ICM 2026 Invited Section Lecture - Moritz Hardt

The Future of Prediction: in Conversation with Oskar Morgenstern A century ago, economist Oskar Morgenstern studied what he called one of the most central and most difficult problems in the theory of prediction: predictions that actively influence outcomes. Self-fulfilling and self-negating prophecies are two extreme cases of a fundamental problem that arises with all predictions about social outcomes. People act on predictions in ways that change the course of events, a phenomenon called performativity. Morgenstern argued that performativity doomed prediction in the social world to guesswork with unforeseeable results. This talk gives an entry point to the emerging area of performative prediction that provides counterpoints to Morgenstern's pessimistic prophecy. The standard formulation of machine learning as risk minimization assumes a fixed data-generating distribution. Performative prediction extends risk minimization toward data-generating distributions that depend on the predictive model. This extension leads to the powerful algorithmic principle of repeated risk minimization, alternating risk minimization and model deployment indefinitely. Under suitable conditions, repeated risk minimization converges to a stable point, a kind of equilibrium at which the predictive model minimizes risk on the distribution it entails. We'll discuss three perspectives on performativity using ideas from economic modeling, optimization, and the theory of forecasting that contribute principled ways to reason about predictions that influences outcomes. Uploaded 06-10-26 Learn more about your ad choices. Visit megaphone.fm/adchoices