Large-scale statistical forecasting models reassess the unpredictability of chaotic systems

Chaos and unpredictability are often considered synonymous, yet recent advances in statistical forecasting suggest that large machine learning models gain unexpected insight from extended observation of complex systems. We perform a large-scale comparison of 24 state-of-the-art multivariate forecasting methods on a crowdsourced database of 135 distinct low-dimensional chaotic systems. Large, domain-agnostic time series forecasting methods consistently exhibit the strongest performance, producing accurate predictions lasting up to two dozen Lyapunov times. The best-performing models contain no inductive biases for dynamical systems, and include hierarchical neural basis functions, transformers, and recurrent neural networks. However, physics-based hybrid methods like neural ordinary differential equations and reservoir computers perform more strongly in data-limited settings. Diverse forecasting methods correlate despite their widely-varying architectures, yet the Lyapunov exponent fails to fully explain variation in the predictability of different chaotic systems over long time horizons. Our results show that a key advantage of modern forecasting methods stems not from their architectural details, but rather from their capacity to learn the large-scale structure of chaotic attractors.