Item

Abstract

The risk and ambiguity inherent to many decisions tend to resolve over time as more information becomes available. How a decision maker values their options at different times, facing different information, depends on their preferences over, and beliefs about, outcomes. The canonical explanation is subjective expected utility theory (Savage 1954), but only non-expected utility theories can account for phenomena such as ambiguity aversion (Ellsberg 1961). Furthermore, new methods such as distributional reinforcement learning (Bellemare et al. 2017) show how it is readily possible to maintain more information about outcome distributions than merely their expectation. A popular proposal inspired by robust control as well as finance is that decision makers use such information to hedge against model uncertainty by optimizing for the worst case, as with robustness preferences (Hansen & Sargent 2000), variational preferences (Maccheroni et al. 2006), or by using convex risk measures (Föllmer & Knispel 2012) such as the Conditional Value at Risk (Gagne & Dayan 2022). Based on behavior alone it is often not possible, however, to disambiguate between ambiguity preferences and distorted beliefs. The goal of this project is to inform (and distinguish between) these models, by making theory-driven inferences about participants' preferences and beliefs from their behavior and prediction-error-related responses (as measured in BOLD signal). We introduce a theoretical framework that makes precise what these behavioral and prediction error measurements reveal about preferences and beliefs. We will also present behavioral pilot data from a lottery task in which the uncertainty about the winning odds gradually resolves. Eliciting participants' lottery valuations at different stages allows us to measure their evolution as ambiguity resolves - as a first application of the theory. Better identification of preferences and beliefs will become possible when supplementing the behavioral data with neural measurements.