← News

ResearchSources: arXiv · 18 Nov 2026

New benchmark and simulation-trained model target seasonal rain forecasts

Researchers behind SeasonRainBench say their model, trained partly on climate simulations, beats the strongest operational or data-driven baseline on several scores for monthly precipitation one to five months ahead.

In plain words

Seasonal rain forecasting still depends on costly physics-based ensemble systems, and no shared global test existed to judge cheaper, data-driven alternatives. The new benchmark supplies that test, and a companion model learns from simulated climates to offset the shortness of real-world observations. Its authors report gains over the best competing method on four verification measures.

What was announced

The arXiv abstract describes three obstacles slowing the shift from numerical ensembles to learned forecasting: no global benchmark, scoring methods that miss operational priorities, and training data limited to one short observational record. SeasonRainBench addresses the first two by pairing satellite-based global precipitation at one-degree resolution with ocean and atmosphere predictors, covering forecasts one to five months out. It draws in seven operational centers and 310,944 simulated months from 52 climate models, and adds two metrics aimed at anomaly strength and the monthly position of the rain belt.

For the third obstacle, the authors built a model that transfers knowledge from simulations to Earth through three components: one aligning the simulated climate background, one reusing a history encoder pretrained on simulations, and one supplementing real labels with analogous simulated outcomes. They report improvements over the strongest operational or data-driven competitor of 12.0 percent in anomaly correlation, 3.6 percent in CRPS, 2.9 percent in anomaly strength, and 0.3 percent in rain-belt position. The figures come from the authors' own experiments and have not been independently verified.

What it does not show

  • Evidence is limited to the paper's abstract; methods, baselines and results cannot be checked in full.
  • The reported performance gains are the authors' own experimental results, not independently confirmed.
  • The abstract does not state whether the model is available for operational use.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.