Inscrutable decision makers: Knightian uncertainty in machine learning

Abstract

In building models that causally explain observed data and future data, econometricians must grapple with quantifiable uncertainty, or risk, and unquantifiable Knightian uncertainty, or ambiguity. In contrast, machine learning practitioners work with statistical models for a data set that enable predictions about data items imputed to be in the data set. Recently these two distinct modeling concepts have become topics of mutual interest in economics and machine learning. We take the viewpoint here that a data set implicitly embodies the ambiguity of the generating processes from which it arises. We present a data model incorporating ambiguity that we dub the Inscrutable Decision Maker (IDM) derived from the Anscombe-Aumann model of subjective utility.