TL;DR
This paper introduces a game-theoretic model to analyze prediction markets where participants can directly manipulate the final outcome they are betting on. The authors discover a sharp threshold: if a key participant's market attendance is highly uncertain, the market remains informative and resistant to manipulation; however, if participation is highly likely, traders successfully collude to manipulate both the market price and the final outcome for profit.
What Problem Does This Paper Address?
Standard prediction market theory assumes that traders are passive observers who possess unique information but have no direct control over the actual event (e.g., predicting the weather). In reality, prediction markets are frequently used in environments where traders have significant skin in the game and direct influence—such as corporate employees betting on product release dates, or politicians trading on election outcomes.
When outcomes are endogenous (controlled by the participants), it distorts incentives. Traders might alter their real-world actions purely to win their market bets. This paper addresses a critical gap: understanding when these markets are prone to manipulation, how much we can trust their prices, and how the uncertainty of participant turnout affects strategic behavior.
Methods
The authors construct a two-stage, two-player game featuring two strategic, Bayesian agents: Alice and Bob.
- Stage 1 (Trading): Alice trades first against a Market Scoring Rule (MSR) market maker, shifting the market price to $p_A$. Bob gets the turn to trade next, but he only shows up with a probability of $1-\pi$ (making $\pi$ his non-participation probability). If he trades, he shifts the price to $p_B$.
- Stage 2 (Voting/Action): Both agents simultaneously cast a "vote" ($v_A, v_B \in \{0,1\}$) that determines the final liquidation value of the market security, defined as the average vote: $v = (v_A + v_B)/2$. If Bob did not participate in the trading stage, he is assumed to vote truthfully according to his private signal.
The authors analyze the Perfect Bayesian Equilibria (PBE) of this game across three popular market-making scoring rules: Logarithmic (LMSR), Quadratic (QMSR), and Spherical (SMSR).
Datasets
Not discussed in the paper. This is a purely theoretical, game-theoretic paper and does not utilize empirical datasets or experimental benchmarks.
Key Insights and Ideas
- The Participation Threshold ($\pi_c$): The market's integrity hinges entirely on how likely Bob is to skip the trading stage. This defines a clear boundary between two regimes: Low Participation Probability (LPP) and High Participation Probability (HPP) equilibria.
- Uncertainty Enforces Truthfulness (LPP Regime): Counterintuitively, when there is a high chance that Bob won't trade ($\pi > \pi_c$), Alice cannot safely plan a collusion. To maximize her expected payoff against a potentially truthful Bob, she is forced to trade in a way that reflects her actual beliefs. As a result, her true private signal can be mathematically recovered from her price update, making the market price an effective information disseminator.
- Certainty Breeds Collusion (HPP Regime): If Bob is highly likely to trade ($\pi < \pi_c$), the market collapses into a manipulative state. Alice deliberately sets a predetermined price ($p^L$ or $p^H$) that signals Bob to join her. They coordinate their subsequent votes to artificially maximize and split the market profits, rendering the market price completely uninformative of their underlying private signals.
- Scoring Rule Vulnerabilities: The threshold for collusion varies by the market mechanism used. As shown in the paper's equilibrium analysis, LMSR displays a unique "dipped" threshold curve compared to the smooth parabolic shapes of QMSR and SMSR due to the mathematical properties of the logarithmic function.
Key Results
- The authors prove the existence of unique threshold bounds ($p^L$ and $p^H$) that dictate Bob's rational voting response. For example, under a Quadratic MSR, these bounds are exactly $p^L = 0.25$ and $p^H = 0.75$ (see Table I).
- In the LPP domain, the market price perfectly tracks the Bayesian estimate of the actual outcome.
- The cross-over probability $\pi_c$ required to break collusion depends heavily on Alice's posterior belief about Bob's signal ($q_0$). For QMSR and SMSR, this threshold peaks when $q_0 = 0.5$, meaning collusion is hardest to sustain when Alice faces maximum uncertainty about Bob's information.
Limitations and Future Work
The model is intentionally restricted to a simplified two-player, two-stage setting where both agents hold equal power over the final outcome ($\alpha = 0.5$). Real-world environments are far more complex, usually consisting of many traders with asymmetric levels of influence. The authors note that expanding this framework to multi-player models and exploring varying degrees of individual control over the final outcome are necessary next steps to fully understand "worst-case" manipulative behaviors in large-scale prediction markets.
Summary updated