Probability Matching, a classical violation of expected utility maximization, refers to people's tendency to randomize, or even match their choice frequency to the outcome probability, when choosing over binary lotteries that differ only in their probabilities. Why? I present an experiment designed to distinguish between several broad classes of explanations: (1) models of Correlation-Invariant Stochastic Choice --- randomizing due to factors orthogonal to the correlation between lotteries, such as non-standard preferences or errors, and (2) models of Correlation-Sensitive Stochastic Choice --- deliberately randomizing due to misperceived hedging opportunities, especially when lotteries are negatively correlated. My experimental design differentiates between their testable predictions by varying the correlation between lottery outcomes. The findings indicate that the first class, despite being home to most existing theories, has limited explanatory power. Using additional treatment, I rule out Similarity Heuristics as a competing explanation with the second class. The results indicate that a vast majority of individuals deliberately randomize due to misperceived hedging opportunities.

[Link]

When updating beliefs, people often receive reports of the same underlying signals that differ both in their informational content and in their qualitative features. We isolate two distinct dimensions of statistical reports -- Instrumental Value (IV) and Information Richness (IR) – and ask how each shapes information demand and information use. In a pre-registered online experiment (N=200), participants see five canonical reports on a series of common binary signals that vary along both dimensions: (1) Majority, which signal is in the majority; (2) Percentage, the fraction of signals out of 100; (3) Difference, how many more dominant signals there are; (4) Count, the number of each signal; and (5) Sequence, the raw sequence of signals. Using a strategy method, we elicit participants’ willingness-to-pay (WTP) for each report and measure their Bayesian belief-updating performance on the same tasks. WTP rises with IR conditional on IV, sharply violating the IV-only benchmark. Performance shows the opposite pattern: it improves with IV but deteriorates with IR conditional on IV, opening a wedge between perceived and actual usefulness. Within participants, higher WTP for a report predicts larger deviations from Bayes on that same report, indicating substantial imperfect meta-cognition. To isolate the mechanism, we add sample-size information to Percentage and Difference and find that, relative to Count, the worse performance under Percentage comes from participants’ overreaction to 0% and 100% without inferring sample size. Deviations from Bayesian updating thus reflect not only how people process information but which features they choose to attend to in the first place.

Diversification is fundamental to optimal investing, enabling investors to reduce portfolio risk without sacrificing expected returns. Standard theory assumes that investors understand the joint distribution of asset returns and adjust how much to diversify in response to correlation, optimally exploiting the resulting hedging opportunities. However, empirical evidence on whether people do is mixed. I show experimentally that complexity drives both correlation insensitivity and correlation mis-response. I independently vary Cancellation Complexity (identifying decision-irrelevant states), and Tradeoff Complexity (evaluating cross-state tradeoffs) via a series of experiments. I find that both reduce correlation sensitivity, but Tradeoff Complexity additionally pushes negative-correlation allocations off the mean-variance frontier. A structural model identifies distinct channels: Cancellation Complexity operates through choice noise on a flatter evaluation landscape, while Tradeoff Complexity induces an allocation-level complexity penalty. Using U.S. stock data, I find that Tradeoff Complexity amplifies the behavioral, but not the rational, component of the comovement premium, particularly in small stocks.

We study how the format of belief elicitation shapes elicited beliefs. In two experiments on Prolific (N=251), we compare the slider-based Classical Method (CM), which requires an absolute magnitude judgment, to a Dynamic Binary Method (DBM) that elicits the same belief through a sequence of comparative judgments over nested subintervals. A cognitive-friction model nests both: CM updates a noisy signal against a prior mean and compresses reports toward the prior mean; DBM applies a fixed threshold rule to per-step raw signals that never references the prior. Aggregate medians are statistically tied, but DBM produces 25–47% higher variance. The mechanism is directly observable within the DBM trajectory: conditional on a correct first binary choice, the report-on-truth slope is 0.98 and indistinguishable from one; conditional on a wrong first choice, it is 0.79—precisely the compression CM exhibits unconditionally. DBM exit intervals function as calibrated imprecision measures, and DBM-recovered perceptual noise predicts the same participant's CM accuracy on held-out tasks. Elicitation format is an active input into belief formation, not a neutral measurement device.