Abstract
This paper explores the intersection of prediction markets, mechanism design, and cooperative game theory. By transitioning from traditional security-trading models to a state-based information revelation framework, the research addresses the challenge of strategic manipulation by non-myopic agents. The proposed mechanisms utilize proper scoring rules and the Shapley value to ensure incentive compatibility and fair reward distribution. A key contribution is a hybrid algorithm that separates raw information reporting from probabilistic modeling, providing a robust architecture for information aggregation.
Introduction
Traditional prediction markets typically operate on the assumption of myopic behavior, where participants trade based on their current beliefs relative to market prices. However, strategic agents may deviate from this behavior, intentionally misleading the market to manipulate prices for future gain. This paper, authored by Vincent Conitzer, bridges the gap between prediction markets and mechanism design to create systems that are robust against such manipulation. By framing information aggregation as a cooperative game, the author introduces a state-based model where agents report private signals directly, allowing for the application of formal incentive-compatibility theories.
Problem Statement
The primary challenge in existing prediction markets is the lack of "truthfulness" among sophisticated participants. In a market for binary outcomes, an agent with significant information might trade against their signal to depress prices before buying a larger volume at a discount. While mechanism design provides tools like "incentive compatibility" to encourage truth-telling, mapping these concepts to iterative, multi-agent prediction environments is non-trivial. Conitzer seeks to unify these fields by moving beyond "trading securities" toward "direct information revelation."
Methodology
The proposed framework utilizes a state-based model for prediction. Instead of reporting a single probability, agents report their "private signal"—the set of possible world-states they have not yet ruled out. The paper evaluates several reward mechanisms, including individual rewarding, marginal information rewarding (similar to Market Scoring Rules), and a Shapley Value mechanism that averages contributions across reporting orders.
Hybrid Algorithm for Practical Implementation
To address the practical difficulties of natural language information processing, the paper proposes a hybrid design. This design distinguishes between Information Agents (who provide raw data) and Probability Agents (who translate that data into mathematical probabilities). This separation allows the market to leverage human intuition alongside formal modeling.
Key Insights
A central finding is that information aggregation is fundamentally a cooperative game. The value generated by a group is the accuracy of their collective prediction. By applying the Shapley value, rewards can be distributed in a way that is both fair and incentivizes truth-telling. Furthermore, the "Pivotal" mechanism is shown to be ex-post incentive compatible, meaning an agent's best strategy remains truth-telling even if they possess knowledge of other agents' signals.
Results
The theoretical analysis yields several key results:
- Theorem 1: Demonstrates the flexibility of "pre-characteristic functions," showing that information games do not possess a restrictive hidden structure.
- Theorem 2: Provides an axiomatization of the pivotal mechanism, proving it is the unique mechanism that satisfies specific fairness and "dummy agent" criteria.
- The Hybrid Algorithm: The paper formalizes a concrete algorithm (Figure 1) that uses probability agents as subroutines to price natural-language information provided by information agents, effectively automating the marginal information reward structure.
Discussion and Limitations
While the framework provides a strong theoretical foundation, several challenges remain. The model assumes a "common prior"—a starting probability agreed upon by all—which is often the very unknown the market is intended to discover. Additionally, the computational complexity of intersecting complex natural language signals remains a hurdle. Future research may explore how combinatorial auction concepts can be adapted to handle more complex, multi-state events.
Summary updated