acceptodds
Under review as a conference paper at ICLR 2027

Marchuk-S2S: Compact Latent Flow Matching for Six-Hourly Subseasonal Ensemble Weather Forecasting

Abstract

Subseasonal forecasts two to six weeks ahead must draw predictability from slowly evolving modes such as the Madden–Julian Oscillation (MJO) after the memory of the initial state has faded, and are therefore issued as large ensembles. Diffusion- and flow-based weather models generate sharp six-hourly ensembles but have been evaluated mainly in the medium range. We introduce Marchuk-S2S, a 276M-parameter Transformer trained with conditional flow matching in the latent space of a frozen pretrained weather autoencoder. Conditioned on one atmospheric state and time-of-year labels, it jointly samples multi-day chunks of six-hourly states; a single checkpoint trained on 1- to 8-day chunks is composed autoregressively to 48 days, without lead-time-specific networks or gridded climatology. A 50-member, 48-day ensemble takes 9.7 minutes on one H100 GPU, faster than LaDCast. Over 2019–2021, Marchuk-S2S is competitive with FuXi-S2S and the 51-member ECMWF ensemble in weekly anomaly correlation and MJO skill. Training on six-hourly states and aggregating to daily means lowers CRPS on all five evaluated fields compared with training on daily means. Treating each call as a stochastic transition, we show that composing two chunks keeps the ensemble scale of one direct chunk of twice the length and lowers its CRPS, without a consistency loss.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
flow matching, generative models, weather forecasting, subseasonal prediction, probabilistic forecasting, latent models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_Wp4wOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Marchuk-S2S is a compact latent flow-matching model that generates multi-day chunks of six-hourly weather states from a single initial condition and time-of-year labels. A single network is trained on variable horizons and composed autoregressively to produce long subseasonal ensembles. The main contributions are showing that training on six-hourly states improves daily forecast skill compared with training on daily means, that composing shorter chunks can outperform direct longer-horizon generation, and that the resulting system is competitive with strong operational and learned subseasonal baselines while remaining computationally efficient.

Critical strengths
  1. The model produces long six-hourly forecasts that match strong baselines (ECMWF S2S and FuXi-S2S) at much lower computational cost. This makes high-resolution subseasonal ensembles more practical.

  2. Training on six-hourly states and then averaging gives better daily CRPS than training directly on daily means. This is a clear and useful finding.

Critical weaknesses
  1. The core components (latent-space generation, flow matching, multi-step/chunk generation, and autoregressive rollouts) are largely combinations of existing ideas. The paper’s main advance is system integration and efficiency rather than a new modeling principle. For a top-tier venue, this level of novelty is borderline.

  2. Results are mostly competitive rather than clearly superior, while precipitation skill and late-lead MJO skill, both highly relevant for subseasonal use cases, are noticeably worse than the baselines. Without stronger overall gains or clear progress on these important targets, the practical impact is limited.

Questions for authors
  1. Table 2 shows that composing two shorter chunks yields lower CRPS than generating one longer chunk directly, even without a consistency loss. Do you have any analysis or hypothesis for why the direct longer-horizon generation underperforms?

  2. The model shows clearly weaker precipitation anomaly correlation than both ECMWF and FuXi-S2S from week 3 onward. Is this mainly caused by information loss in the frozen latent autoencoder, by the flow-matching objective struggling with precipitation’s intermittent distribution, or by something else?

2 comments
AuthorsOpenReview ↗

Author Responses (1/2)

We thank the reviewer for the helpful suggestions. Our point-by-point responses follow.

W1

We agree that the individual components are established. However, dense six-hourly probabilistic forecasting at subseasonal lead times remains underexplored. Marchuk-S2S uniquely combines latent flow matching, variable-horizon joint chunk generation, and autoregressive composition to produce stable 48-day forecasts using a single checkpoint, without lead-time-specific models or gridded climatology. This integration enables efficient dense S2S ensembles and reveals that six-hourly training improves daily forecast skill. In the following, we provide further analysis of these design choices.

W2

On most fields Marchuk-S2S is competitive with FuXi-S2S and the ECMWF S2S ensemble, while precipitation anomaly skill and MJO skill are weaker. This is the question posed in Section 1: whether one compact latent flow-matching model, trained as a single variable-horizon checkpoint on multi-day chunks of six-hourly states and composed to 48 days, can remain competitive with dedicated S2S systems. The experiments show that this architecture, this training pipeline, and six-hourly forecasting at subseasonal leads work well on the variables represented in the latent space. The corresponding scores, significance tests, six-hourly training comparison, and inference cost are reported in Figures 1–2, Table 2, and Section 3.6.

The precipitation and late-lead MJO gaps come from the frozen LaDCast latent space, which does not include outgoing longwave radiation and has limited precipitation reconstruction fidelity. Appendix A.7 specifies the shared precipitation-based RMM-like index and the matched comparison. Replacing OLR with precipitation is established practice for ML weather models [1], and it is noisier, so the absolute COR values lie below the OLR-based correlations reported by FuXi-S2S. The frozen DC-AE already introduces substantial precipitation error before any forecast is made; our response to Q2 shows how this error enters the predicted field. These scores can be improved by an autoencoder with higher precipitation fidelity and with OLR. Improving the autoencoder is beyond the scope of this work.

[1] Chien, M.-T., Barnes, E. A., and Maloney, E. D. Modulation of tropical cyclogenesis on subseasonal-to-interannual timescales in the deep-learning climate emulator ACE2, 2025.

AuthorsOpenReview ↗

Author Responses (2/2)

Q1

We hypothesize that the result reflects a trade-off between autoregressive error accumulation and single-transition difficulty. Direct long-horizon generation must represent a broader and more nonlinear conditional distribution, whereas composition factorizes it into shorter, better-learned transitions. Conversely, very short chunks require more transitions, which can accumulate errors and introduce additional noise realizations.

To examine this trade-off, we evaluated 2-, 4-, and 8-day inference chunks over full rollouts. The 4-day chunk achieves the lowest CRPS in 8 of the 10 variable–lead-time comparisons. The 2-day chunk produces the largest ensemble spread in 9 cases, and the 8-day configuration is optimal under neither metric.

CRPS ↓

Inference chunkT2M, 15T2M, 45T500, 15T500, 45T850, 15T850, 45Z500, 15Z500, 45Z850, 15Z850, 45
2 days1.0391.0771.3781.4891.5111.631324.84349.97236.54238.11
4 days (ours)0.9911.0481.3601.4491.5001.591326.98348.58227.04238.97
8 days1.0031.0971.4131.4951.5631.654342.44357.96233.85247.22

Spread ↑

Inference chunkT2M, 15T2M, 45T500, 15T500, 45T850, 15T850, 45Z500, 15Z500, 45Z850, 15Z850, 45
2 days1.8151.8132.7392.9632.9493.152668.07713.05449.50493.42
4 days (ours)1.7261.8792.6162.8222.8503.065629.53682.09436.84470.45
8 days1.5701.7502.4262.7482.7982.784589.74664.45400.31443.23

Short chunks require more model transitions, so errors can accumulate [1, 2]. Long chunks require the model to represent a broader and more nonlinear conditional distribution in one step. The 4-day chunk is the best balance between these effects. The larger spread of the 2-day rollout is consistent with more frequent stochastic sampling: each transition draws an additional noise realization [3]. CRPS jointly evaluates accuracy, calibration, and sharpness [4], so the wider 2-day ensemble is not additional probabilistic skill. The 4-day chunk gives the best overall probabilistic accuracy.

CRPS and ensemble spread by inference chunk horizon. The 4-day inference chunk gives the best overall CRPS, whereas shorter chunks generally produce larger ensemble spread. Columns are lead times in days. Bold marks the best value in each column (lowest CRPS, largest spread).

[1] Pangu-Weather. Bi, Kaifeng, Lingxi Xie, Hengheng Zhang, Xin Chen, Xiaotao Gu, and Qi Tian. Accurate medium-range global weather forecasting with 3D neural networks. Nature, 2023.

[2] Nguyen, Tung, Rohan Shah, Hritik Bansal, Troy Arcomano, Romit Maulik, Veerabhadra Kotamarthi, Ian Foster, Sandeep Madireddy, and Aditya Grover. Scaling transformer neural networks for skillful and reliable medium-range weather forecasting. NeurIPS, 2024.

[3] Price, Ilan, Alvaro Sanchez-Gonzalez, Ferran Alet, Tom R. Andersson, Andrew El-Kadi, Dominic Masters, Timo Ewalds, Jacklynn Stott, Shakir Mohamed, Peter Battaglia, Remi Lam, and Matthew Willson. Probabilistic weather forecasting with machine learning. Nature, 2025.

[4] Gneiting, Tilmann and Adrian E. Raftery. Strictly proper scoring rules, prediction, and estimation. Journal of the American Statistical Association, 2007.

Q2

The weaker precipitation anomaly correlation from week 3 onward is a property of the frozen LaDCast representation. The flow-matching objective is applied only after that representation is fixed. The reconstruction error below is the encode–decode error of the frozen DC-AE on 6-hour precipitation, so it does not depend on the flow model or on lead time.

Lead timeDC-AE RMSE, mmMarchuk-S2S RMSE, mmDC-AE / Marchuk-S2S
1 day0.2020.31963%
5 days0.2020.46344%
10 days0.2020.50840%
15 days0.2020.51939%
30 days0.2020.52638%
45 days0.2020.53038%

The frozen DC-AE reconstructs 6-hour precipitation with an RMSE of 0.202 mm at every lead. At 1.5° on 2018, that reconstruction is 63% of the Marchuk-S2S precipitation RMSE at day 1, 44% at day 5, 39% at day 15, and 38% on days 30–45. The decoded forecast therefore starts from a distorted precipitation field, and week-3 anomaly correlation is computed from fields that still pass through the same decoder.

The last column is a ratio of RMSE values, rr. Independent errors add in the square, so if the reconstruction error were orthogonal to the remaining forecast error, removing it would lower the forecast RMSE by \sqrt[\]{1 - r^2}: about 23% at day 1 and about 8% from day 15 onward.

Reviewer_x2AVOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

It presents a generative weather forecasting model (Marchuk-S2S) to produce global ensemble predictions at six-hourly resolution up to 48 days ahead. The framework uses a 276M-parameter Transformer with conditional flow matching in a compressed latent space and jointly generates multi-day forecast sequences, which are combined through autoregressive rollout, allowing a single model to handle different forecasting horizons. The key contribution is an efficient approach to high-resolution subseasonal ensemble forecasting rather than a fundamentally new generative architecture. The model is evaluated using ERA5 data and compared with established forecasting systems, such as ECMWF S2S, FuXi-S2S, and LaDCast. Results show competitive subseasonal forecasting skill, particularly for geopotential height, while requiring significantly less computation. Additional experiments investigate the benefits of six-hourly training, forecast chunk lengths, rollout consistency, and ensemble calibration. However, precipitation forecasting remains relatively weak at longer lead times.

Critical strengths
  • Practical contribution with significant computational efficiency. It demonstrates that a relatively compact model can generate 50-member, six-hourly global ensemble forecasts up to 48 days ahead in just 9.7 minutes on a single H100 GPU. Its competitive performance against well-known subseasonal forecasting systems makes the efficiency gains particularly meaningful. This is a key strength, where computational cost and practical forecasting capability are important considerations.

  • Experimental analysis of key design choices. Beyond forecasting accuracy, it investigates the effects of 6-hourly vs daily training, forecast chunk lengths, rollout consistency, and ensemble calibration. The consistent CRPS improvements from 6-hourly training are particularly encouraging. These experiments justify the proposed design and provide useful insights for developing efficient subseasonal forecasting models.

Critical weaknesses
  • Limited benchmarking and statistical validation. Comparisons with recent subseasonal forecasting models, such as FengWu-W2S (Ling et al., 2024) and AIFS-SUBS (Schloer et al., 2026), remain incomplete. Differences in evaluation settings also make it difficult to establish a clear advantage over existing approaches. Furthermore, the evaluation covers only three test years without confidence intervals or statistical significance analysis.

  • While Marchuk-S2S performs well for several atmospheric variables, precipitation forecasting remains weaker than ECMWF and FuXi-S2S beyond week three, and MJO prediction skill deteriorates more rapidly at longer horizons. The ensembles are also mildly underdispersive. These are important for a subseasonal forecasting application, as they raise questions about the reliability and practical usefulness of the generated forecasts, particularly for high-impact weather events.

Questions for authors
  • The results are promising, but comparisons with recent subseasonal forecasting models, such asFengWu-W2S and AIFS-SUBS, are missing. Could you explain why these methods were not included and how Marchuk-S2S compares with them in terms of forecasting accuracy, temporal resolution and computational efficiency?

  • The evaluation is based on only 3 test years (2019–2021), without confidence intervals or statistical significance analysis. How convincing are the reported improvements over ECMWF and FuXi-S2S consistent across different seasons and forecast initialization dates?

  • The reported 7.6× speedup over LaDCast is very good. However, are both methods evaluated under comparable hardware, ensemble size and forecasting settings? Does the reported 9.7-minute runtime include all processing steps, particularly latent encoding and decoding?

  • Precipitation prediction is noticeably weaker beyond week three, while MJO skill also declines faster at longer lead times. Could you clarify how reliable the probabilistic forecasts are at weeks 4–6, particularly for high-impact weather conditions?

5 comments
AuthorsOpenReview ↗

Author Responses (1/2)

We thank the reviewer for these constructive comments. We address each point in turn below.

W1 & Q1

We agree that matched comparisons with FengWu-W2S and AIFS-SUBS would be valuable. However, we could not run these systems under our evaluation protocol. For FengWu-W2S, complete checkpoints and matched hindcasts were unavailable during our experiments. AIFS-SUBS provides a public 20-year reforecast archive (10 members, 46 days, instantaneous 00 UTC fields), but no period is out of sample for both models: its only held-out years, 2007–2011, lie within our training period (1979–2017), while our test years, 2019–2021, lie within its training data [2]. The archive also differs from our evaluation in initialization, resolution, ensemble size, and use of model-specific reforecast climatology, preventing a controlled numerical comparison.

Published results indicate that FengWu-W2S performs strongly for T2M, Z500, and MJO, while AIFS-SUBS matches ECMWF probabilistic skill over weeks 2–6. However, these numbers use different periods, climatologies, grids, and verification protocols, so they do not establish a reliable ranking against Marchuk-S2S.

As a controlled comparison, Figure 4 evaluate Marchuk-S2S against LaDCast-1.6B and Climatology through day 30 using the same 1.5° grid, 2019–2021 initializations, and dense six-hourly RMSE and CRPS. This comparison is relevant because ChaosBench shows that weather-scale models such as GraphCast, Pangu-Weather, and FourCastNetV2 can collapse to climatology at S2S horizons [3]. Swift demonstrates stable 75-day rollouts but explicitly interprets them as stability evidence rather than verified S2S skill [4].

W2 & Q4

Marchuk-S2S uses the frozen pretrained representation from LaDCast, which does not include OLR and has limited precipitation reconstruction fidelity. We therefore compute an RMM-like index from joint EOFs of U850, U200, and precipitation anomalies, projecting every system onto the same ERA5-derived EOFs.

Replacing OLR with precipitation is established practice for ML weather models because precipitation is a more common model output [5]. However, precipitation is noisier, resulting in lower absolute COR values for all evaluated systems than the OLR-based values reported by FuXi-S2S. The within-study comparison remains fair because all systems use identical EOFs, indices, and matched dates.

The weaker precipitation anomaly correlation from week 3 onward is a property of the frozen LaDCast representation. The flow-matching objective is applied only after that representation is fixed. The reconstruction error below is the encode–decode error of the frozen DC-AE on 6-hour precipitation, so it does not depend on the flow model or on lead time.

Lead timeDC-AE RMSE, mmMarchuk-S2S RMSE, mmDC-AE / Marchuk-S2S
1 day0.2020.31963%
5 days0.2020.46344%
10 days0.2020.50840%
15 days0.2020.51939%
30 days0.2020.52638%
45 days0.2020.53038%

The frozen DC-AE reconstructs 6-hour precipitation with an RMSE of 0.202 mm at every lead. At 1.5° on 2018, that reconstruction is 63% of the Marchuk-S2S precipitation RMSE at day 1, 44% at day 5, 40% at day 10, 39% at day 15, and 38% on days 30–45. The decoded forecast therefore starts from a distorted precipitation field, and week-3 anomaly correlation is computed from fields that still pass through the same decoder.

Regarding calibration, the underdispersion is mild: the subseasonal spread–skill ratio remains approximately 0.93–0.99 through days 15–45, rather than collapsing with lead time (Appendix Figure 15). Nevertheless, we agree that further calibration is required for operational use. Despite these limitations, Marchuk-S2S remains below climatology through day 30 and competitive with ECMWF and FuXi-S2S on the evaluated non-precipitation fields, while providing dense six-hourly ensembles.

[1] Ling, F., Chen, K., Wu, J., Han, T., Luo, J.-J., Ouyang, W., and Bai, L. FengWu-W2S: A deep learning model for seamless weather-to-subseasonal forecast of global atmosphere, 2024.

[2] Schloer, J., Tietsche, S., Roberts, C. D., Zampieri, L., Lang, S., Mertes, G., Jones, G., Chantry, M., and Vitart, F. AIFS-SUBS: Extending Data-Driven Forecasting to Sub-Seasonal Timescales, 2026.

[3] Nathaniel, J., Qu, Y., Nguyen, T., Yu, S., Busecke, J., Grover, A., and Gentine, P. ChaosBench: A Multi-Channel, Physics-Based Benchmark for Subseasonal-to-Seasonal Climate Prediction, 2024.

[4] Stock, J., Arcomano, T., and Kotamarthi, R. Swift: An Autoregressive Consistency Model for Efficient Weather Forecasting, 2025.

[5] Chien, M.-T., Barnes, E. A., and Maloney, E. D. Modulation of tropical cyclogenesis on subseasonal-to-interannual timescales in the deep-learning climate emulator ACE2, 2025.

AuthorsOpenReview ↗

Author Responses (2/2)

Q2

 We strengthened the statistical analysis in the revision. In Figure 1 the shaded bands are bootstrap standard errors: for each lead time, 1,000 resamples draw the 313 initializations from 2019–2021 with replacement. In Figure 2 a pale color marks lead times at which Marchuk-S2S does not improve on the ECMWF ensemble at the 97.5% confidence level, again from a bootstrap repeated 1,000 times. Appendix Tables 10–11 give paired 95% moving-block bootstrap intervals by season and initialization year.

The improvements are not uniform. Every difference whose 95% interval excludes zero favors Marchuk-S2S, but many do not: Z500 and Z850 RMSE are indistinguishable from zero in every season and year, and the significant RMSE reductions are limited to T850 and T2M in part of the sample. Fair CRPS is significant more often, especially in spring, summer, and across 2019–2021, and less often in winter and autumn. We therefore describe Marchuk-S2S as competitive, rather than consistently superior. Tables 10–11 test ECMWF only. The comparison with FuXi-S2S uses the same matched dates in Figure 2, and that sample does not support the same seasonal significance claims.

Appendix A.10, “Sensitivity to model drift and to the ERA5 climatology,” measures both changes on the 2019 and 2020 starts. Table 13 replaces the 1990–2017 ERA5 climatology of the forecast with each model’s own cross-year climatology, and Table 14 changes only the ERA5 base period, from 1980–2007 to the 1990–2017 period used in the main evaluation. Both changes lower weeks 3–6 TCC. The drop is shared by Marchuk-S2S, FuXi-S2S, and ECMWF, and it is larger for Z500, T850, and T2M than for U850 and precipitation. Absolute late-lead TCC is therefore sensitive to the climatology, and this sensitivity is not specific to Marchuk-S2S.

Absolute late-lead TCC is therefore sensitive to the climatology definition, and this sensitivity is shared by Marchuk-S2S, FuXi-S2S, and ECMWF. A longer independent evaluation remains necessary, and we retain the three-year period as an explicit limitation.

Q3

This has been clarified in the revision. The comparison uses the same hardware (one NVIDIA H100), ensemble size (50 members), forecast horizon (48 days), and output resolution (192 six-hourly states per member). LaDCast requires 73.3 minutes versus 9.7 minutes for Marchuk-S2S, giving the reported 7.6× speedup.

The 9.7-minute runtime includes encoding the initial state, all flow-matching evaluations, chunk transitions, and decoding all 50×192 forecast states. Both models use the same frozen DC-AE, making their encoding and decoding costs directly comparable. LaDCast is extended to 48 days solely to match the generated output volume; this timing comparison does not claim validated LaDCast accuracy beyond its reported 15-day horizon.

AuthorsOpenReview ↗

Official Comment by Authors

Dear Reviewer x2AV, one short addition to our response above.

W2 & Q4. In the discussion with Reviewer Wp4w we scored encode–decode reconstructions of ERA5 with the weekly TCC of Figure 2. The reconstruction keeps a precipitation TCC of 0.87, while all three forecasts stay below 0.21 from week 3. We therefore no longer attribute the weaker precipitation skill to the frozen representation, as our response above did, and describe it as a limitation of the full system.

We would be glad to answer any remaining questions before the discussion closes.

Reviewer_FXPUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Marchuk-S2S, a generative model for six-hourly subseasonal ensemble forecasting. It combines conditional flow matching in a frozen pretrained latent space with variable-horizon training and autoregressive composition to produce 48-day forecasts. The method is computationally efficient and achieves competitive results against several forecasting systems. However, some experimental comparisons and methodological claims require further clarification and validation.

Critical strengths
  1. The proposed framework efficiently generates six-hourly ensemble forecasts up to 48 days using a single model.

  2. The paper includes comparisons with numerical and AI-based forecasting systems, together with ablations on temporal resolution, positional encoding, and forecast chunk length.

Critical weaknesses
  1. LaDCast is mentioned in the abstract without brief explanation. Readers unfamiliar with this model may not understand its relationship to Marchuk-S2S or the significance of the reported 7.6× speedup.

  2. The MJO forecast skill comparison between FuXi-S2S and ECMWF S2S in Figure 3 is inconsistent with the results reported by Chen et al. (2024, Nature Communications). The original study reported skillful MJO predictions up to 36 days for FuXi-S2S and 30 days for ECMWF S2S, whereas Figure 3 shows both models losing skill at approximately 30 days. The authors use a precipitation-based RMM-like index rather than the standard OLR-based RMM index. The evaluation periods and ECMWF ensemble configurations also differ.

  3. Appendix A.8 does not sufficiently describe how the precipitation-based RMM-like index is constructed, including the EOF basis, normalization, and treatment of the seasonal cycle. Replacing OLR with precipitation changes the definition of the MJO index, and the resulting skill is not directly comparable with standard RMM-based evaluations. More methodological details and a sensitivity analysis are needed.

  4. The main advantage of Marchuk-S2S is its six-hourly subseasonal forecasting capability, but most subseasonal evaluations are performed using daily averages. Table 1 shows that six-hourly training improves daily CRPS, but it also provides four times more target frames per training example than daily training. Moreover, the daily targets differ from the distribution used to train the frozen autoencoder. The experiment therefore does not isolate the benefit of higher temporal resolution. Direct verification of six-hourly forecasts would better support this contribution.

  5. The subseasonal comparison is limited to ECMWF S2S and FuXi-S2S, although several relevant models, including FengWu-W2S and AIFS-CRPS, are discussed. More importantly, the ECMWF S2S forecasts are evaluated without bias correction, and anomalies are calculated relative to ERA5 climatology rather than model-specific reforecast climatology. These choices may influence subseasonal skill comparisons. The authors should clarify their effects and provide uncertainty estimates for the reported differences.

  6. The ablations in Table 3 mainly evaluate RMSE at a 15-day lead time, whereas the central contribution is forecasting up to 48 days. It remains unclear whether the benefits of variable-horizon training, positional encoding, and the selected four-day chunk length persist at subseasonal lead times. Additional ablations for weeks 3–6 would be more convincing.

  7. The direct-versus-composed transition analysis in Table 2 shows distributional discrepancies of 0.13–0.18, substantially above the sampling reference of approximately 0.04. Although composition reduces CRPS and maintains similar ensemble spread, the two transition distributions are not consistent. The current analysis covers only compositions of two chunks, while the 48-day forecasts require twelve consecutive four-day transitions. The implications for long-range ensemble calibration and forecast stability remain unclear.

  8. The computational efficiency comparison requires clearer interpretation. Table 7 reports 9.7 minutes for a 48-day Marchuk-S2S ensemble and 5.1 minutes for a 42-day FuXi-S2S ensemble. The distinction in temporal resolution of these two models should be clearly stated when discussing efficiency.

Questions for authors
  1. Can the authors reproduce the MJO comparison using the standard OLR-based RMM index and explain the discrepancy with Chen et al. (2024)? Which ECMWF S2S model cycle, forecast product, and ensemble configuration were used?

  2. How does Marchuk-S2S perform at six-hourly resolution in terms of forecast accuracy, temporal variability, and the representation of diurnal cycles, compared with relevant six-hourly forecasting models?

  3. Do the benefits of variable-horizon training and four-day chunk composition remain consistent over weeks 3–6? How does changing the inference chunk length affect long-range CRPS and ensemble spread?

  4. How sensitive are the subseasonal comparisons to bias correction, climatology definitions, and the selection of initialization dates? Can the authors report confidence intervals for the main comparisons?

7 comments
AuthorsOpenReview ↗

Author Responses (1/3)

We sincerely thank the reviewer for the helpful comments. We respond to each point below.

W1

We clarified this relationship in the revised abstract: “Sharing LaDCast's frozen autoencoder and six-hourly latent-space ensemble setting allows a direct comparison: Marchuk-S2S achieves comparable medium-range forecast skill with a 276M-parameter backbone, versus 1.6B parameters for LaDCast.“

W2 & W3 & Q1

We cannot reproduce the three-system comparison with the standard OLR-based RMM index. OLR is neither an input nor an output of Marchuk-S2S: the frozen LaDCast representation does not include it. FuXi-S2S and ECMWF in Figure 3 are therefore scored with the same precipitation-based index as Marchuk-S2S. The curves are not a reproduction of the OLR-based leads in [2].

The ECMWF forecast in Figure 3 is the 51-member real-time operational S2S product on the matched 2019–2021 dates: one control member and 50 perturbed members, initialized at 00 UTC on Mondays and Thursdays and integrated to day 46. The cycle on a given date is one of CY45R1, CY46R1, CY47R1, CY47R2, and CY47R3 [3]. [4] report skillful OLR-based predictions out to 36 days for FuXi-S2S and 30 days for ECMWF, on 2017–2021 data, using 11-member Cycle 47r3 reforecasts. In Figure 3 all three systems fall below COR = 0.5 at roughly 30 days. That crossing is specific to our index.

The construction is now specified in Appendix A.7. The index is a modified RMM [4], built from joint EOFs of daily U850, U200, and 24-hour total precipitation, each averaged with equal latitude weight over 15°S–15°N. The basis is computed once from ERA5 for 1990–2017, after the seasonal cycle is removed, and then held fixed. Each field is normalized by its ERA5 standard deviation, and the leading two joint EOFs are retained. Ensemble-mean forecasts and ERA5 are projected onto that same basis. Skill is the bivariate correlation of the ensemble-mean index, with COR = 0.5 as the conventional threshold.

Replacing OLR with precipitation follows the practice used for ML models that do not predict OLR [1]. Precipitation is a noisier convection proxy, so the absolute COR values in Figure 3 lie below the OLR-based correlations reported by [2]. The frozen DC-AE adds a further precipitation error before any forecast is made. On 2018 at 1.5°, its 6-hour precipitation RMSE is 0.202 mm at every lead (Table 9):

Lead timeDC-AE RMSE, mmMarchuk-S2S RMSE, mmDC-AE / Marchuk-S2S
1 day0.2020.31963%
5 days0.2020.46344%
10 days0.2020.50840%
15 days0.2020.51939%
30 days0.2020.52638%
45 days0.2020.53038%

This reconstruction error is 63% of the Marchuk-S2S precipitation RMSE at day 1 and 38% on days 30–45. The lower precipitation-based COR therefore combines the noisier convection proxy with this fixed latent error. The gap relative to [2] also includes the evaluation period, the ECMWF product and cycle, and the ensemble size, as stated in Appendix A.7. We do not present the COR = 0.5 crossing in Figure 3 as the OLR-based result.

[1] Chien, M.-T., Barnes, E. A., and Maloney, E. D. Modulation of tropical cyclogenesis on subseasonal-to-interannual timescales in the deep-learning climate emulator ACE2, 2025.

[2] Chen, L., Zhong, X., Li, H., Wu, J., Lu, B., Chen, D., Xie, S.-P., Wu, L., Chao, Q., Lin, C., Hu, Z., and Qi, Y. A machine learning model that outperforms conventional global subseasonal forecast models, 2024.

[3] ECMWF S2S model configurations. confluence.ecmwf.int/spaces/S2S/pages/47288128/ECMWF+Model

[4] Wheeler, M. C., and Hendon, H. H. An All-Season Real-Time Multivariate MJO Index: Development of an Index for Monitoring and Prediction, 2004.

AuthorsOpenReview ↗

Author Responses (2/3)

W4 & W5 & Q2

We agree that Table 2 does not isolate temporal resolution. The six-hourly model receives four times more target frames per example, and the daily run is trained on daily means, especially 24-hour precipitation totals, that lie partly outside the distribution of the frozen DC-AE, which was trained on instantaneous states.

Appendix Table 12 shows how that mismatch appears in reconstruction error. The daily-mean RMSE is lower overall, because averaging suppresses short-term variability, but it is higher for T2M (0.938 K versus 0.769 K at 6 hours) and for precipitation scaled to a 6-hour accumulation (3.9×10⁻⁴ m versus 2.02×10⁻⁴ m). These are also the two fields whose ensemble-mean RMSE falls most under six-hourly training, by 4.1% for T2M and 1.5% for precipitation. CRPS falls on all five fields, by 1.6–6.9% (Table 2). The daily scores therefore improve under this training setup. The time step is not established as the sole cause, and the autoencoder error does not favor the daily model on every field.

The native six-hourly output is scored directly in Figure 4 and Appendix Figures 7–8. Both report latitude-weighted RMSE and CRPS from days 15 to 30 on the 1.5° grid for the 2019–2021 initializations, against LaDCast-1.6B on matched initializations and ERA5 targets, and against climatology. The two models share the frozen autoencoder, the variables, and the 6-hour step, so the comparison tests forecasting and sampling. LaDCast was validated only to 15 days, and its extension to day 30 is a rollout diagnostic. These scores complement the daily evaluation and do not verify the diurnal cycle.

Matched six-hourly comparisons at this horizon are limited, so we did not add FengWu-W2S or AIFS-CRPS as controlled baselines. FengWu-W2S has no public checkpoint, and we had no matched hindcasts for our 313 dates. AIFS ENS, the operational AIFS-CRPS, has open weights, but it is fine-tuned on operational analyses for 2016–2023, so 2019–2021 are inside that window, and a 46-day ensemble at about 30 km on 313 dates was beyond our compute. Its published 46-day reforecasts are a different O96 run, trained on ERA5 for 1979–2017 and issued weekly with 8 members. AIFS-SUBS provides 10-member, 46-day, 24-hour reforecasts [2], but the period, initialization, resolution, ensemble size, and climatology do not match our protocol.

The controlled comparisons are therefore ECMWF and FuXi-S2S at daily resolution, and LaDCast-1.6B with climatology at six-hourly resolution (Figure 4, days 15–30). Many six-hourly models are scored only in the medium range, and their longer rollouts often lose skill relative to climatology [1].

[1] Nathaniel, J., Qu, Y., Nguyen, T., Yu, S., Busecke, J., Grover, A., and Gentine, P. ChaosBench: A Multi-Channel, Physics-Based Benchmark for Subseasonal-to-Seasonal Climate Prediction, 2024.

[2] European Centre for Medium-Range Weather Forecasts. AIFS-SUBS reforecasts.

W6 & Q3

We agree that these choices have to be checked at a subseasonal lead. The manuscript reports them at day 15 and day 45, which are the two leads in the ablation tables. The RMSE results are in Table 4, on 2018 initializations with 20 members. CRPS and ensemble spread for the 2-, 4-, and 8-day inference chunks are in Table 5.

The benefits are still present at day 45. Variable-horizon training is best or tied for T500, Z500, MSLP, and T2M, and second-best for 10 m zonal wind (Table 4b). Our positional encoding has the lowest RMSE on all five fields at both day 15 and day 45 (Table 4a).

With that variable-horizon checkpoint fixed, four-day inference chunks give the lowest day-45 RMSE for MSLP and T2M, tie the two-day chunk for T500, and tie the one- and two-day chunks for 10 m zonal wind. Two-day chunks are lower for Z500 (Table 4c). Four days is a strong overall choice. It is not the lowest RMSE for every variable.

Table 5 has five fields and two lead times, so the count is 10 comparisons. Four-day chunks have the lowest CRPS in 8 of these 10, and in 4 of the 5 fields at day 45. The two-day chunk is lower for Z500 at day 15 and for Z850 at day 45. Two-day chunks produce the largest spread in 9 of the 10 cases; at day 45 the T2M spread is larger for the four-day chunk. The extra spread does not make the two-day chunk best overall. Eight-day chunks are not optimal for CRPS or for spread.

This is the trade-off in Section 4.5. Short chunks require more autoregressive transitions, while long chunks have to represent a broader and more nonlinear conditional distribution. Four-day composition is the best overall balance at both reported leads, including day 45.

AuthorsOpenReview ↗

Author Responses (3/3)

W7

We agree that Table 3 does not establish exact transition consistency. The distributional discrepancy between a direct transition and a composition of two chunks is 0.13–0.18, against a sampling floor of about 0.04, so the endpoint laws differ. Section 3.4 treats this as a departure from the Chapman–Kolmogorov relation: long forecasts use the composed path and do not require exact equality. Because the two laws differ, the rollout depends on the chunk length (Tables 4c and 5).

Table 3 compares only two chunks with one chunk of twice the length, and only up to 8 days. A 48-day forecast is twelve successive four-day calls, so Table 3 does not show whether the mismatch accumulates over the full rollout.

The twelve-call forecasts are checked directly. Their spread–skill ratio stays about 0.93–0.99 from days 15 to 45 on the fields of Figure 1 (Appendix Figure 13). The ensemble is mildly underdispersive over this window, and the ratio does not collapse. RMSE and CRPS on those fields remain below climatology through day 45 (Figure 1). In the inference-chunk ablation, four-day chunks have the lowest CRPS in 4 of the 5 fields at day 45; the two-day chunk is lower for Z850 (Table 5).

Within Table 3 itself, the two-chunk compositions keep a similar ensemble scale, with spread ratios 0.91–0.94, and lower CRPS by 3–9%, with no consistency loss. That result and the twelve-call diagnostics are both finite-horizon evidence. They do not establish long-run calibration or climate fidelity, which the limitations state explicitly. An explicit transition-consistency loss and rollout-based calibration remain important future directions.

W8

We clarified the distinction in the revision. The comparison is in Section 3.6 and Table 1. Both timings are for a 50-member ensemble on one H100. Marchuk-S2S takes 9.7 minutes for a 48-day forecast and FuXi-S2S takes 5.1 minutes for a 42-day forecast. For Marchuk-S2S the time includes encoding the initial state and decoding all outputs.

FuXi-S2S requires less total time because it produces 42 daily states, whereas Marchuk-S2S produces 192 six-hourly states, four per day. Table 1 therefore also reports time per forecast day and time per output state. FuXi-S2S is 1.7× faster per forecast day (7.3 s versus 12.1 s). Marchuk-S2S is 2.4× faster per output state (3.0 s versus 7.3 s). Section 3.6 states that the total runtimes are not directly comparable for this reason. The matched six-hourly comparison in the same table is LaDCast, which produces the same 192 states.

Q4

We agree that these choices affect absolute subseasonal skill. Section 4.1 applies no bias correction, and Section 4.2 removes one ERA5 day-of-year climatology, computed over 1990–2017, from every system. All three models are scored on the same 313 twice-weekly ECMWF initialization dates from 2019–2021 and on the same ERA5 targets. Anomalies are taken from this observed climatology, so a model’s own systematic error remains in its anomaly field. The limitations state this explicitly.

Appendix A.10, “Sensitivity to model drift and to the ERA5 climatology,” measures both changes on the 2019 and 2020 starts. Table 13 replaces the 1990–2017 ERA5 climatology of the forecast with each model’s own cross-year climatology, and Table 14 changes only the ERA5 base period, from 1980–2007 to the 1990–2017 period used in the main evaluation. Both changes lower weeks 3–6 TCC. The drop is shared by Marchuk-S2S, FuXi-S2S, and ECMWF, and it is larger for Z500, T850, and T2M than for U850 and precipitation. Absolute late-lead TCC is therefore sensitive to the climatology, and this sensitivity is not specific to Marchuk-S2S.

Appendix A.8 Tables 10 and 11 give paired 95% moving-block bootstrap intervals by season and initialization year, from 1,000 resamples. Every difference whose interval excludes zero favors Marchuk-S2S, but many do not: Z500 and Z850 RMSE are indistinguishable from zero in every season and year, and the significant RMSE reductions are limited to T850 and T2M in part of the sample. Fair CRPS is significant more often, especially in spring, summer, and across 2019–2021, and less often in winter and autumn. We therefore describe Marchuk-S2S as competitive.

Uncertainty for the pooled comparison is reported with those tables. In Section 4.1, Figure 1 shows bootstrap standard errors from 1,000 resamples of the 313 dates. In Section 4.2, Figure 2 marks in pale the lead times at which Marchuk-S2S does not improve on ECMWF at the 97.5% level, again from a bootstrap repeated 1,000 times. The three-year sample remains an explicit limitation.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response to All Reviewers

Dear Reviewers,

We sincerely thank all the reviewers (FXPU, x2AV, Wp4w) for their valuable feedback. We are glad that the reviewers appreciated the efficient generation of six-hourly ensemble forecasts up to 48 days with a single model (FXPU, x2AV, Wp4w), the practical computational efficiency of a compact model that remains competitive with ECMWF S2S and FuXi-S2S (x2AV, Wp4w), the comparisons and analyses of temporal resolution, positional encoding, forecast chunk length, rollout consistency, and ensemble calibration (FXPU, x2AV), and the consistent improvement in daily CRPS from training on six-hourly states rather than on daily means (x2AV, Wp4w).

We have made every effort to faithfully address your comments in the individual responses. As suggested by the reviewers, we add:

  • Statistical uncertainty and significance analysis: Figure 1, Figure 2, and Appendix A.8 (Tables 10–11). Requested by Reviewers FXPU and x2AV.
  • Sensitivity to model drift and to the ERA5 climatology: Appendix A.10 (Tables 13–14). Raised by Reviewers FXPU and x2AV.
  • Dense six-hourly subseasonal evaluation, and what Table 2 does not isolate: Section 4.6 (Figure 4), Appendix A.11 (Figures 7–8), and Appendix A.9 (Table 12). Requested by Reviewer FXPU, also relevant to Reviewer x2AV.
  • Ablations at day 15 and day 45: Table 4 in Section 4.5. Requested by Reviewer FXPU.
  • Effect of the inference chunk horizon on CRPS and ensemble spread: Table 5 in Section 4.5. Raised by Reviewers FXPU and Wp4w.
  • Clarified inference cost comparison: Table 1 and Section 3.6. Requested by Reviewers FXPU and x2AV.
  • Clarified relationship to LaDCast: the abstract. Raised by Reviewer FXPU.
  • MJO evaluation and precipitation skill: Appendix A.7 and Table 9. Raised by Reviewers FXPU, x2AV, and Wp4w.

Marchuk-S2S is described as competitive with ECMWF and FuXi-S2S, and the remaining limitations are stated explicitly.

We have incorporated the suggested modifications in the revised version, which are highlighted in red. Added tables and figures have shifted the numbering relative to the original submission; the references above follow the revised manuscript. If you are satisfied with the revisions, we kindly request your consideration of an improved score.

Thanks again to all the reviewers for their time.

Best regards,

Authors

AuthorsOpenReview ↗

Summary of the Discussion

We thank the reviewers for the constructive discussion. For the Area Chair, a brief summary:

Reviewer Wp4w raised the score from 2 to 3. A new reconstruction analysis (encode–decode ERA5 keeps a weekly precipitation TCC of 0.87, against forecasts below 0.21 from week 3) led us to withdraw our earlier attribution of the precipitation and late-lead MJO gaps to the frozen autoencoder; we now describe the shortfall as a limitation of the full system. We also clarified how variable-horizon training differs from LaDCast, Stormer and Continuous Ensemble Forecasting. The reviewer noted clear practical and efficiency benefits; the remaining concern is methodological novelty.

Reviewer x2AV (score 3) noted that the added experiments and statistical analysis addressed most concerns. The remaining point is that differences in evaluation settings prevent claiming an advantage over systems such as FengWu-W2S and AIFS-SUBS (which lack public checkpoints or overlapping out-of-sample periods); our comparative claims are therefore limited to ECMWF S2S and FuXi-S2S on the same 313 initialization dates (Figure 2, Tables 10–11), and to LaDCast-1.6B with the same autoencoder, 1.5° grid, and 6-hour step through day 30 (Figure 4).

Reviewer FXPU (score 2) followed-up only on the precipitation-based MJO index. OLR is not represented in the frozen autoencoder, which we reuse to keep the comparison with LaDCast controlled, and we will describe the result as skill on a precipitation-based RMM-like index. The other points of the review are covered by additions to the revision: bootstrap uncertainty and paired significance tests (Figures 1–2, Tables 10–11), sensitivity to drift and climatology (Appendix A.10), ablations at day 45 (Tables 4–5), direct six-hourly verification against LaDCast (Figure 4), and the cost comparison (Table 1).

Beyond the additions already in the revised PDF, the final version will include: the TCC reconstruction table and revised precipitation discussion; the precipitation-based RMM-like index wording in the abstract, contributions list, and Conclusion; OLR and the scope of S2S model comparisons in the Limitations; Section 4.5 text aligned with the day-45 numbers of Table 4; and a citation to Continuous Ensemble Forecasting (Andrae et al., 2025).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.