Probabilistic forecasting for high-dimensional nonlinear dynamical systems is computationally challenging because propagating predictive distributions requires solving distributional evolution equations, which is often prohibitive. Conventional approaches therefore approximate multi-step distributions using ensembles or sampled trajectories.
The authors propose a distribution-to-distribution (D2D) neural probabilistic forecasting framework. It encodes input distributions via kernel mean embeddings and parameterises outputs with mixture density networks, enabling recursive multi-step forecasting directly in distributional form. The neural forecasting module itself is replaceable, making the framework flexible.
Tests on the Lorenz63 system show that the D2D model captures nontrivial distributional evolution and produces skillful probabilistic forecasts without explicit ensemble simulation. Its performance remains competitive with a naively initialised perfect-model benchmark, offering proof-of-concept that predictive distributions can be recursively propagated through a neural model and evaluated in terms of probabilistic forecast skill.