Robust Estimation of Risk Preferences

Obtaining quantitative measures of risk preferences is important because it enables researchers to formulate predictions. However, the stringent assumptions on which these predictions rely often raise concerns about external validity: can these estimated models produce reliable predictions beyond the specific settings from which they were derived? In this paper, we study preferences under risk that may deviate from Expected Utility (EU) by relaxing its fundamental and most disputed assumption: the independence axiom. Rather than committing to a specific non-EU model, we estimate the set of utility functions that characterize the largest subrelation of a preference that satisfies the independence axiom. Without making further assumptions on preferences, this approach allows us to predict choices between two lotteries when one lottery is better for all utilities.

We will conduct an experiment on choices under risk, testing whether predictions generated with our approach, despite being less granular, are more accurate out-of-sample than those obtained with specific economic models or machine learning algorithms.