Horizon-Constrained Rashomon Sets for Chaotic Forecasting

A new paper from researchers Gauri Kale, Rahul Vishwakarma, Holly Diamond, Ava Hedayatipour, and Amin Rezaei establishes the first rigorous connection between chaos theory and predictive multiplicity—the phenomenon where many equally accurate models exist for the same prediction task. Traditionally, Rashomon sets (the collection of near-optimal models) are considered static, but for chaotic systems like weather or power grids, tiny initial differences among models explode exponentially over time. The authors formalize this as horizon-constrained Rashomon sets, proving that the effective set of models shrinks exponentially with prediction horizon at a rate set by the maximum Lyapunov exponent. They also introduce Lyapunov-weighted metrics that give tighter bounds on how much models disagree as forecasts extend further into the future.

To make this theory actionable, the team develops decision-aligned selection algorithms that pick models based on downstream utility rather than raw forecast accuracy. In experiments on synthetic systems (Lorenz-96, Kuramoto–Sivashinsky) and real-world datasets for wind power, traffic, and weather, their approach improves decision quality by 18–34% while preserving competitive predictive performance. This work offers principled guidance for deploying machine learning in safety-critical chaotic domains—like energy load forecasting, extreme weather warnings, or autonomous navigation—where understanding model uncertainty over time is essential. The paper appears in AIP Advances and on arXiv (2605.05218).