Abstract
Prediction markets are designed to elicit information from multiple agents in order to predict (obtain probabilities for) future events. A good prediction market incentivizes agents to reveal their information truthfully; such incentive compatibility considerations are commonly studied in mechanism design. While this relation between prediction markets and mechanism design is well understood at a high level, the models used in prediction markets tend to be somewhat different from those used in mechanism design. This paper considers a model for prediction markets that fits more straightforwardly into the mechanism design framework. We consider a number of mechanisms within this model, all based on proper scoring rules. We discuss basic properties of these mechanisms, such as incentive compatibility. We also draw connections between some of these mechanisms and cooperative game theory. Finally, we speculate how one might build a practical prediction market based on some of these ideas.
Introduction
Prediction markets are powerful tools for information aggregation, yet their theoretical foundations often sit at an awkward intersection between finance and game theory. Traditional models frequently assume "myopic" agents—those who trade immediately based on their current beliefs. However, real-world participants are strategic; they may withhold information or mislead the market to secure better prices later.
In this work, originally presented at UAI 2009, Vincent Conitzer proposes a framework that bridges this gap. By moving away from the "trading" metaphor and toward a "direct revelation" mechanism design model, the paper explores how to incentivize agents to reveal their private information truthfully and how to distribute rewards fairly using cooperative game theory.
The Problem: Strategic Manipulation and Redundancy
Standard prediction markets face two primary hurdles. First is the strategic manipulation problem: an agent with significant information might trade in the wrong direction to manipulate the price, only to "correct" it later for a larger profit. Second is the information redundancy problem: if multiple agents possess the same piece of information, a market should not pay for that information multiple times.
To address these, Conitzer introduces a state-based model. In this model, there is a set of possible world states $\Omega$. Each agent $i$ receives a private signal $S_i \subseteq \Omega$, representing the states they believe are still possible. The goal of the mechanism is to aggregate these signals to determine the true state.
Mechanism Design and Proper Scoring Rules
The paper evaluates several mechanisms rooted in Proper Scoring Rules (e.g., logarithmic or quadratic scores), which are mathematical functions designed to reward accurate probability estimates.
1. Individual Rewarding
In this baseline, agents are rewarded based on their own reports. While simple, this fails to account for what others have already contributed, leading to excessive payouts for redundant information.
2. Marginal Information Mechanism
Here, agents are rewarded only for the new information they provide relative to previous reporters. While this solves the redundancy problem, it introduces a "path dependency" where the order of reporting drastically changes an agent's payout.
3. Shapley Value Information Mechanism
To resolve the ordering issue, Conitzer applies the Shapley Value from cooperative game theory. This mechanism calculates an agent's reward by averaging their marginal contribution across all possible permutations of agent arrivals. This ensures that the total reward is distributed fairly based on the unique value each agent brings to the collective prediction.
4. Pivotal Information Mechanism
Drawing inspiration from the Vickrey-Clarke-Groves (VCG) mechanism, this approach ensures ex-post incentive compatibility. This is a high bar: it means that even if an agent knew exactly what every other agent knew, their best strategy would still be to tell the truth.
Key Insights: Information vs. Probability
One of the paper's most practical insights is the distinction between two types of participants:
- Information Agents: These agents possess raw data or "signals" (e.g., "I saw a supply chain disruption").
- Probability Agents: These agents possess the expertise to translate raw data into numerical probabilities (e.g., "A supply chain disruption of this type reduces quarterly earnings by 5%").
Conitzer suggests a hybrid market where Information Agents provide the "what" and Probability Agents provide the "so what." This division of labor could simplify the implementation of complex prediction markets by allowing specialists to focus on their strengths.
Theoretical Results and Limitations
The paper provides several key proofs:
- Incentive Compatibility: Under a "consistent" model, both the marginal and Shapley-based mechanisms are ex-interim incentive compatible (truth-telling is the best strategy on average).
- Robustness: Theorem 1 demonstrates that these "information games" can represent almost any characteristic function game, proving the framework's versatility.
However, challenges remain. Calculating the Shapley value is computationally expensive ($n!$ permutations), making it difficult to scale to large markets without approximations. Furthermore, the model relies on a common prior, assuming all agents agree on the initial probabilities before receiving their private signals—an assumption that rarely holds in the messy reality of human forecasting.
Conclusion
By integrating cooperative game theory into prediction market design, Conitzer provides a rigorous path toward truthful information aggregation. While computational and practical hurdles exist, the "Information vs. Probability" split and the use of the Shapley value offer a sophisticated blueprint for the next generation of decentralized forecasting tools.
Summary updated