ICM 2026 Invited Section Lecture - Moritz Hardt

Value of the Information & Strength of the Argument

The talk provides a rigorous and insightful analysis of a fundamental issue in predictive modeling. Hardt clearly motivates the problem with relatable examples, then formalizes it mathematically. He discusses both the limitations of existing convergence guarantees and introduces a conceptually simpler ‘stable signals’ perspective that addresses the two main shortcomings of prior work: the assumption of small performative effects and the lack of interpretability of stable points. The argumentation is well-structured, logical, and builds on a rich historical context. The speaker effectively balances technical depth with accessibility, making the session valuable for both specialists and interested non-specialists.

Scientific Rigor, Source Quality, Title Accuracy

The lecture demonstrates high scientific rigor: definitions are precise, assumptions are clearly stated, and known hardness results are referenced. Hardt cites key works including Morgenstern’s habilitation, Grunberg and Modigliani, Herbert Simon, and recent papers in performative prediction. He openly discusses the limitations and open questions. The title is adequate: it is a standard academic lecture title, not misleading. The content is consistent with the expected scope of an ICM invited talk. No commercial or promotional content is present.

The title is generic but accurately reflects the content: an invited lecture at ICM 2026 by Moritz Hardt on performative prediction.

9/10

The talk presents rigorous formal definitions and results from peer-reviewed research, with clear arguments and appropriate caveats. The speaker is a recognized expert in the field.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 00:00Introduction and setup of risk minimization
• 05:00The astronomical conception of society and performativity examples (YouTube, Google Maps)
• 10:00Oskar Morgenstern's 1928 habilitation and his pessimism about prediction
• 15:00Early counterarguments by Grunberg, Modigliani, and Simon; fixed-point argument
• 20:00Formalization of performative prediction and repeated risk minimization
• 25:00Convergence theorems and hardness results
• 30:00Limitations of existing assumptions and introduction of the 'stable signals' perspective
• 35:00Discussion and conclusion

• Performative Prediction (Perdomo, Zrnić, Mendler-Dünner, Hardt, 2020) — Introduced the formal framework of performative prediction and repeated risk minimization.
• Morgenstern's Habilitation (1928) — Original argument that predictions affecting outcomes invalidate their empirical basis.
• Grunberg & Modigliani: The Predictability of Social Events — Early attempt to show that predictions can be made despite performativity using continuity arguments.
• Simon: Bandwagon and Underdog Effects and the Possibility of Election Predictions — Used intermediate value theorem to argue for existence of self-fulfilling predictions.

• Performative Prediction (Perdomo et al.) — Foundational work cited and extended in the talk.
• Grunberg-Modigliani paper — Historical counterpoint to Morgenstern's pessimism.

• Rational expectations literature — Assumes agents' expectations are consistent with model outcomes, but does not fully address feedback from predictions.