GenCast-Persisted: Seasonal Forecasting

• The paper demonstrates that GenCast, a medium‐range probabilistic forecast model, can be extended into seasonal predictions by using fixed sea-surface temperature anomalies.
• It employs an autoregressive rollout from November through February to isolate atmospheric dynamics from evolving ocean conditions.
• Quantitative evaluations reveal improved skill in higher latitudes and mountainous regions, while also highlighting overconfidence in some probabilistic estimates.

GenCast-Persisted is a seasonal forecasting configuration built from GenCast, a probabilistic machine-learned weather prediction model, in which sea-surface temperature anomalies are prescribed above a daily climatology for the full forecast horizon rather than being interactively predicted. In the explicit formulation reported in the seasonal forecasting study, GenCast-Persisted rolls a medium-range diffusion-based weather model forward autoregressively over December–February, using persisted SST anomaly forcing from a 1 November initialization to probe how far short-timescale learned atmospheric dynamics extend into the seasonal regime (Antonio et al., 8 Sep 2025). This seasonal use sits within a broader GenCast literature in which the base model was introduced as a 15-day global ensemble forecaster at 12-hour steps and 0.25 degree resolution (Price et al., 2023), while several adjacent papers either do not define a GenCast-Persisted variant or use “persistence” in different senses (Rathkopf, 11 Apr 2025).

1. Model lineage and scope of the term

GenCast itself is a diffusion-based ensemble weather forecasting model. In its original medium-range form, it generates stochastic 15-day global forecasts at 12-hour steps and 0.25 degree latitude–longitude resolution for over 80 surface and atmospheric variables, and it was reported to have greater skill than ECMWF ENS on 97.4% of 1320 evaluated targets (Price et al., 2023). The seasonal study does not retrain a distinct seasonal foundation model; instead, it takes the state-of-the-art probabilistic MLWP model and applies it beyond its native medium-range regime by autoregressive rollout over four months, using the 1° GenCast model because the 0.25° version could not be run due to GPU limits (Antonio et al., 8 Sep 2025).

In this literature, “GenCast-Persisted” has a specific technical meaning in the seasonal paper: prescribed SST anomalies persisted above daily climatology for the entire forecast period. That meaning is narrower than generic meteorological “persistence” baselines, and it is narrower than later interpretive uses of “persisted” for deployment artifacts, conformal padding, or ensemble state storage. A common misconception is therefore to read GenCast-Persisted as either a trivial last-observation-carried-forward forecast or a universal GenCast baseline. The seasonal paper does neither; it defines an atmospheric-only seasonal experiment with prescribed SST boundary conditions (Antonio et al., 8 Sep 2025).

2. Persisted SST forcing and seasonal adaptation

The seasonal configuration starts from the fact that GenCast was trained on short timescales and lacks an interactive ocean. To extend it to seasonal forecasting, the model is rolled out autoregressively over four months, with sea-surface temperature prescribed as a boundary condition. In GenCast-Persisted, SST anomalies are held fixed relative to climatology throughout the December–February forecast. Initialisation time is $t_0 = 1$ November, and the daily SST climatology is computed from ERA5 for 1 Jan 1979–12 Dec 2018 (Antonio et al., 8 Sep 2025).

The forcing is defined by

$$\mathrm{SST}(x,t)=\mathrm{SST}_{\mathrm{clim}}(x,t)+a(x),$$

with

$$a(x)=\mathrm{SST}_{\mathrm{obs}}(x,t_0)-\mathrm{SST}_{\mathrm{clim}}(x,t_0).$$

The anomaly $a(x)$ is time-invariant over the DJF forecast and is added to the appropriate daily climatological cycle $\mathrm{SST}_{\mathrm{clim}}(x,t)$ for each forecast day. No sea-ice boundary information is provided to GenCast. The contrasting configuration, GenCast-Forced, replaces persisted anomalies with the evolving ERA5 SST time series itself during DJF, while leaving the ensemble size, rollout, variables, and resolution unchanged (Antonio et al., 8 Sep 2025).

This construction matters because it isolates one of the principal bottlenecks in extending medium-range atmospheric ML models to seasonal horizons: the treatment of slow boundary conditions. GenCast-Persisted asks how much seasonal signal can be recovered when the atmospheric model is supplied only with an SST anomaly frozen above climatology, whereas GenCast-Forced asks how much error is attributable to SST evolution rather than to the atmospheric model.

3. Experimental design and verification framework

The seasonal experiments are initialized on 1 November for each year 2004–2024, described as 20 winters, and rolled out to the end of February, with DJF seasonal means verified afterward. Although 2004–2018 overlaps GenCast’s training era, the seasonal autoregressive rollout is far beyond training timescales and is treated as out-of-sample for this experiment. The model advances at 12-hour steps and uses 20 stochastic members for each GenCast configuration; the SEAS5 comparator also uses 20 ensemble members (Antonio et al., 8 Sep 2025).

The evaluated variables are precipitation, 2-metre temperature, and mean sea level pressure. Precipitation is taken as 12-hour accumulation anomalies aggregated to DJF; 2mT and MSLP are evaluated as DJF means. All DJF fields are aggregated by season, and ACC and reliability are computed on detrended anomalies to remove the climate change signal. Teleconnection case studies focus on strong La Niña DJF 2010–2011 and strong El Niño DJF 2015–2016, while subregional precipitation distributions are examined over Indonesia, southern–central Africa, and Mexico–USA (Antonio et al., 8 Sep 2025).

For 2mT and MSLP, the anomaly correlation coefficient is

$$\mathrm{ACC}=\frac{\sum_i f'_i o'_i}{\sqrt{\sum_i (f'_i)^2}\sqrt{\sum_i (o'_i)^2}},$$

where $f'$ and $o'$ are anomalies relative to climatology after detrending. For significance of $\mathrm{ACC}>0$, a one-sided $t$-test uses the statistic $$\mathrm{SST}(x,t)=\mathrm{SST}_{\mathrm{clim}}(x,t)+a(x),$$0 at 95% confidence, and differences between forecast correlations are tested with the Siegert et al. (2017) method to account for shared correlation between competing forecasts (Antonio et al., 8 Sep 2025).

The observational reference is ERA5 reanalysis, treated as “truth,” which is methodologically consequential because GenCast is also trained on ERA5. The seasonal paper explicitly did not report CRPS or Brier scores, even though formulas for both are listed for reference in the synthesis. Its main probabilistic diagnostics instead emphasize ensemble distributions and reliability diagrams for tercile events (Antonio et al., 8 Sep 2025).

4. Skill, calibration, and uncertainty structure

The central performance picture is regionally differentiated. SEAS5 is superior in the tropics for 2-metre temperature and mean sea level pressure, whereas GenCast-Persisted achieves significantly higher skill in some higher-latitude areas, including mountainous regions, and shows notable improvements for MSLP in particular. The uncertainty in precipitation response, as represented by the ensemble, compares favourably to SEAS5, but reliability diagrams indicate that GenCast-Persisted is overconfident relative to SEAS5 for seasonal 2mT probabilities (Antonio et al., 8 Sep 2025).

Region	2mT ACC (Persisted / Forced / SEAS5)	MSLP ACC (Persisted / Forced / SEAS5)
Tropics	0.62 / 0.88 / 0.74	0.62 / 0.71 / 0.76
Northern extratropics	0.25 / 0.51 / 0.28	0.22 / 0.21 / 0.21
Southern extratropics	0.42 / 0.76 / 0.54	0.48 / 0.54 / 0.50

These values show the dual character of GenCast-Persisted. On the one hand, it retains meaningful seasonal signal despite short-timescale training, especially in extra-tropical circulation and over complex terrain. On the other hand, prescribing SST anomalies by persistence leaves substantial tropical skill on the table, especially for 2mT over ocean, where GenCast-Forced rises to 0.88 ACC in the tropics and 0.76 in the Southern extratropics. The seasonal study interprets this as evidence that realistic SST evolution is sufficient to produce well-calibrated seasonal 2mT probabilities in GenCast, while the persisted configuration remains overconfident, especially at low predicted-probability bins (Antonio et al., 8 Sep 2025).

Spatially, SEAS5 is significantly better for 2mT in the tropics and maritime continent, whereas GenCast-Persisted is comparable across large parts of the extra-tropics and significantly better in mountainous regions such as the Andes, Rockies, and Alps, as well as slightly better in the NW Atlantic. For MSLP, SEAS5 significantly outperforms GenCast-Persisted across the tropics, northern Africa, South America, and the Sea of Okhotsk, but GenCast-Persisted shows pockets of improvement in northern Canada, northern Asia, and parts of the North Atlantic (Antonio et al., 8 Sep 2025).

5. Teleconnections, circulation, and seasonal error modes

A major result is that GenCast-Persisted reproduces much of the correct precipitation response to ENSO. In La Niña 2010–2011, both GenCast-Persisted and GenCast-Forced reproduce drying over the tropical Pacific and wetting over the maritime continent, consistent with ERA5 and SEAS5; drying around the Gulf of Mexico is stronger in GenCast than in SEAS5. In El Niño 2015–2016, all models capture the correct wetting and drying over the tropical Pacific and maritime continent, but GenCast-Persisted erroneously wets the tropical Atlantic, and GenCast-Forced corrects this error. Additional mismatches in Hudson Bay and parts of the Indian Ocean are likewise reduced when the evolving ERA5 SST boundary is used (Antonio et al., 8 Sep 2025).

These case studies show that the persisted-SST approximation is not merely a weak baseline. It preserves enough large-scale forcing to recover canonical ENSO teleconnections and realistic ensemble spreads in subregional precipitation distributions over Indonesia, Mexico–USA, and southern–central Africa, with ERA5 lying within the forecast distributions. At the same time, the correction from GenCast-Persisted to GenCast-Forced identifies a concrete class of ocean-boundary errors: where SST evolution matters materially, the fixed-anomaly approximation can create regionally coherent but wrong precipitation signals (Antonio et al., 8 Sep 2025).

The circulation result most often highlighted is the North Atlantic Oscillation. The study defines

$$\mathrm{SST}(x,t)=\mathrm{SST}_{\mathrm{clim}}(x,t)+a(x),$$1

with the Azores box $$\mathrm{SST}(x,t)=\mathrm{SST}_{\mathrm{clim}}(x,t)+a(x),$$2 and Iceland box $$\mathrm{SST}(x,t)=\mathrm{SST}_{\mathrm{clim}}(x,t)+a(x),$$3. Each time series is centred over the 20-year period and normalised by the ERA5 NAO standard deviation. Correlations with ERA5 DJF NAO are $$\mathrm{SST}(x,t)=\mathrm{SST}_{\mathrm{clim}}(x,t)+a(x),$$4 for SEAS5, $$\mathrm{SST}(x,t)=\mathrm{SST}_{\mathrm{clim}}(x,t)+a(x),$$5 for GenCast-Persisted, and $$\mathrm{SST}(x,t)=\mathrm{SST}_{\mathrm{clim}}(x,t)+a(x),$$6 for GenCast-Forced; the latter two are significant at 5%, while SEAS5 is not. All systems nonetheless exhibit the signal-to-noise paradox by underestimating NAO variance relative to observations (Antonio et al., 8 Sep 2025).

6. Reliability, epistemic context, and boundaries of the label

The broader GenCast literature places this seasonal configuration inside a larger discussion of reliability, calibration, and workflow design. In the hallucination and reliability analysis, GenCast is used as a case study for how theory-rich training data, physics-informed losses, and ensemble-based inference can neutralize “corrosive hallucinations”; that paper emphasizes ERA5 reanalysis, physics-informed constraints, stochastic ensemble generation, and spread–skill calibration, but it explicitly does not mention “GenCast-Persisted,” define “persistence,” or report any persistence baseline (Rathkopf, 11 Apr 2025). This matters because it separates the seasonal boundary-condition experiment from the epistemic argument about reliability in generative scientific models.

Several adjacent papers make the terminological boundary even sharper. TCBench includes persistence as a simple reference forecast and also uses persistence-fill for fair comparison when a model misses a verification pair, but it states that there is no named “GenCast-Persisted” configuration in that benchmark (Gomez et al., 30 Jan 2026). The ATLAS medium-range comparison paper treats GenCast as a baseline diffusion model and contains no mention of a “GenCast-Persisted” variant (Kossaifi et al., 26 Jan 2026). The conformal prediction study likewise states that it does not define or report a GenCast-Persisted model, although it proposes persisted conformal-padding variants as a possible operational construction external to the paper’s baseline setup (Asch et al., 17 Jun 2026).

Related but non-equivalent uses of “persisted” also appear in post-training and data-assimilation contexts. The RMMD distillation work applies reward-guided moment matching to GenCast and states that, if one adopted a “GenCast-Persisted” operational concept there, it would mean saving and deploying the RMMD-distilled checkpoint for production use rather than defining a persistence forecast metric (Jacq et al., 29 Jun 2026). The training-free data-assimilation paper likewise does not use the term, but its particle-filter framework naturally supports persisting ensemble states and weights across assimilation cycles (Savary et al., 23 Sep 2025). A plausible implication is that the label “GenCast-Persisted” is stable only in the seasonal SST-forcing paper; elsewhere, persistence refers either to a generic meteorological baseline or to operational state persistence rather than to a single canonical GenCast variant.

7. Scientific significance and limitations

GenCast-Persisted is scientifically significant because it demonstrates that a probabilistic ML weather model trained for the medium range can recover substantial seasonal structure under a simple prescribed-SST forcing scheme. It captures canonical ENSO precipitation teleconnections, achieves competitive or superior skill in higher latitudes and mountainous regions for 2mT and MSLP, and correlates more strongly than SEAS5 with the observed DJF NAO over 2004–2024. The seasonal paper therefore uses it as evidence that MLWP models similar to GenCast have potential for the “full” seasonal forecasting problem, especially once coupled to ocean, land, and cryosphere models (Antonio et al., 8 Sep 2025).

The limitations are equally explicit. GenCast-Persisted has no interactive ocean, land, or cryosphere; it receives no land-surface state and no sea-ice input; it was evaluated without the hindcast bias correction typically used in seasonal systems; and it is overconfident in tercile-based seasonal 2mT probabilities. Some of its salient errors, such as tropical Atlantic wetting during El Niño, are corrected when the model is supplied with evolving ERA5 SST rather than persisted SST anomalies. These results imply that the persisted-SST configuration is informative but incomplete: it is a controlled seasonal extension of a medium-range atmospheric diffusion model, not a finished coupled seasonal forecasting system (Antonio et al., 8 Sep 2025).

Within the synthesized literature, the most precise definition is therefore narrow. GenCast-Persisted is the prescribed-SST seasonal adaptation of GenCast introduced in the seasonal forecasting study, and it should not be conflated with persistence baselines, persistence-filled verification protocols, persisted conformal post-processing, persisted data-assimilation states, or persisted distilled checkpoints.