Horizon-Constrained Rashomon Sets for Chaotic Forecasting

A novel theoretical framework connects predictive multiplicity with chaotic dynamics within the realm of machine learning. This study presents horizon-constrained Rashomon sets, which illustrate the evolution of model multiplicity in chaotic systems as the prediction horizon changes. In contrast to static tasks, chaos leads to an exponential divergence among models that start similarly. The effective Rashomon set decreases exponentially with increased lead time, influenced by the maximum Lyapunov exponent. Additionally, Lyapunov-weighted metrics offer more precise limits on predictive discrepancies. Algorithms for decision-aligned selection evaluate nearly optimal models based on their downstream utility. This research was published on arXiv (2605.05218).