Learning by Convex Combination,
2026-07-10
We study how an agent evaluates an action when she observes a sample of its outcomes rather than its outcome-generating distribution. We propose a model where the agent’s estimated value for the action is a convex combination of her average utility over the sample outcomes and an ex ante utility reflecting her prior information. The weight put on the average utility increases with sample size, reflecting the inferential advantages of larger samples. The model nests certain forms of Bayesian behaviour and, more generally, identifies parameters quantifying departures from Bayesian updating, such as conservatism and the Law of Small Numbers.