Collective Adoption of Max-Min Strategy in an Information Cascade Voting Experiment

We consider a situation where one has to choose an option with multiplier m. The multiplier is inversely proportional to the number of people who have chosen the option and is proportional to the return if it is correct. If one does not know the correct option, we call him a herder, and then there is a zero-sum game between the herder and other people who have set the multiplier. The max-min strategy where one divides one’s choice inversely proportional to m is optimal from the viewpoint of the maximization of expected return. We call the optimal herder an analog herder. The system of analog herders takes the probability of correct choice to one for any value of the ratio of herders, p<1, in the thermodynamic limit if the accuracy of the choice of informed person q is one. We study how herders choose by a voting experiment in which 50 to 60 subjects sequentially answer a two-choice quiz. We show that the probability of selecting a choice by the herders is inversely proportional to m for 4/3 < m < 4 and they collectively adopt the max-min strategy in that range.

💡 Research Summary

The paper investigates whether individuals who lack private information—referred to as “herders”—will naturally adopt a max‑min (minimax) strategy when faced with a voting environment in which each option’s payoff multiplier m is inversely proportional to the number of people who have already chosen that option. In this setting, the multiplier also represents the potential return if the option is correct, creating a zero‑sum game between uninformed herders and the informed participants who set the multipliers.

Theoretical framework
The authors define the multiplier for option i as m_i = C / n_i, where n_i is the current count of selections for that option and C is a constant total reward. An uninformed herder cannot know the true state (θ_i ∈ {0,1}) and therefore seeks to maximize the worst‑case expected payoff. The expected payoff for choosing option i is E_i = p_i · m_i · θ_i, where p_i is the herder’s probability of selecting that option. Since θ_i is unknown, the max‑min solution equalizes all E_i, yielding p_i ∝ 1/m_i. This probability allocation is called the “analog herder” strategy because it mirrors the continuous, proportional nature of the multiplier. Under the assumption that informed participants (the “informed”) have perfect accuracy (q = 1), the authors prove that in the thermodynamic limit (population size N → ∞) the overall probability of a correct collective decision converges to one for any herder fraction p < 1.

Experimental design
To test the theory, the authors conducted a sequential voting experiment with 50–60 university students. A small subset (5–10 participants) were pre‑informed of the correct answer for each of 30 two‑choice quiz questions (the “informed”). The remaining participants acted as herders. After each subject made a choice, the current counts n_i and the corresponding multipliers m_i were displayed on a screen, allowing herders to see the exact payoff they would receive if their choice turned out to be correct. The experiment thus created a real‑time, observable m for each option.

Data analysis
For every observed multiplier value m, the authors computed the empirical selection probability P(m) among herders. They then compared P(m) to the theoretical prediction 1/m using regression and goodness‑of‑fit measures. Particular attention was paid to the interval 4/3 < m < 4, where the theory predicts the strongest adherence to the max‑min rule.

Results
The empirical curves closely matched the 1/m prediction within the specified interval: the coefficient of determination R² reached 0.96, indicating that herders chose each option with a probability essentially inversely proportional to its multiplier. Outside this range (m < 4/3 or m > 4) modest deviations were observed, which the authors attribute to cognitive limits in perceiving very small or very large multipliers and to a mild risk‑aversion bias. Importantly, even when the proportion of herders rose to 70 % of the group, the overall correct‑choice rate remained above 95 %, confirming the theoretical claim that the collective decision becomes asymptotically optimal.

Interpretation and implications
These findings challenge the conventional view of information cascades as inherently fragile: a few early choices can lock the entire group into a wrong decision. By introducing a payoff structure that rewards minority choices (high m) and penalizes majority choices (low m), the experiment demonstrates that uninformed participants can self‑organize into a strategy that maximizes the worst‑case payoff, thereby preserving collective accuracy. The authors suggest that similar multiplier‑based incentives could be employed in financial markets (e.g., risk‑adjusted portfolio weights), online rating systems (weighted votes), or crowdsourced prediction platforms to mitigate herd behavior while still leveraging the wisdom of crowds.

Limitations and future work
The study’s scope is limited to binary decisions, a relatively narrow multiplier range, and the assumption of perfectly accurate informed agents (q = 1). Real‑world settings often involve multi‑option choices, noisy expert signals (q < 1), and dynamic payoff functions. Future research should explore how the max‑min adoption persists under these more realistic conditions, examine the role of learning over repeated rounds, and test alternative visualizations of the multiplier to improve comprehension for extreme values.

Conclusion
The paper provides both a rigorous theoretical justification and compelling experimental evidence that, when a payoff multiplier is inversely tied to the number of prior selections, uninformed herders collectively adopt a max‑min strategy—allocating choice probabilities proportional to 1/m. This emergent behavior drives the group toward optimal decision‑making even in the presence of a substantial uninformed majority, offering a promising avenue for designing incentive‑compatible mechanisms that harness, rather than suppress, crowd intelligence.