Randomized reservoir-computing ensembles extended the reported short-term forecast horizon in a Hénon-map test and produced slower error growth than individual members in Lorenz 96. In the Hénon case, the ensemble mean reached about 25 forecast steps, while individual-model horizons ranged from about 20 to 25 steps; in Lorenz 96, ensemble forecasts were reported to remain successful for at least 2.5 Lyapunov times.
The work also tested the folded towel map, where the reported finding concerned the dimension of forecast vectors after several steps. The experiments covered deterministic numerical systems, so they do not directly validate operational weather or climate models or establish a universal forecast horizon across chaotic systems.
How the forecasts were combined
Each reservoir computer was trained on the same input data with identical hyperparameters. The models differed in randomly chosen connections inside the reservoir and in their input weights. Researchers compared a plain arithmetic average, a performance-weighted average and a closed-loop version that fed the previous-step ensemble average back to each reservoir.
One Hénon-map experiment examined 157,500 individual reservoir computers. Their errors varied widely and formed long, power-law-like tails toward exceptionally large errors, with rare very poor models.
Errors in iterated Hénon-map forecasts grew roughly exponentially with forecast step, at a rate slightly above the maximal Lyapunov exponent.
The clearest result came from the Hénon map
Among 100 Hénon-map ensemble members, individual errors varied by almost two orders of magnitude. The arithmetic mean was almost as accurate as the best individual model, and the weighted mean performed better still.
Forecast-error clouds concentrated along the system’s unstable manifold, the direction associated with growing errors. Around step 10, spread along that manifold was about 65 times larger than spread across it, and the concentration weakened after roughly step 17 as curvature became relevant.
Weighting helped, but bigger was not always better
The performance-weighted arithmetic mean performed better than the plain mean and narrowed the error distributions by reducing the contribution of poor forecasts. The rule gave each model a weight equal to the inverse square root of its five-step-ahead test error; the authors described it as ad hoc and did not claim it was optimal.
Researchers evaluated 250 randomly constructed ensembles at each fixed ensemble size. Larger groups shifted errors toward smaller values, but doubling the ensemble size reduced errors by less than the ideal 1/√2 benchmark.
The authors associated that slower-than-ideal pattern with heavy-tailed individual-model errors. The analysis reports no formal uncertainty for the tail behavior or inferential statistical test.
Results varied across the test systems
In the folded towel map, forecast vectors after more than five steps showed a crossover to a correlation dimension of 2, matching the dimension associated with the system’s two positive Lyapunov exponents.
For Lorenz 96, hyperparameters were selected by grid search. Ensemble forecasts were reported to remain successful for at least 2.5 Lyapunov times before pointwise error became noticeable, with slower error growth than individual members. Exact comparative effect sizes were not reported.
The evidence is limited to numerical reservoir-computing experiments in three deterministic chaotic systems. It does not show that the same strategy improves real-world weather or climate forecasts or establish a universal forecast horizon across chaotic systems.
The weighting rule was not shown to be optimal, and the paper reports no confidence intervals or formal statistical tests. The size and reliability of the reported gains therefore remain uncertain.
Paper data and sources
Original title: Understanding the superiority of multi-model ensemble forecasts through reservoir computing
Authors: Daniel Estevez Moya, Francesco Martinuzzi, Edmilson Roque dos Santos et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text