acceptodds
Under review as a conference paper at ICLR 2027

HurdleCast: Hurdle-Gated Mixture of Experts for Extreme Precipitation Nowcasting

Abstract

Recent AI weather forecasting models have achieved substantial gains in forecast skill across lead times ranging from hours to weeks. Nonetheless, record-breaking extremes remain a persistent challenge, with far less consistent improvements in predictive performance. This limitation is especially consequential for precipitation nowcasting, where short-lead forecasts support severe-weather warning, flash-flood management, and dam operations, and extremes pose the greatest operational risk. Even modern generative nowcasting models achieve strong forecast quality on light-to-moderate precipitation but struggle to predict extremes. High-intensity extremes are sparsely represented in training data, while AI models typically optimize an average loss across the full precipitation range. We hypothesize that modeling this distribution through a single shared set of generative dynamics limits the representation of rare, high-intensity events. We introduce HurdleCast, a flow-matching generative model that explicitly decomposes precipitation prediction into occurrence and conditional severity through a supervised mixture-of-experts velocity field. Two probabilistic gates route latent flow-matching dynamics among experts specialized for precipitation intensity attributes, implemented here as dry, moderate, and heavy precipitation. On SEVIR and MeteoNet, HurdleCast improves over state-of-the-art AI nowcasting models at extreme precipitation intensities, increasing Critical Success Index (CSI) at the highest intensity thresholds by 11–30% and 9–31%, respectively.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
precipitation nowcasting, extreme weather, mixture of experts, conditional flow matching, hurdle model, spatiotemporal forecasting

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_LZUTOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The authors introduce HurdleCast, a generative precipitation nowcasting model combining latent flow matching with a supervised mixture-of-experts (MoE) for dry, moderate, and heavy precipitation regimes. Experiments on SEVIR and MeteoNet demonstrate improved performance, particularly for extreme precipitation events.

Critical strengths
  1. Clear presentation and thorough evaluation. The paper is well written, clearly motivated, and easy to follow. The experimental evaluation is comprehensive, with competitive baselines and informative ablation studies.
  2. Simple, effective contribution. The supervised hurdle-gated MoE is intuitive and well-motivated. The ablations convincingly support its effectiveness, especially for extreme precipitation.
Critical weaknesses
  1. Limited forecasting horizon and high computational cost. Evaluation is restricted to 60-minute forecasts, while training takes approximately 11 days on two B200 GPUs per dataset. This raises questions about efficiency and applicability to longer horizons, especially when compared to other recent work like DiffCast or PercpCast.
  2. Limited probabilistic and qualitative evaluation. Beyond CRPS, ensemble calibration and diversity receive little attention. The visual improvements in Figures 4 and 5 are also difficult to appreciate, and the local SSIM maps are not particularly informative.
Questions for authors
  1. Could the authors compare against PercpCast and DiffCast, including training efficiency? PercpCast (3-hour forecasting) reportedly trains in 16 hours on an A100, and DiffCast (100-minutes forecasting) in under a day on an A6000. They both achieve SOTA performance in SEVIR and have public code.
  2. Could the authors provide additional ensemble calibration and diversity analysis, such as spread-skill plots and rank histograms?
  3. Could the authors clarify the local SSIM maps, report LPIPS in the main comparisons, and explain the white regions in exPreCast’s prediction in Figure 5?
  4. Could the authors consider an anonymized code release?
5 comments
AuthorsOpenReview ↗

W1/Q1: Forecasting Horizon, Baselines, and Computational Cost

We thank the reviewer for their positive assessment of the paper's presentation, comprehensive evaluation, and the effectiveness of our supervised hurdle-gated MoE. We appreciate the suggestions concerning forecasting horizons, efficiency, probabilistic evaluation, and reproducibility, and address them in separate responses.

Benchmark protocol and forecasting horizon. Our evaluation compares HurdleCast against eight baselines: six ML-based approaches (U-Net, EarthFormer, SimVP, exPreCast, CasCast, and FlowCast) and two conventional nowcasting approaches (Persistence and pySTEPS). All are evaluated under the same protocol at SEVIR's native 384×384 resolution and a 60-minute forecasting horizon (Table 1), consistent with the benchmark setting established in SEVIR's introductory paper [1].

PercpCast and DiffCast evaluate different forecasting configurations:

PropertyHurdleCastPercpCastDiffCast
Input context13 frames13 frames5 frames
Horizon60 min180 min100 min
Resolution384×384128×128128×128
EvaluationNative SEVIRModified SEVIRModified SEVIR

These differences in forecasting horizons, spatial resolutions, and evaluation protocols prevent direct comparison of the published benchmark results.

Reproducibility of additional baselines. We examined both publicly released implementations and identified practical obstacles to reproducing them under our evaluation protocol:

  1. DiffCast: Although the paper describes joint training of the diffusion component and deterministic predictor (Section 4.4), the released implementation raises NotImplementedError when training is invoked, stating that the training process is unsupported. The available code supports inference with the pretrained 128×128 checkpoint using 5 input and 20 output frames, but not retraining under our configuration.

  2. PercpCast: Its repository includes a configuration file matching our resolution and frame counts, but our inspection indicates that this configuration is not used by the executable pipeline. The released pipeline instead uses 128×128 resolution and a 36-frame, 180-minute forecasting output.

Consequently, reproducing either method under the native SEVIR protocol requires substantial implementation changes beyond adjusting configuration parameters. We therefore prioritize comparisons with eight established baselines under identical evaluation conditions. We recognize that longer-horizon forecasting remains an important direction that our current experiments do not assess.

Training and inference efficiency. For completeness, we summarize the reported training costs of the closest generative baselines alongside our measured inference latency.

ModelTraining costHardware
HurdleCastApproximately 11 days (528 GPU-hours)2 NVIDIA B200
FlowCastApproximately 7 days (672 GPU-hours)4 NVIDIA H100
CasCast18 hours (72 GPU-hours), diffusion stage only4 NVIDIA A100

CasCast's reported training time covers only its diffusion stage, excluding the separately trained autoencoder and EarthFormer compression model. As these approaches use different hardware and training configurations, the figures provide context rather than a hardware-normalized comparison.

We additionally measured inference latency for an eight-member ensemble at batch size one under identical conditions:

ModelLatency per input (8-member ensemble)
HurdleCast11.30 s
FlowCast11.28 s
CasCast5.83 s

HurdleCast introduces a negligible inference overhead of 0.02 s per input (approximately 0.18%) relative to FlowCast, while improving extreme-threshold forecasting skill (Table 4). CasCast achieves lower inference latency, illustrating the trade-off between computational efficiency and extreme-event forecasting performance across these approaches.

We will incorporate these comparisons and clarify the computational considerations in the revised manuscript.

AuthorsOpenReview ↗

Q2: Ensemble Calibration and Reliability

We thank the reviewer for suggesting additional probabilistic diagnostics. We have conducted the requested reliability analysis on the full SEVIR test set, using eight-member forecast ensembles.

Rank histogram. We computed a rank histogram with random jitter to handle ties arising from VIL's substantial zero inflation. The resulting histogram is U-shaped, indicating ensemble underdispersion. In particular, the lowest rank bin contains approximately 27.5% of observations, compared with 11.1% expected across nine equally likely rank bins for an eight-member ensemble.

Exceedance-probability reliability. We additionally evaluated predicted exceedance probabilities against observed event frequencies at the two highest precipitation thresholds, VIL ≥ 181 and VIL ≥ 219.

Predicted probabilityObserved frequency (VIL ≥ 181)Observed frequency (VIL ≥ 219)
0.0000.0010.000
0.1250.1050.079
0.2500.1770.136
0.3750.2390.189
0.5000.3010.245
0.6250.3680.307
0.7500.4470.384
0.8750.5530.492
1.0000.7690.717

The observed frequencies increase monotonically with predicted exceedance probabilities at both thresholds. However, predicted probabilities generally exceed observed frequencies, indicating overforecasting of extreme-event probabilities.

Empirical ensemble CRPS. To complement these calibration diagnostics, we also computed CRPS directly from the raw ensemble members, without assuming a parametric predictive distribution, using the bias-corrected estimator [4]. On SEVIR, HurdleCast's empirical CRPS is 6–7% lower than its Gaussian-fitted CRPS across lead times. It also improves upon FlowCast's empirical CRPS by 0.00046, with a 95% paired bootstrap confidence interval for HurdleCast minus FlowCast of [−0.00048, −0.00045].

These findings distinguish improvements in overall probabilistic accuracy from the remaining calibration limitations. The empirical CRPS results support HurdleCast's probabilistic advantage, while the rank histogram and reliability diagrams identify underdispersion and overforecasting of extreme-event probabilities. Ensemble size influences the interpretation of probabilistic verification diagnostics [2], and underdispersion is a recognized challenge in precipitation ensemble prediction [3].

We acknowledge these limitations and will include the additional diagnostics and their interpretation in the appendix.

AuthorsOpenReview ↗

Q3: LPIPS and Qualitative Evaluation

We thank the reviewer for requesting additional perceptual evaluation and clarification of the qualitative figures.

LPIPS comparisons. HurdleCast's LPIPS is already reported in Appendix A.13 through the ensemble-size analysis (LPIPS = 0.2744 for eight ensemble members on SEVIR). We provide the comparative results below and will incorporate them into the main experimental comparisons.

SEVIR LPIPS (lower is better):

ModelLPIPS
U-Net0.4010
EarthFormer0.3946
SimVP0.3622
exPreCast (deterministic)0.1964
CasCast (N=8)0.2811
FlowCast (N=8)0.2767
HurdleCast (N=1)0.2349
HurdleCast (N=8)0.2744

MeteoNet LPIPS (lower is better):

ModelLPIPS
U-Net0.3102
EarthFormer0.3081
SimVP0.2739
exPreCast (deterministic)0.1824
FlowCast (N=8)0.2510
HurdleCast (N=1)0.2164
HurdleCast (N=8)0.2458

The three deterministic regression baselines (U-Net, EarthFormer, and SimVP) exhibit higher LPIPS than the generative methods on both datasets, consistent with the smoothing often associated with pixel-wise reconstruction objectives.

Among the remaining methods, exPreCast achieves the lowest LPIPS. Ensemble averaging also affects perceptual similarity: HurdleCast's single-member prediction (N=1) achieves substantially lower LPIPS than its eight-member ensemble mean, consistent with the smoothing introduced by averaging generated forecasts.

For eight-member ensembles, HurdleCast achieves LPIPS of 0.2744 on SEVIR, compared with 0.2767 for FlowCast and 0.2811 for CasCast. On MeteoNet, HurdleCast achieves 0.2458, compared with 0.2510 for FlowCast. These results place HurdleCast among the stronger probabilistic baselines in perceptual similarity, while its principal advantage remains extreme-event forecasting skill.

Local SSIM maps (Figures 4 and 5). The local SSIM maps identify spatial regions where forecast structure differs from the ground truth. Two forecasts may appear visually similar at the full-image scale while exhibiting meaningful local differences in storm boundaries, spatial texture, and precipitation-core placement. The maps are intended to highlight these discrepancies rather than serve as an additional aggregate performance metric.

We agree that their interpretation could be clearer and will expand the figure captions to explain the local structural differences they illustrate.

White regions in exPreCast's Figure 5. These regions correspond to negative dBZ values. Because dBZ is logarithmic, negative reflectivity values can represent very weak radar returns. The white regions arise because these values fall below the colormap's 0 dBZ display floor.

Unlike the other models, exPreCast's output was not clipped to the 0 dBZ display floor to remain faithful to its released implementation.

Reviewer_2K7MOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper leverages the ideas common in econometrics, based on hurdle models, to model different rates of precipitation events. To this end, the authors formulate a model which uses a proposed hurdle-gated Mixtures-of-Experts (MoE), and leverages flow matching.

Critical strengths

The paper offers a sound approach to precipitation forecasting, rooted in a familiar idea of improving the model by using different experts to model the precipitation severity.

The paper links the problem of precipitation forecasting to hurdle models, common in econometrics, but not regularly linked to precipitation forecasting.

Critical weaknesses
  1. The paper does not sufficiently engage with the existing literature on extreme precipitation forecasting, which leverage the ideas of separation of precipitation levels, and no precipitation, as well as directly use MoEs. These works include:

Jiang et al (2025) Knowledge-guided adaptive mixture of experts for precipitation prediction (directly introduces MoEs)

Xiao et al (2026) PA-Net: Precipitation-Adaptive Mixture-of-Experts for Long-Tail Rainfall Nowcasting

Most importantly, the methods the authors compare with are to my knowledge not like-for-like models using different levels of precipitation forecasting. The authors need to explain why this comparison shows the full picture and compares with the methods which use different expert models suitable for different precipitation levels.

  1. Crucially, the model does not show the link between the hurdle models and models, common in machine learning for precipitation nowcasting, which directly separate the rainfall level event prediction from severity. Addressing the aspect why this is not mathematically the same is not sufficient either.
Questions for authors

Addressing the questions above is crucial for confirming the claims-and-evidence aspect of this work.

3 comments
AuthorsOpenReview ↗

W1/Q1: Related Literature and Baseline Comparability

We thank the reviewer for their positive assessment of the paper's soundness and for recognizing the value of connecting precipitation forecasting with hurdle models. We appreciate the suggested references and address the literature comparison and architectural distinction in two separate responses.

Jiang et al. [1]. This work employs 16 MLP experts to partition input feature categories for single-site precipitation regression. It does not incorporate spatial convolution or attention, and its evaluation uses a nine-day Hurricane Ian window with training and testing samples drawn from the same event. Its specialization axis is therefore input features or modalities, rather than the occurrence/severity decomposition we propose. The single-event evaluation also does not directly assess generalization across independent storms.

Xiao et al. [2] (PA-Net). This work has greater thematic overlap with HurdleCast, employing a spatiotemporal architecture with precipitation-adaptive routing. However, PA-Net uses sparse top-k routing to allocate computational capacity according to rainfall intensity, whereas HurdleCast combines fixed regime-specialist velocity experts through supervised occurrence and conditional-severity probabilities at each flow-matching step.

PA-Net is also evaluated on ERA5 reanalysis rather than direct radar observations. Precipitation in ERA5 is a model-derived diagnostic quantity rather than a directly assimilated observational field, introducing important differences when evaluating extreme precipitation [3]. HurdleCast is evaluated on radar-based SEVIR and MeteoNet benchmarks.

Comparison of specialization mechanisms and evaluation settings:

PropertyJiang et al. [1]PA-Net [2]HurdleCast
Architecture16 MLP expertsTransformer MoEHurdle-gated velocity experts
SpecializationInput featuresSparse top-k compute allocationSupervised occurrence/severity
DataSingle-site precipitation featuresERA5 reanalysisSEVIR/MeteoNet radar
TaskSingle-site regressionSpatiotemporal forecastingProbabilistic short-term nowcasting
GenerationDeterministicDeterministicConditional flow matching
EvaluationOne 9-day event, within-event split2016–18/2019/2020 splitHeld-out benchmark splits and extreme-event cases

These approaches share the motivation of modeling heterogeneous precipitation behavior but differ in their specialization mechanisms, predictive tasks, and evaluation settings. Jiang et al.'s within-event split does not directly assess cross-event generalization, while PA-Net's chronological ERA5 evaluation addresses a different data setting.

Baseline selection. To our knowledge, neither method provides publicly available implementation code. Reimplementing these architectures within the discussion period would risk introducing uncontrolled differences in training and optimization. Our empirical comparison therefore focuses on established deterministic and generative nowcasting baselines evaluated under common benchmark conditions.

We will incorporate both studies into Related Work, explicitly distinguish their specialization mechanisms from ours, and clarify the scope of our empirical comparisons.

AuthorsOpenReview ↗

W2/Q2: Relationship to Existing Hurdle and Precipitation Models

We thank the reviewer for highlighting the importance of connecting our formulation to existing machine-learning approaches that separate precipitation occurrence from intensity. We fully agree that the contribution must be established through its architectural and empirical implications, rather than mathematical distinctions alone. Accordingly, we have explicitly developed these connections throughout the manuscript and substantiated our architectural contribution through controlled ablations and empirical analyses.

Connections to existing literature. This relationship is developed across the manuscript. The Introduction (Section 1) motivates the occurrence/severity decomposition through its econometric origins (Cragg, 1971; Mullahy, 1986) and hydrological applications (Richardson, 1981; Stern and Coe, 1984). Related Work (Section 2.3) discusses existing ML approaches that separate precipitation occurrence from severity, including Deep Extreme Mixture Models (DEMM) [4], which explicitly model zero-inflated and heavy-tailed spatiotemporal outcomes.

Architectural distinction. We do not claim novelty for hurdle modeling or mixture-of-experts individually. Our contribution lies in integrating supervised occurrence/severity specialization directly into the generative flow-matching dynamics.

As described in Section 3.3 (Eq. 4), the hurdle factorization

p(y)=(1−π)1[y=0]+πf(y∣y>0)p(y)=(1-\pi)\mathbf{1}[y=0]+\pi f(y\mid y>0)

motivates two supervised gates: an occurrence gate distinguishing dry from non-dry conditions, and a conditional-severity gate distinguishing moderate from heavy precipitation. These gates combine regime-specific velocity experts at every flow-matching integration step.

This differs from approaches such as DEMM, which parameterize the predictive output distribution using separate precipitation regimes. In HurdleCast, the decomposition instead governs the evolution of the generative velocity field itself.

Controlled empirical evidence. Section 4.5 isolates the effect of this architectural choice through two ablations.

First, against Vanilla-CFM, a single-velocity-head control matched on backbone, autoencoder, and training budget (Table 3), HurdleCast improves CSI-219-P16 by 153% (0.1116 to 0.2824) and CSI-181-P16 by 32%.

Second, against a capacity-matched unsupervised three-expert MoE with identical architecture but without occurrence/severity supervision (Table 4):

ModelCSI-MCSI-219
Unsupervised 3-expert MoE0.45120.1346
HurdleCast0.47000.2072
Relative improvement+4.2%+53.9%

These results demonstrate that the gains cannot be explained solely by increased expert count or model capacity. The supervised hurdle decomposition provides an additional benefit, particularly at the extreme threshold, where CSI-219 improves by approximately 54% over the unsupervised MoE.

Together with the routing and frequency-bias analyses, these controlled experiments support our central contribution: using physically interpretable occurrence/severity supervision to guide expert specialization within the generative dynamics, rather than exclusively at the predictive output.

We will strengthen the discussion of this relationship in the Introduction, Related Work, and Methods.

AuthorsOpenReview ↗

References for our responses W1–W2

[1] Jiang, C., Osei, K., Yeddula, S. D., Feng, D., and Ku, W.-S., 2025: Knowledge-Guided Adaptive Mixture of Experts for Precipitation Prediction. arXiv:2509.11459.

[2] Xiao, X., Lei, S., Liu, E., Xiang, S., Li, H., Yuan, C., Qi, Y., and Jin, Q., 2026: PA-Net: Precipitation-Adaptive Mixture-of-Experts for Long-Tail Rainfall Nowcasting. arXiv:2603.13818.

[3] Rasp, S., Hoyer, S., Merose, A., Langmore, I., Battaglia, P., Russell, T., Sanchez-Gonzalez, A., Yang, V., Carver, R., Agrawal, S., Chantry, M., Bouallègue, Z. B., Dueben, P., Bromberg, C., Sisk, J., Barrington, L., Bell, A., and Sha, F., 2024: WeatherBench 2: A benchmark for the next generation of data-driven global weather models. Journal of Advances in Modeling Earth Systems, 16(6), e2023MS004019. https://doi.org/10.1029/2023MS004019

[4] Wilson, T., McDonald, A., Galib, A. H., Tan, P.-N., and Luo, L., 2022: Beyond Point Prediction: Capturing Zero-Inflated & Heavy-Tailed Spatiotemporal Data with Deep Extreme Mixture Models. KDD 2022, 2020–2028.

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

This paper proposes HurdleCast, a novel generative framework for extreme precipitation nowcasting that integrates a supervised hurdle-gated mixture-of-experts (MoE) mechanism into conditional flow matching. To address the zero-inflated and heavy-tailed distribution of precipitation, HurdleCast employs two hierarchical probabilistic gates to model precipitation occurrence and conditional severity, dynamically combining three specialized velocity experts for dry, moderate, and heavy precipitation regimes. The framework further incorporates an intensity-aware autoencoder reconstruction objective and supervised gating losses to improve the representation of rare, high-intensity events. Extensive experiments on the SEVIR and MeteoNet datasets demonstrate that HurdleCast consistently outperforms competitive deterministic and generative baselines, particularly at extreme precipitation thresholds. Ablation studies further validate the effectiveness of supervised expert specialization in improving extreme precipitation forecasting performance.

Critical strengths
  1. The paper addresses the challenging problem of extreme precipitation nowcasting by explicitly considering the zero-inflated and heavy-tailed characteristics of precipitation data. The proposed hurdle-based formulation provides a reasonable foundation for modeling different precipitation intensity regimes.
  2. The proposed hurdle-gated MoE mechanism integrates precipitation occurrence and severity information into the flow-matching velocity field. The ablation studies demonstrate that supervised gating effectively improves extreme precipitation prediction beyond simply increasing model capacity.
  3. Extensive experiments on SEVIR and MeteoNet demonstrate consistent improvements over competitive baselines, particularly at high precipitation intensity thresholds. The additional ablation studies and extreme-event analyses further support the effectiveness of the proposed approach.
Critical weaknesses
  1. The proposed gating mechanism relies on predefined precipitation intensity thresholds to distinguish different regimes. However, the paper lacks a systematic analysis of how these threshold settings affect model performance, leaving their robustness insufficiently validated.
  2. HurdleCast requires substantial computational resources for training and ensemble-based inference. The paper lacks a detailed efficiency comparison with competing methods, making it difficult to assess the practical cost-effectiveness of the proposed framework.
  3. Although HurdleCast improves extreme precipitation detection, it exhibits higher MSE than the single-head baseline and does not consistently achieve the lowest false-alarm ratio. These trade-offs raise concerns about its overall prediction reliability in practical applications.
Questions for authors

Please responds to my concerns in the weaknesses.

4 comments
AuthorsOpenReview ↗

W1: Precipitation Threshold Selection and Sensitivity

We thank the reviewer for their thoughtful feedback and positive assessment of HurdleCast's motivation, architecture, experiments, and ablations. We address the three concerns in separate comments.

Threshold selection. The thresholds are domain-informed modeling choices, not parameters tuned through a hyperparameter search. For SEVIR, we use r=16r=16 and τ=133\tau=133 VIL, informed by the intensity conventions established in the SEVIR dataset paper [1]. As reported in Appendix A.10 (Table 11), these correspond to the 74.59th and 96.98th percentiles of test-set pixel intensities. For MeteoNet, we use r=3.5r=3.5 and τ=28\tau=28 dBZ as the hurdle thresholds. The latter corresponds to the 98.76th percentile of test-set pixel intensities (Appendix A.10, Table 12). Unlike SEVIR, MeteoNet does not provide an equivalent benchmark-defining threshold convention. We therefore selected the hurdle thresholds based on the empirical precipitation-intensity distribution of MeteoNet, while keeping data preprocessing and spatial configurations identical across all evaluated methods.

Learned routing. These thresholds supervise the gates during training; they do not enforce hard expert assignments at inference. Figure 10 illustrates that learned routing weights vary smoothly with precipitation intensity, allowing gradual transitions among the dry, moderate, and heavy experts.

Sensitivity analysis. We agree that a systematic threshold sweep would strengthen the robustness assessment. Changing rr and τ\tau changes the gate-supervision targets (Appendix A.4) and thus requires retraining for each configuration. With approximately 11 days of training per run, this is infeasible during the discussion period. We will clarify the threshold-selection rationale and the distinction between fixed supervision thresholds and soft, learned inference-time routing in the revised manuscript.

AuthorsOpenReview ↗

W2: Computational Efficiency and Inference Cost

Training cost. HurdleCast requires approximately 11 days on two NVIDIA B200 GPUs per dataset (528 GPU-hours). This is a one-time training cost, as the resulting checkpoint can be reused for subsequent forecasts. For context, FlowCast reports approximately seven days on four NVIDIA H100 GPUs (672 GPU-hours). CasCast reports 18 hours on four NVIDIA A100 GPUs (72 GPU-hours) for its diffusion stage alone (its separately trained autoencoder and EarthFormer compression model are not included in that figure).

Inference efficiency. We additionally measured inference latency for an eight-member ensemble at batch size one under identical measurement conditions:

ModelLatency per input (8-member ensemble)
HurdleCast11.30 s
FlowCast11.28 s
CasCast5.83 s

HurdleCast introduces a negligible inference overhead of 0.02 s per input (approximately 0.18%) relative to FlowCast, while improving extreme-threshold forecasting skill (Table 4). CasCast achieves lower inference latency, highlighting the trade-off between computational efficiency and extreme-event forecasting performance across these approaches.

Deterministic baselines (U-Net, SimVP, and EarthFormer) produce a single forecast per input rather than an eight-member ensemble. Their inference costs are therefore not directly equivalent to those of ensemble-based probabilistic methods. For operational deployment, the relevant consideration is the balance between inference latency, forecast uncertainty quantification, and extreme-event detection skill.

We will incorporate these efficiency comparisons and clarify the computational costs in the revised manuscript.

AuthorsOpenReview ↗

W3: MSE, False Alarms, and Extreme-Event Detection

MSE and false-alarm trade-offs. We agree that the extreme-event detection gains come with trade-offs in average MSE and, relative to some baselines, false-alarm ratio (FAR). Because intense precipitation is rare, optimizing overall MSE can favor conservative forecasts that underpredict extremes.

Section 4.5 (following Table 3) documents this directly: the single-head Vanilla-CFM has lower MSE (185.99 vs. 232.88 for HurdleCast), but much lower CSI-219-P16 (0.1116 vs. 0.2824). Its frequency bias drops to 0.19 at the highest threshold, whereas HurdleCast remains within 0.94–1.04 across thresholds (Table 7, Appendix A.3). This suggests that Vanilla-CFM's lower aggregate MSE accompanies substantial underforecasting of the rarest events.

FAR-M in context. In Table 1, HurdleCast's FAR-M is 0.330, versus EarthFormer 0.285, SimVP 0.298, U-Net 0.308, FlowCast 0.325, CasCast 0.383, exPreCast 0.422, pySTEPS 0.498, and Persistence 0.623. The three deterministic models with lower FAR-M also have lower CSI-219: 0.145 (EarthFormer), 0.136 (SimVP), and 0.086 (U-Net), compared with 0.206 for HurdleCast (a 29–58% relative shortfall). Thus, HurdleCast does not universally minimize false alarms. It improves detection of extreme events while accepting some additional false alarms.

Operational relevance. The relative costs of misses and false alarms are application-dependent. Bouttier and Marchal [2] discuss decision thresholds for high-precipitation warnings, including practices that tolerate more false alarms than misses in high-impact settings. Flood-warning studies [3,4] likewise motivate evaluating missed events alongside FAR. Our introduction also cites Das et al. [8] on underdetection of extremes in operationally relevant flood and hydropower forecasting. The broader impacts of extreme rainfall and flooding are documented in [5–7]. These studies motivate our emphasis on detecting high-impact extremes.

We will clarify this trade-off in the revised manuscript: HurdleCast improves extreme-event CSI and mitigates severe underforecasting at high thresholds, without claiming to minimize overall MSE or FAR.

AuthorsOpenReview ↗

References for our responses W1-W3

[1] Veillette, M., Samsi, S., and Mattioli, C., 2020: SEVIR: A Storm Event Imagery Dataset for Deep Learning Applications in Radar and Satellite Meteorology. NeurIPS 2020.

[2] Bouttier, F., and Marchal, H., 2024: Probabilistic short-range forecasts of high-precipitation events: optimal decision thresholds and predictability limits. Natural Hazards and Earth System Sciences, 24, 2793–2816. https://doi.org/10.5194/nhess-24-2793-2024

[3] Imhoff, R. O., Brauer, C. C., van Heeringen, K. J., Uijlenhoet, R., and Weerts, A. H., 2022: Large-sample evaluation of radar rainfall nowcasting for flood early warning. Water Resources Research, 58, e2021WR031591.

[4] Busker, T., van den Hurk, B., de Moel, H., and Aerts, J. C. J. H., 2025: The Value of Precipitation Forecasts to Anticipate Floods. Bulletin of the American Meteorological Society, 106(3), E473–E491.

[5] Fowler, H. J., et al., 2021: Anthropogenic intensification of short-duration rainfall extremes. Nature Reviews Earth & Environment, 2, 107–122.

[6] Merz, B., et al., 2021: Causes, impacts and patterns of disastrous river floods. Nature Reviews Earth & Environment, 2, 592–609.

[7] Kotz, M., Levermann, A., and Wenz, L., 2022: The effect of rainfall changes on economic production. Nature, 601, 223–227.

[8] Das, P., Posch, A., Barber, N., Hicks, M., Duffy, K., Vandal, T., Singh, D., van Werkhoven, K., and Ganguly, A. R., 2024: Hybrid physics-AI outperforms numerical weather prediction for extreme precipitation nowcasting. npj Climate and Atmospheric Science, 7, 35.

Reviewer_KXMROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

HurdleCast replaces the single velocity head in a latent flow-matching nowcaster with three experts for dry, moderate, and heavy precipitation. Supervised occurrence and severity gates combine their outputs at each integration step. An intensity-weighted autoencoder preserves high-intensity structures. Experiments on SEVIR and MeteoNet report improved extreme-threshold forecast skill and examine the contributions of routing and supervision.

Critical strengths
  1. The method addresses a clear weakness of nowcasting models with a small architectural change. The hurdle decomposition gives the routing mechanism an interpretable role, and the additional heads contribute fewer than 0.1% of the backbone parameters. The gains at the highest thresholds on both datasets make the approach practically relevant.
  2. The main ablations are well chosen. The single-head comparison holds the backbone, autoencoder, magnitude-weighted objective, and training budget fixed. The capacity-matched unsupervised mixture then tests the value of hurdle supervision. Together with the routing and frequency-bias diagnostics, these results provide useful evidence that the improvement involves learned regime specialization, rather than simply more parameters.
Critical weaknesses
  1. The probabilistic evaluation does not directly assess the distribution the model is designed to capture. Appendices A.8 and B.2 compute CRPS for a Gaussian fitted to the ensemble mean and standard deviation, rather than for the forecast ensemble itself. That approximation cannot represent the dry-point mass and skewed tail motivating the method. The lower reported CRPS therefore gives limited support to the claim of better calibration. Near-unbiased event counts are useful, but they also do not establish that predicted event probabilities are reliable.
  2. The stability of the benchmark improvements is difficult to judge. Table 10 specifies a single training seed, and Tables 1–2 provide no uncertainty on the differences. This matters for the roughly 2–4% aggregate CSI gains and for estimates at very rare thresholds. The four storm cases per dataset add useful examples, but their repeated lead times are correlated and do not resolve uncertainty across storms or training runs.
Questions for authors
  1. Does the probabilistic advantage persist when CRPS is computed directly from the ensemble samples? How reliable are ensemble exceedance probabilities at the heavy and extreme thresholds?
  2. Can the main comparisons include paired confidence intervals obtained by resampling storms, together with repeated training runs for HurdleCast and the matched controls? This would help distinguish stable gains from variation in training and the test events.
2 comments
AuthorsOpenReview ↗

W1/Q1: Empirical CRPS and Probabilistic Reliability

We thank the reviewer for their detailed and constructive feedback. We appreciate their recognition of our architectural contribution, the effectiveness of the ablations, and the routing and frequency-bias analyses. We have conducted additional experiments to address their concerns regarding probabilistic performance and statistical robustness.

We ran the direct empirical-CRPS comparison the reviewer suggested, and HurdleCast's advantage holds.

Evaluation protocol. Our Gaussian-fitted CRPS follows FlowCast's own convention (Appendix A.1.3), applied identically to every method for consistency with the existing benchmark protocol. This is an evaluation choice, not a constraint on the predictive distribution HurdleCast generates. The Gaussian approximation does not fully characterize the zero-inflated, heavy-tailed distribution that HurdleCast is designed to capture.

Empirical ensemble CRPS. We computed CRPS directly from the raw ensemble members, without parametric assumptions, using the bias-corrected estimator [1]. On the SEVIR test set, HurdleCast's empirical CRPS is 6–7% lower than its own Gaussian-fitted value across lead times. It also remains lower than FlowCast's empirical CRPS by −0.00046 (95% paired bootstrap CI [−0.00048, −0.00045]). This confirms that the probabilistic advantage is not an artifact of the Gaussian approximation.

Exceedance-probability reliability. The exceedance relationship is monotonic at both VIL ≥ 181 and VIL ≥ 219, indicating that higher predicted probabilities correspond to higher observed event frequencies. However, predicted probabilities exceed observed frequencies, indicating overforecasting. The rank histogram is also U-shaped, with observations falling in the lowest rank bin 27.5% of the time, compared with 11.1% expected across nine equally likely rank bins for an eight-member ensemble.

Predicted probabilityObserved frequency (VIL ≥ 181)Observed frequency (VIL ≥ 219)
0.0000.0010.000
0.1250.1050.079
0.2500.1770.136
0.3750.2390.189
0.5000.3010.245
0.6250.3680.307
0.7500.4470.384
0.8750.5530.492
1.0000.7690.717

These findings indicate a calibration limitation distinct from the improvements in empirical CRPS and ensemble-mean CSI/HSS. The rank histogram suggests ensemble underdispersion, while the reliability results show overforecasting of extreme-event probabilities. Ensemble size can affect probabilistic diagnostics [2], and ensemble reliability remains a recognized challenge in precipitation forecasting [3]. We acknowledge this limitation and will include the additional diagnostics and their interpretation in the appendix.

References

[1] Ferro, C. A. T., Richardson, D. S., and Weigel, A. P. (2008). On the effect of ensemble size on the discrete and continuous ranked probability scores. Meteorological Applications, 15(1), 19–24. https://doi.org/10.1002/met.45

[2] Richardson, D. S. (2001). Measures of skill and value of ensemble prediction systems, their interrelationship and the effect of ensemble size. Quarterly Journal of the Royal Meteorological Society, 127, 2473–2489.

[3] Mullen, S. L., and Buizza, R. (2002). The impact of horizontal resolution and ensemble size on probabilistic forecasts of precipitation by the ECMWF ensemble prediction system. Weather and Forecasting, 17(2), 173–191.

AuthorsOpenReview ↗

W2/Q2: Statistical Robustness and Repeated Runs

We thank the reviewer for raising the important question of statistical robustness. Performance stability across held-out test events and across independent training runs represents two distinct sources of variability, which we address separately.

Paired confidence intervals. To quantify uncertainty arising from the evaluation sample, we conducted paired bootstrap comparisons between HurdleCast and FlowCast using matched SEVIR test examples (10,000 resamples, 95% confidence intervals). Identical resampled indices were applied to both models to preserve the paired comparison.

MetricHurdleCast − FlowCast95% CI
CSI-M ↑+0.00586[+0.00554, +0.00621]
HSS-M ↑+0.00802[+0.00758, +0.00845]
Empirical CRPS ↓−0.00046[−0.00048, −0.00045]

All confidence intervals exclude zero under the specified resampling procedure. These results support improvements in aggregate CSI, HSS, and empirical CRPS relative to FlowCast.

Repeated training seeds. We agree that independent training runs would provide additional evidence regarding the stability of the reported improvements. However, multi-seed evaluation is computationally demanding for large generative spatiotemporal forecasting models. A single HurdleCast training run requires approximately 11 days on two NVIDIA B200 GPUs (528 GPU-hours), making repeated training of HurdleCast and matched baselines infeasible within the discussion period.

This computational constraint is not unique to HurdleCast. Closely related generative nowcasting approaches, including FlowCast and CasCast, likewise report their principal benchmark results without multi-seed uncertainty estimates, reflecting a broader practical challenge in evaluating large-scale generative forecasting models.

To partially address the reviewer's concern, we instead conducted paired bootstrap analyses to quantify uncertainty in performance differences across held-out evaluation examples, conditional on the trained model instances. We recognize that these confidence intervals characterize evaluation-sample variability, and will explicitly acknowledge this distinction in the revised manuscript.

Extreme-event case studies. The individual extreme-event case studies were included as an additional stress test under particularly challenging precipitation scenarios. They allow us to examine model behavior during specific high-impact events beyond the aggregate metrics, rather than serving as a substitute for statistical uncertainty analysis.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Revised Manuscript Uploaded

We thank the reviewers for their careful reading and detailed feedback. We have revised the paper in response to their comments and provided detailed responses to each reviewer separately. Changes are marked in blue\textcolor{blue}{\text{blue}} in the revised paper.

AuthorsOpenReview ↗

General Comments and Revision Summary

Dear Reviewers and Area Chair,

With less than 48 hours remaining in the discussion period, we would like to thank all four reviewers again before it closes.

Our aim with HurdleCast is to directly address extreme precipitation, one of the most consequential sources of forecast error, through hurdle-gated supervised regime specialization.

  • Process-Informed Generative Dynamics: HurdleCast embeds hurdle structures directly into the continuous-time vector field of a latent conditional flow-matching model: a process-informed generative dynamics formulation.
  • Supervised Regime Specialization: Our ablations demonstrate that supervised occurrence/severity gating, not added model capacity, is what recovers extreme-event skill.
  • State-of-the-Art Extreme-Precipitation Skill: The resulting method achieves state-of-the-art performance on both SEVIR and MeteoNet, with a 9–31% relative improvement at the highest precipitation intensity thresholds, without sacrificing general predictive skill. The aggregate CSI gain (2–4%) shows this comes alongside, not instead of, overall performance (Table 1).

All four reviewers found the paper informative (Informativeness ≥ 2, KXMR, voVq, and 2K7M: "I learned something interesting or useful"; LZUT: "I gained a substantially deeper understanding of the problem or question"). Three of four (KXMR, voVq, LZUT) specifically praised the ablations and experimental results, including the gains at extreme precipitation thresholds; 2K7M's strengths focus instead on the hurdle-econometrics framing, as detailed below.

Strengths the reviewers highlighted:

  • KXMR: "the hurdle decomposition gives the routing mechanism an interpretable role"; our ablations "provide useful evidence that the improvement involves learned regime specialization, rather than simply more parameters."
  • voVq: "the ablation studies demonstrate that supervised gating effectively improves extreme precipitation prediction beyond simply increasing model capacity."
  • 2K7M: a "sound approach... rooted in a familiar idea of improving the model by using different experts," linking "the problem of precipitation forecasting to hurdle models, common in econometrics, but not regularly linked to precipitation forecasting."
  • LZUT: "clear presentation and thorough evaluation"; the hurdle-gated MoE is "intuitive and well-motivated," with ablations that "convincingly support its effectiveness, especially for extreme precipitation."

Our rebuttal, now reflected in a revised manuscript with changes highlighted in blue:

  • KXMR (probabilistic evaluation): The empirical CRPS computed directly from the raw ensemble, the paired bootstrap confidence intervals on CSI-M, HSS-M, and CRPS, and the calibration diagnostics are now in the appendix, including the full rank histogram as a table rather than a single summary statistic.
  • voVq (threshold sensitivity, computational cost, MSE and FAR trade-off): The domain basis for our threshold choices, the measured inference latency against FlowCast and CasCast, and an explicit statement of the MSE and FAR trade-off we are making are now in the manuscript.
  • 2K7M (positioning relative to prior mixture-of-experts work): Jiang et al. and PA-Net are now cited in Related Work with the distinctions from our response, and Methods now elaborates further on the correspondence to the classical hurdle density.
  • LZUT (baseline comparability, qualitative evaluation): The PercpCast and DiffCast code inspection and the rank-histogram and reliability diagnostics are in the appendix, and the comparative LPIPS evaluation is now a column in the main SEVIR and MeteoNet comparison tables themselves.

We remain grateful for the time and effort invested in these reviews and are happy to clarify any open points before the discussion period ends.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.