Bounded Rationality in Wagering Mechanisms

David M. Pennock, Vasilis Syrgkanis, Jennifer Wortman VaughanUAI 2016 · OpenReview

Introduction

In traditional game theory, the "no-trade theorem" presents a significant paradox for the design of wagering mechanisms. It posits that if all agents are perfectly rational and share a common prior, no trade or bet should ever occur. The logic is simple: if a counterparty is willing to accept your bet, they must possess information that makes the bet unfavorable for you. Conversely, many existing wagering models rely on a "naivety assumption," where agents ignore the strategic reasoning of their opponents entirely.

This paper bridges these two extremes by applying a level-$k$ reasoning framework to wagering mechanisms. By modeling agents as boundedly rational—strategic but not infinitely so—the authors demonstrate how participation can be sustained through a specific form of overconfidence, while also identifying the inherent limits of these mechanisms as agents become more sophisticated.

Modeling Bounded Rationality

The researchers utilize the level-$k$ framework from behavioral game theory to analyze agent behavior:

  • Level-0 Agents: These participants have immutable beliefs and always participate truthfully, oblivious to the strategic nature of the game.
  • Level-$k$ Agents: These participants assume all other players are at level-$(k-1)$ and choose their best response based on that assumption.

The study focuses on the Kilgour-Gerchak mechanism, a budget-balanced system where players are compensated based on how their forecast accuracy (measured by a scoring rule like the Brier score) compares to the average accuracy of the group. The goal is to design mechanisms that achieve unbiased participation—where the decision to bet is independent of the specific private information an expert holds—and truthful reporting.

Key Insights: Overconfidence and the "Rationality Tax"

The paper reveals that participation in these mechanisms is driven by a "small dose of overconfidence." Because every level-$k$ player believes they are exactly one level more sophisticated than their opponents, they perceive an edge that overcomes the no-trade theorem.

However, this participation is subject to what might be termed a Rationality Tax. The authors prove that as the level of rationality ($k$) increases, participation rates inevitably decline toward zero. Sophisticated agents eventually realize that only the most informed experts remain in the pool, leading them to exit the mechanism unless their own information is of exceptionally high quality.

Unbiasedness and Truthfulness

A central finding of the research is the structural link between participation and honesty. The authors demonstrate that if a mechanism is designed to be unbiased (meaning the decision to participate does not leak information about the agent's private signal), then those who choose to participate are naturally incentivized to report their beliefs truthfully.

In symmetric information environments—where errors are independent and symmetric—the authors prove that a wagering mechanism achieves this unbiasedness if and only if the underlying scoring rule is symmetric (treating all outcomes, such as "heads" or "tails," identically).

Results and Mechanism Design

The paper provides several formal characterizations for optimal mechanism design:

  1. Participation Conditions: Theorem 3 establishes a recursive condition showing that a player participates if their expected score exceeds the average expected score of the level-$(k-1)$ players they believe they are facing.
  2. Maximizing Participation: To maintain the highest possible participation rates, the entropy function of the scoring rule should be as close to linear as possible (Theorem 6).
  3. The Weakly Proper Limit: Interestingly, mechanisms that maximize participation are often those that provide the smallest marginal incentive for truthfulness. By rewarding players more evenly, the mechanism avoids scaring off less-informed experts, though this involves a trade-off with the strength of the incentive for precision.

Conclusion and Future Directions

While this work provides a robust mathematical foundation for boundedly rational wagering, it operates under specific assumptions, such as a fixed $k$ level for all agents and symmetric information structures. Real-world forecasting often involves a "cognitive hierarchy" (a distribution of different $k$ levels) and asymmetric or correlated information.

Future research is required to validate these theoretical findings through empirical experiments, determining how well the level-$k$ model predicts actual human behavior in complex wagering environments like the Kilgour-Gerchak mechanism.

Summary updated

BibTeX
@inproceedings{pennock2016bounded,
  author = {David M. Pennock and Vasilis Syrgkanis and Jennifer Wortman Vaughan},
  title = {{Bounded Rationality in Wagering Mechanisms}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {2016},
  url = {https://openreview.net/forum?id=QRdIJHWZtmU},
}
Figure