus on I games in the discounted LMSR. we will show that. truthful betting is still the unique WPBE in the discounted LMSR. the market price will converge towards the optimal probability for the particular realized set of information signals. 9. 2 . with Alice being the first player. Suppose a trader moves the prediction for outcome Y from p to q in the i-th stage of the market. On the other hand. We note that for CI games. Alice and Bob. from now on. We will show that the discounted LMSR can have better non-myopic strategic properties for I games. We do not. Chen et al. it does surely happen in the long run at any WPBE. This decomposition is possible in equilibrium analysis. because the actual firststage strategy σ1 is common knowledge in an equilibrium profile. although complete aggregation of information may not happen in two rounds. we assume a strict alternating sequence of trades between Alice and Bob. They alternately play in the market. and the outcome Y is later observed to happen. the same arguments in Sections 7. the value (positive or negative) of trade is discounted over time.2 Continuation Games It is useful to think about an infinite-stage market game G as a first-stage move followed by a subsequent game H that perhaps depends on the first-stage move. Alice does not want to deviate from the truthful betting strategy as in the non-discounted LMSR because for any deviating strategy her expected profit in the first stage is the same in both the non-discounted LMSR and the discounted LMSR but her expected profit in the third stage in the discounted LMSR is strictly less than that in the non-discounted LMSR. the myopic strategic properties of the market scoring rule are retained in the discounted form. The regular market scoring rule corresponds to δ = 1. we are proposing a mechanism that explicitly discounts gains and losses in later rounds. if the outcome Y did not happen. Intuitively. in any WPBE strategy profile σ. As truthful betting is the unique WPBE in an Alice-Bob-Alice CI game in the discounted LMSR. For simplicity. Thus. the beliefs of the players after observing the first move x are consistent with private signals It is possible that traders intrinsically discount future profits. Then. and the trader moves the prediction from p to q in the i-th stage the trader would earn a payoff of δ i−1 (log(1 − q) − log(1 − p)).30 Y. assume any intrinsic discounting by traders. Clearly. rather. making deviation even more undesirable. they may be uncertain about how long their information will remain private. Even though that the discounted LMSR may admit equilibria in which players bluff with some probability.

.Gaming Prediction Markets: Equilibrium Strategies with a Market Maker 31 and a common prior p that depends on the initial prior p0 . the prior p in H is consistent with p0 . . Proof We need to show that if Hi → Hi+1 . We denote this G → H. suppose that x is one of the points that the first player in Hi plays with positive probability according to σ. ω = z} = Pi {c1 = k. . Observe that if G → H. We define G → G to always be true. ∗ we write G → H and say G reduces to H if there is a sequence G = ∗ H0 . Let Pi {c1 = k|x. the continuation game Hi+1 has joint probability distribution Pi+1 such that Pi+1 {c1 = k. ∗ We define the relation → as the reflexive transitive closure of →. Definition 3 We say that H is a continuation game of G if there is a first-stage strategy σ1 and a move x such that x is played with nonzero probability under σ. ω = z|c1 = k} Pi {c1 = k|x. c2 = l. b) are the same in G and H. Only the prior over different signals changes after a first-round strategy. i. Without loss of generality. player 2 updates its beliefs after seeing x. ω = z|x. and x. Hi+1 satisfies the same independence property as Hi . – If G satisfies the independence (I) property. signal realization of player 2. then. and the sequence of play in H is the same as the subsequence of play in G after the first move in G. Then. ω = z} Pi {c1 = k|σ} where k. cB = b) for any given pair of signal realizations (a. and realized outcome respectively. Then. so does H. l. Hk = H such that Hi → Hi+1 for all i. We use Pi {·} to represent probability of some outcome under Pi . σ} = Pi {c1 = k. σ} be the probability that player 1 has signal k conditional on that it plays x. Lemma 7 shows that continuation games preserve the independence property of the original game.e. .. the full-information optimal probabilities Pr(Y |cA = a. σ} ∗ ∗ = Pi {c1 = k|x. the first-stage strategy σ1 and the observed move x. not the full-information posterior probabilities. so does H. c2 = l. Let σ be a WPBE strategy for game Hi . σ1 . H1 . and z represent signal realization of player 1. σ}Pi {c2 = l. – If G satisfies the conditional independence (CI) property. Lemma 7 Suppose G → H. c2 = l. H2 . Let Pi be the joint probability distribution for Hi . knowing player 1 follows σ. .

Pi+1 {c2 = l|ω = z} = = = Pi {c1 =k|x. c2 = l|ω = z} = Pi+1 {c1 = k|ω = z}Pi+1 {c2 = l|ω = z}. ω = z} = l|ω = z} Pi+1 {ω = z} Pi {c1 =k|x. Pi+1 {c1 = k. c2 = l. Pi+1 {c1 = k. σ} = Pi {c1 = k|σ}Pi {c2 = l} Pi {c1 = k|σ} = Pi {c1 = k|x.σ} k Pi {c1 =k|σ} Pi {c1 = k. c2 = l. σ} = Pi {c1 = k.σ} k Pi {c1 =k|σ} Pi {c1 Thus. Pi+1 of game Hi+1 also satisfies I condition. c2 Pi+1 {ω = z}/Pi {ω = z} Pi {c1 =k|x.σ} Pi {c1 =k|σ} Pi {c1 = k|ω = z}Pi {c2 Pi {c1 =k|x. ω = z} Pi {c1 = k|σ} Pi {c1 = k|x.σ} Pi {c1 =k|σ} Pi {c1 = k. c2 = l. σ} Pi {c1 = k. Pi+1 {ω = z}/Pi {ω = z} Pi {c1 =k|x.σ} Pi {c1 =k|σ} Pi {c1 = k. σ}Pi {c2 = l} . Pi+1 {c1 = k. c2 = l. c2 = l|ω = z} = Pi+1 {ω = z} = = = Pi {c1 =k|x.σ} k Pi {c1 =k|σ} Pi {c1 = k.σ} Pi {c1 =k|σ} Pi {c1 = l|ω = z} Pi+1 {ω = z}/Pi {ω = z} = k. We first show that if Pi of game Hi satisfies CI condition. c2 = l} Pi {c1 = k|σ} Pi {c1 = k|x.σ} Pi {c1 =k|σ} Pi {c1 = k|ω Pi+1 {ω = z}/Pi {ω = z} = z} . Pi+1 of game Hi+1 also satisfies CI condition. c2 = l|ω = Pi {c1 =k|x. Pi+1 {c1 = k|ω = z} = = Pi+1 {ω = z} Pi {c1 =k|x. ω = z} = k.σ} k Pi {c1 =k|σ} Pi {c1 = k. Chen et al. c2 = l} = = ω ω Pi+1 {c1 = k. ω = z} Pi {c1 = k|x. ω = z} = z} . We then show that if Pi of game Hi satisfies I condition. we also have = Pi {c2 = l|ω = z}. c2 = l. ω = z} Pi+1 {c1 = k.32 Y. ω = z}/Pi {ω CI condition holds for Hi+1 . c2 = l|ω = z} z} = z} Pi+1 {ω Pi {c1 =k|x.

Pi {c1 = k|x.e.. We also have Pi+1 {c1 = k} = = l l Pi+1 {c1 = k. If signals are substitutes.3 Complementarity or Substitution of Signals In this section. 9.Gaming Prediction Markets: Equilibrium Strategies with a Market Maker 33 The second to the last equality is because of the independence condition of Hi . Pi+1 {c1 = k. the expected profit from knowing both signals is greater than the sum of the expected profits from knowing only the individual signals. i. σ} Pi {c1 = k. if signals of Alice and Bob are complements. For a game G participated by Alice and Bob. In other words. σ} Pi {c1 = k. c2 = l} Pi {c1 = k|σ} = l = Pi {c1 = k|x. Hi+1 satisfies independence. Definition 4 The complementarity coefficient C(G) of a market game G is defined as π A (σ N ) + π B (σ N ) C(G) = . π AB = π B (σ N ) + π A (σ T ) = H(ω|cA . and Alice moves first. π AB is the sum of Alice’s and Bob’s expected profit if they both follow the truthful strategy. cB ). ∗ By induction. c2 = l} Pi {c1 = k|σ} = k = Pi {c2 = l}. we introduce the concept of complementarity for signals. σ} and Pi+1 {c2 = l} = = k Pi {c1 = k|x. π AB . σ} Pi {c1 = k|σ}Pi {c2 = l} Pi {c1 = k|σ} k Pi+1 {c1 = k. then H preserves the same independence property as G. σ} Pi {c1 = k|σ}Pi {c2 = l} Pi {c1 = k|σ} Hence. the order of play is not important. if G → H. c2 = l} Pi {c1 = k|x. Intuitively. the total profit potential in G is π AB = π A (σ N ) + π B (σ T ). c2 = l} = Pi+1 {c1 = k}Pi+1 {c2 = l}. the relation is reversed. c2 = l} Pi {c1 = k|x. We note that under either the CI condition or the I condition. Now we define the complementarity coefficient of a game that captures the complementarity of players’ signals.

π A (σ N ) + π B (σ N ) > π A (σ N ) + π B (σ T ) = π AB ⇒ C(G) > 1. any game G that starts with the prior probability ˆ satisfies C(G) < 1. and the associated ex-ante . ˆ ˆ ˆ By definition of C. where the right-hand side profits are implicitly a function of the game G under consideration. Thus. Lemma 8 Under the CI model. σ). If C(G) > 1.34 Y.4 Profit Potential and Entropy We now present a natural metric. Lemma 9 Under the I model. ∗ 9. Thus. that quantifies the degree of aggregation in a prediction market after i trades under strategy profile σ: Di is the expectation. π A (σ N ) + π B (σ N ) < π A (σ N ) + π B (σ T ) = π AB ⇒ C(G) < 1. an information node φ consists of a realization of the signals of the two players and a sequence of t trades in the market. Formally. and thus. By lemma 7. ˆ The complementarity bound Cσ (G) is the minimum value of C(H) over all games H that G could reduce to by following the strategy profile σ: ˆ Cσ (G) = H:G→H by σ min ∗ C(H) We note that if C(G) < 1. Chen et al. we must have Cσ (G) > 1 for any σ. and denote p∗ (φ) = p∗ (φ). it follows that π B (σ N ) > π B (σ T ) for any CI game G with informative and distinct signals. for a strategy profile σ and a number of stages t. Let p∗ (φ) = Pr(Y |realization of signals in φ). and thus we must have Cσ (G) < 1. Proof From Theorem 2. Thus. and denote pi (φ) = pi (φ). The aggregative effect of the strategy profile σ after i trades is summarized by the collection of information nodes φ that can be reached. 1 − p∗ (φ) . Proof From Theorem 1. the information deficit Di . we know that any H such that G → H itself satisfies the CI condition ˆ no matter what strategy the reduction follows. it follows that π B (σ T ) > π B (σ N ) for any I game G starting with the prior probability. Cσ (G) < 1. of the relative entropy between the optimal probability (given the realization of the signals) and the actual probability after i stages. 1 − pi (φ) . any game G with informative and distinct signals ˆ satisfies Cσ (G) > 1 for any strategy profile σ. signals are substitutes. Let pi (φ) = Pr(Y |first i trades in φ. we have Cσ (G) ≤ C(G). over all possible signal realizations and the randomization of moves as dictated by σ. signals are complements. the market probability pi (φ) in stage i.

k∈Z+ ˜ M i = D1 − D2 + δ(D3 − D4 ) + . we can now define M i as the expected profit of the i-th trade in the discounted LMSR. in addition to measuring the distance from full information aggregation. i. Di is always nonnegative. ˜ Given the definition of M i . with positive probability. We now show that. because the relative entropy is always nonnegative.e. φ:φ∈σ When t = 0. The key result is that Di can be related to the expected payoff of the i-th stage move in the non-discounted (standard) LMSR: Lemma 10 Let M i denote the expected profit (over all signal nodes φi ) of the i-th trade under profile σ.. . Then. we define Di (σ) = Eσ [D(p∗ (φ)||pi (φ))] = Pr(φ)D(p∗ (φ)||pi (φ)) . the information nodes φ correspond to different realizations of the signals. If Di > 0. ˜ Mi = δ k (D2k − D2k+1 ) ∀i = 2k + 1 δ k (D2k+1 − D2(k+1) ) ∀i = 2k + 2 (13) ˜ Using the definition of M i we write the total profit for Alice in the δdiscounted LMSR as: Alice’s payoff = i:i=2k. . If Di = 0. Now for i ≤ t. . Di also enables interesting strategic analysis. We can similarly rewrite Bob’s expected payoff as: Bob’s payoff = i:i=2k. . the market has not yet reached the optimal probability. it implies that the market will always have reached its optimal probability for the realized signals by the i-th stages. We assume discounting after every even trade. it indicates that.k∈Z+ ˜ M i−1 = D0 − D1 + δ(D2 − D3 ) + .Gaming Prediction Markets: Equilibrium Strategies with a Market Maker 35 probability Pr(φ) of reaching each such information node. . Note that D0 is the same as π AB defined in the previous section. after every trade by Bob. M i = Di−1 − Di Proof 1 − pi (φ) pi (φ) + (1 − p∗ (φ)) log pi−1 (φ) 1 − pi−1 (φ) Mi = φ:φ∈σ Pr(φ) p∗ (φ) log = φ:φ∈σ Pr(φ) D(p∗ ||pi−1 ) − D(p∗ ||pi ) = Di−1 − Di The first equality holds by the definition of M i and the second by the definition of relative entropy. This suggests another interpretation for Di : Di represents the expected value of the potential profit left for trades after the i-th trade.

We exclude such degenerate cases. because if not. C(G). for I games that start with the prior probability. Recall that after stage 2. This argument generalizes to any even number of rounds. In the first stage. Bob will revise his beliefs consistent with the profile σ and the realized action x. Now. Without loss of generality we assume that Bob moves second in the market. and a particular instance of the two-person market game induced by the δ-discounted LMSR. a simple deviation to the truthful strategy would be beneficial. Alice will take some (perhaps mixed) action x dictated by σ. the stability of a given strategy profile σ may change with δ. Note that D0 = π AB . we bound the value of Dn for large n. the total expected payoff of Bob in equilibrium is also at least π B (σ N ). We will express our convergence bound in terms of complementarity coefficient. we could have lack of information aggregation even with the myopic strategy. fix a discount parameter δ. by looking at the total expected profits within the first 2k moves. Of course. Therefore we can bound D2 as: D2 ≤ D0 − [π A (σ N ) + π B (σ N ) − δD2 ]. and thus.36 Y. the myopic strategy may not involve any movement by either player. 9. We also know that the total expected payoff of Alice in equilibrium is at least π A (σ N ). Let σ be any WPBE of this market game. we can rewrite inequality (15) D2k ≤ ˆ D0 (1 − Cσ (G)) 1 − δk (16) . Because π B (σ) ≥ π B (σ N ). the total expected payoff of both players is at most δD2 . and assume that C(G) > 0. and the remaining profit potential δ k D2k : ⇐⇒ (1 − δ )D2k ≤ D0 − [π A (σ N ) + π B (σ N )] π A (σ N ) + π B (σ N ) = D0 1 − D0 0 D (1 − C(G)) ⇐⇒ D2k ≤ 1 − δk k D2k ≤ D0 − [π A (σ N ) + π B (σ N ) − δ k D2k ] (14) (15) as: ˆ Note that. According to Theorem 2. Chen et al. by definition. Note that if C(G) = 0. We reiterate that the definition of Di is not dependent on δ. Thus. This means that the total payoff of the first two rounds in the market is at least π A (σ N ) + π B (σ N ) − δD2 . C(G) ≥ Cσ (G). Under the equilibrium strategy profile σ. in any WPBE. It is a measure of the informational distance between the prices after i trades in profile σ and the optimal prices. we have π B (σ) ≥ π B (σ N ).5 Asymptotic Convergence in Discounted LMSR for I Games In this part.

for both finite-stage and infinite-stage games. Now consider the remainder of the game after 2k rounds.Gaming Prediction Markets: Equilibrium Strategies with a Market Maker 37 From inequality (16) we see that the bound on D2k depends only on δ. Then. Hence. for any game G with independent signal. then G → H2k . We show that traders with conditionally independent signals may be worse off by bluffing in LMSR. To summarize. as ˆ D and Cσ (G) are both constants for any instance of the market and equilibrium profile σ. . we examine different scenarios where traders at equilibrium bet truthfully or bluff. Specifically. are considered. inequality (17) shows that ˆ 4 log Cσ the relative entropy of the probabilities with respect to the optimal probabilities reduces exponentially over time. we rewrite inequality (16) in terms of δ and for a round n = 2km. ˆ ˆ We set k such that δ k/2 = Cσ (G). ˆσ (G)) ˆ Cσ (G) < 1. Using this value of k and a log δ n value m = 2k we note that: 0 Dn ≤ D0 = D0 = D0 (1 + δ ) n ˆ = D0 (1 + Cσ (G))− 2k − n log δ ˆ ˆ = D0 (1 + Cσ (G)) 4 log Cσ (G) = D0 δ ˆ n ˆ − log(1+Cσ (G)) ˆ 4 log Cσ (G) m ˆ 1−Cσ (G) 1−δ k ˆ 1−Cσ (G) (1−δ k/2 )(1+δ k/2 ) k/2 −m m (17) σ (G)) Note that − log(1+C(G) depends only on the complementarity bound of ˆ 4 log Cσ the game. for all n > 1. etc. the equilibrium profile σ will also induce an equilibrium profile on H2k . According to Lemma 7. We can repeat the argument to bound D4k in terms of D2k .e. For this game. − log(1+C(G) > 0. the information deficit Dn after n rounds of trading is bounded by Dn ≤ D0 δ n ˆ − log(1+Cσ (G)) ˆ 4 log Cσ (G) 10 Conclusion We have investigated the strategic behavior of traders in LMSR prediction markets using dynamic games with incomplete information. H2k preserves the independence property of G. denoted as ∗ H2k . Hence. with a δ-discounted logarithmic market scoring rule. truthful betting is not only a PBE strategy but also a unique WPBE strategy for all traders in LMSR. let σ be any WPBE strategy profile. we have proven the following result: Theorem 13 Let G be a market game in the I model. In this way. Two different information structures. k = 2 log Cσ (G) . i. According to Lemma 9. Moreover. conditional independence and unconditional independence of signals. Both finite-stage and infinite-stage games are investigated. Therefore.

C. A. to ameliorate this potential problem. Nelson. Gale. B. 2008. They might engage in bluffing strategies to bargain over how this subsidy is divided. Healy and Mohammad Mahdian for helpful discussions. A. 2009. We use a simple modification to the market scoring rule. as measured by the relative entropy between full information aggregation and the actual market probability. E. When are signals complements or o a substitutes? Working paper. which includes a form of discounting. Allen and D. Can asset markets be manipulated? A field experiment with race-track betting. Prediction markets as decision support systems. We thank anonymous reviewers for their valuable comments. J. References 1. Kulick. Roust. Handbook of Experimental Economics Results. For infinite-stage games with unconditionally independent signals. North Holland. Berg and T. We are grateful to Paul J. Information Systems Frontier. This allows us to prove a bound on the rate at which the error of the market. Chen et al. Plott. Stock-price manipulation. A. Berg. A. players can pool their information together to make sharper predictions than either could alone and thus extract an even larger profit. Results from a dozen years of election futures markets research. A. 6. Journal of Political Economy. F. Explicit discounting can make this bargaining more efficient. 1992. S. 1998. Journal of Economic Behavior and Organization. B¨ rgers. 2003. T. In C. editors. The exponent depends on the complementarity coefficient of the market instance. when the signals of traders are unconditionally independent. there does not exist any WPBE that fully reveals information of all players in finite stages. Amsterdam. Rietz. Smith. (106):457–482. Acknowledgments Stanko Dimitrov was supported by NSF grants CCF-0728768 and IGERT0114368. Forsythe. Rahul Sami was partially supported by NSF grant CCF-0728768. On the other hand. B. Camerer. Axelrod. The Review of Financial Studies. Kr¨ hmer. Design improved parimutueltype information aggregation mechanisms: Inaccuracies and the long-shot bias as disequilibrium phenomena. and T. 3. 5. E. truthful betting is not a WPBE for finite-stage games. J. reduces exponentially over time. Traders have incentives to strategically mislead other traders with the intent of correcting the errors made by others in a later stage. C. J. . R. 69:170–181. (5):503–529. Hernando-Veciana. 5:79–93. D.38 Y. 4. F. 2. The need for discounting shows a connection to classical bargaining settings in which players bargain over how to divide a surplus they can jointly create. Rietz. and K. Moreover. In a prediction market. and D. We provide a bluffing equilibrium for a restricted twoplayer game. 2007. We thank John Ledyard for pointing us to earlier work on discounting in parimutuel markets. F. informed players can extract a profit from the market. Plott and V. R.

Hanson. R. 26. R. expectations. J. Reeves. 10. Elements of Information Theory. American Economic Review. Logarithmic market scoring rules for modular combinatorial information aggregation. Cover and J. A. T. Zuber. Division of the Humanities and Social Sciences. 9. Hanson. 17. ACM. R. 21. Futures manipulation with cash settlement. E. J.preconditions and evidence. R. Oprea. C. R. Ross. Green. Brown. A. 15. New York. Y. Journal of Finance. The Third Workshop on Internet and Network Economics (WINE 2007). 13. New York. R. D. 2003. 82(5):1142–1161. 60(4):449– 459. Dare. Working paper No. 1(1):1–15. G. S. BC Canada. and R. Theoretical Computer Science. Informed traders and price variations in the betting market for professional basketball games. Pennock. 12. R. R. A. Milgrom and N. 25. Yilmaz. pages 49–56. 343:114–132. Combinatorial information market design. Plott. A. T. Vancouver. and R. F. Porter. Information aggregation mechanisms: Concept. LIII(1):385–401. A. Journal of Economic Behavior and Organization. J. Forsythe. F. Hanson. Anatomy of an experimental political stock market. J. 1131. Chen and D. 26(1):17–27. M. Manipulation in political stock markets . and D. 1999. 2007. 8. L. M. Fortnow. July 2008. Debnath. A. M. and J. In Lecture Notes in Computer Science. R. Pennock. P. M. (A preliminary version appeared in the 2003 ACM Conference on Electronic Commerce). C. Sami. 2001. Information aggregation and manipulation in an experimental market. Fama. 20. Pennock. 18. Kyle. Econometrica. trade and common knowledge. 2007. 24. 11. Dimitrov and R. Microeconomics Theory. 2004. Sami. NY. Seppi. and S. 1970. design and implementation for a sales forecasting problem. Neumann. 1992. 1985. Chakraborty and B. C. K. 14. 25:383–417. In Proceedings of the 23rd Conference on Uncertainty in Artificial Intelligence (UAI 2007). 7(2):187–206. Lawrence. and J. Bluffing and strategic reticence in prediction markets. Inc. 19. W. D. M. Mas-Colell. A utility framework for bounded-loss market makers. M. 5(1):107–119. CA. D. Efficient capital market: A review of theory and empirical work. W. Kumar and D. Journal of Financial Markets. 16. Rietz. Chen and C. Journal of Economic Behavior and Organization. 22. 2003. R. S. Journal of Finance. pages 70–81. Strobel.-Y. Hansen. 2005. Nelson. Continuous auctions and insider trading. Gonen. and T. Wright. Gandar. H. San Diego. 1992. Information Systems Frontiers. (47):1485–1502. Stokey. S. 1995. Technical Report. In Proceedings of the Fourth Annual ACM Conference on Electronic Commerce (EC’03). D. and R. 23.Gaming Prediction Markets: Equilibrium Strategies with a Market Maker 39 7. D. Non-myopic strategies in prediction markets. R. California Institute of Technology. Thomas. Fortnow. and actions: A survey on price formation in election stock markets. John Wiley & Sons. 39:83–110.. D. and M. Computation in a distributed information market. 2007. Information incorporation in online in-game sports betting markets. D. Forsythe. Whinston. 53(6):1315–1336. Wishes. 2007. Oxford University Press. 1999. L. Feigenbaum. 2002. Information. L. M. Manipulation in market order models. Hanson. L. P. M. 1982. Journal of Prediction Markets. . Y. Giles. Chen. D. 1991. Schmidt. In Proceedings of the Nineth ACM Conference on Electronic Commerce (EC’08). Journal of Economic Theory. D. Journal of Finance. Pennock.

Nikolova and R. Journal of Economic Perspectives. 18(2):127–142. 28. K. R. A. 30. Sami. Anomalies: Parimutuel betting markets: Racetracks and lotteries. In Proceedings of the Seventh Annual ACM Conference on Electronic Commerce (EC ’07). S. Working Paper. 1988. P. Thaler and W. 27. 50(3):336–351. W. 2(2):161–174. A strategic model for information markets. Strumpf. Historical presidential betting markets. W. Rhode and K. Chen et al. Rietz. T. 2004. Ziemba. 29. Journal of Economic Perspectives. Manipulating political stock markets: A field experiment and a century of observational data. but markets are made: Individual rationality. Oliven and T. H.40 Y. Strumpf. 31. Management Science. 2007. 2004. 2007. arbitrage and market efficiency on an electronic futures market. E. P. . Suckers are born. S. Rhode and K.