Ceiling of Predictability in Real-World Social Systems

Determine where the ceiling of predictability lies in real-world social systems by identifying the fundamental upper bound on forecasting accuracy for real-world events under realistic information constraints.

Background

The paper situates its contribution within decades of research on human forecasting and the emerging potential of LLMs to reason about uncertain future events. Although prior work acknowledges that social systems possess an inherent limit to predictability, the precise location of this upper bound remains unknown.

Establishing the ceiling of predictability is essential for interpreting progress in forecasting and for understanding how far training LLMs can improve predictive accuracy and calibration in complex real-world domains.