Efficient “Risk Attitudes”

Speaker: A/Prof Rafael Polania

Affiliation: University of Melbourne

Location: Room 112, Chamberlain Building (#35), St Lucia Campus

Zoom: https://uqz.zoom.us/j/82603079317

Abstract: Risk attitudes and deviations from expected utility maximization, such as Allais violations, are usually interpreted as properties of preferences. We show that the same patterns can instead arise from interdependent constraints on attribute representations in an optimal but resource-limited agent that simply maximizes expected monetary reward. Specifically, we posit that reward and probability attributes are represented with finite precision and decoded by Bayesian inference using limited resources drawn from a common representation invariant capacity. Importantly, the reward environment determines the optimal allocation of probability precision and how capacity is best divided across the two attributes. Intuitively, this is because errors in probability representations matter more when attached to more consequential rewards. This single efficient representation mechanism generates stake and probability dependent risk attitudes and a rational agent using it may violate both common ratio and common consequence invariance through distinct properties of the probability representation. This mechanism further predicts how these violations change with probability scaling and payoff structure. Across several experiments in humans, the observed Allais paradox patterns follow these predictions, and in two large risky choice datasets the efficient representation model predicts held-out choices better than prospect theory, salience, random utility, and noisy cognition benchmarks. Apparent risk preferences can therefore arise from rational adaptation of a finite representation.