Seasonal rain forecasts get a common arena: a data-driven model matches the strongest operational center across eleven metrics
Synopsis
Seasonal precipitation forecasting still depends on computationally expensive numerical ensemble systems and has lacked a global benchmark for comparing data-driven models against them. SeasonRainBench places seven operational centers, 310,944 simulated months from 52 CMIP6 climate models and GPM IMERG satellite precipitation under one protocol, adding metrics for anomaly strength and rain-belt position. Built on it, Sim2Earth pretrains on simulations and transfers to Earth observations, improving ACC, CRPS, anomaly strength and rain-belt position by 12.0%, 3.6%, 2.9% and 0.3% over the strongest operational or data-driven model.
Interpretation
For the first time, global monthly precipitation forecasts one to five months ahead can be compared under one protocol, pitting data-driven models against seven operational centers. Earlier benchmarks covered only medium-range, subseasonal or regional seasonal forecasting, leaving no unified protocol for comparing global data-driven precipitation forecasts with operational systems. SeasonRainBench verifies against GPM IMERG satellite precipitation, includes the monthly ensembles of seven Copernicus centers for 2020 to 2025, and uses a test set of 345 issue-lead pairs issued from January 2020 with targets through December 2025.
Scoring no longer rests only on anomaly correlation and squared error; it reports the strength of flood-season precipitation anomalies and the latitude of the monthly main rain belt. ACC and RMSE summarize association and squared error but do not directly state the anomaly strength and rain-belt location that operational flood-season outlooks report. On paired forecasts built from IMERG observations that differ in only one of these properties, standard scores reward damped anomalies and prefer differently placed belts than operational metrics do, motivating the AGSS and RBLE metrics.
Sim2Earth treats climate simulations as the source of what the observational record lacks, then transfers to the real Earth in three stages. Satellite precipitation offers fewer than thirty samples per calendar month, each the only observed outcome of its ocean state. Training uses 158 of 52 CMIP6 models' members that supply simulated months, 310,944 training months in total, with nine members held out; the model is then fine-tuned in turn on reanalysis and satellite observations.
Perspective
The demonstrated scope is global monthly precipitation at leads of one to five months, verified against GPM IMERG satellite precipitation and compared with seven Copernicus operational centers over 2020 to 2025; the test set is 345 issue-lead pairs issued from January 2020 with targets through December 2025.
The native ensembles of all seven centers score below climatology on the probabilistic scores, with negative CRPSS and RPSS; Sim2Earth is the only system whose tercile probabilities improve on climatology, at an RPSS of 0.0619. Extending the historical simulations beyond 2014 relies on SSP2-4.5 runs of 38 models, and not every member has such an extension.
