acceptodds
Under review as a conference paper at ICLR 2027

MarCO: Propagating Covariate Uncertainty for Time Series Forecasting with Exogenous Variables

Abstract

Exogenous variables (i.e., covariates) often govern how a target time series evolves: electricity prices follow load and renewable generation, and wind power follows wind speed. Existing methods either forecast deterministically, encoding historical covariates as features or forecasting the covariates first and using that single trajectory, or forecast the target probabilistically while treating the covariates as fixed context. Neither models the future covariates, which are unknown when a forecast is issued, probabilistically, nor captures how their individual uncertainties combine into a joint distribution over covariate futures to which the target should respond. We propose MarCO, which treats the future covariates as a random vector: it forecasts their marginal distributions, couples the marginals into a joint distribution through a copula, draws coherent covariate scenarios from it, and lets every scenario modulate the target forecast, which is obtained as the mixture over scenarios. Using historical inputs only, MarCO ranks first in MSE on 11 of 12 real-world datasets against state-of-the-art deterministic and probabilistic methods and reduces their average MSE by 5.2% and 5.9%, while additionally quantifying the uncertainty of the covariates themselves.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Forecasting, Exogenous Variables, Probabilistic Modeling

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

The authors present MarCO, a method to incorporate uncertainty about future covariates to make better probabilistic time series forecasts. Specifically, they first produce multivariate marginal forecasts for the future covariates, then apply a learned Gaussian copula to structure the dependence between these, and lastly modulate the base target forecasts by conditioning on sampled future covariates from the copula.

Critical strengths
  • Favorable evaluation against a wide suite of baseline methods
Critical weaknesses
  • The motivation is rather unclear. The introduction suggests that the main goal is to "model the joint distribution of the covariate future across covariates and horizon steps", whereas Figure 1 and other parts of the text imply that the distinguishing features are probabilistic target forecasting that incorporates uncertainty about future covariates. However, it is not clear whether the joint distribution for the target forecast is of interest, and why we would expect MarCO to improve the target forecasts at all over a well-trained base / direct forecasting approach.

  • The main results do not evaluate the paper's supposed contribution. The main evaluations in Table 1 are on point forecasts (MSE and MAE) of the target variable, which does not measure the paper's contribution of improved probabilistic forecasts. At the very least, evaluations should compare CRPS against baselines. And if the joint distribution is of interest, then energy score and variogram score should be evaluated against baselines. For methods that only produce marginals, the method form [1] could be used.

[1] Baron et al. Efficiently Generating Correlated Sample Paths from Multi-step Time Series Foundation Models. https://arxiv.org/abs/2510.02224

Questions for authors
  • Why would we expect MarCO to improve at all over multivariate forecasting methods or direct forecasting, as there is no new information being added? It seems to me like a properly-trained and well-calibrated direct forecasting method should already give us optimal marginals in the base target forecasts, so it's not clear how the extra modulation could help.

  • How does MarCO's performance compare to time series foundation models (Chronos-2, TimesFM-3, etc.)? These models are state-of-the-art for time series forecasting and support multviarate forecasts with exogenous variables.

12 comments
AuthorsOpenReview ↗

W1 & Q1: Motivation Clarification and Gains over Direct and Multivariate Forecasting

We sincerely thank Reviewer NBgB for the detailed feedback. We apologize that our original presentation did not make our motivation sufficiently clear and lacked full-benchmark probabilistic metrics in the main text. Below, we clarify our formulation and present new cross-backbone ablations, 12-dataset CRPS evaluations, and foundation-model comparisons to address W1 & Q1, W2, and Q2.


W1 & Q1: Motivation Clarification and Gains over Direct and Multivariate Forecasting

1. Clarification of Motivation

Our primary goal is improving target (endogenous) forecasting by accounting for uncertainty in future covariates. Because multiple covariates jointly influence the target across horizon steps, we need to characterize their plausible future combinations rather than their marginal distributions alone. Modeling the joint distribution of future covariates therefore serves as the means to improve target forecasts, rather than a separate objective.

2. Why Our Method Improves over Direct and Multivariate Forecasting

We agree that an ideal direct forecaster can theoretically learn the optimal target distribution. However, achieving this is difficult with limited and noisy real-world data. Recent methods such as DAG [1] and GCGNet [2] have demonstrated the benefits of leveraging future covariate forecasting to improve target prediction. Building on this principle, MarCO offers two advantages:

  • Training: Unlike direct forecasting, which discards future covariate labels, MarCO uses observed future covariates as additional reliable supervision to learn their predictive distributions; and unlike conventional multivariate forecasting, which predicts all variables in parallel, MarCO uses target supervision to train a dedicated modulation module that explicitly leverages covariate predictions to improve target forecasting.
  • Inference: MarCO uses the learned covariate forecaster and modulation module to refine the base target distribution, without requiring observed future covariates or any additional information.

3. Direct Empirical Evidence Across Backbones

To verify that leveraging predicted future covariates improves over direct forecasting, we compare target test MSE across three methods with and without their dedicated future-covariate enhancement mechanisms (both variants retrained on the same machine under official per-dataset hyperparameters)—DAG (injecting temporal and channel correlations learned from covariate forecasting into the target forecaster), GCGNet (aligning future covariate and target trajectories through a shared variational graph structure), and MarCO (modulating the base target distribution over copula-coupled covariate scenarios):

DatasetModelw/o Future Covariate Enhancementw/ Future Covariate EnhancementImprovement (%)
DEMarCO0.4140.3974.11
DAG0.4430.4224.83
GCGNet0.4290.4153.26
FRMarCO0.3510.3373.99
DAG0.3630.3454.96
GCGNet0.3540.3384.52
Sdwpfh1MarCO0.6610.6324.39
DAG0.6650.6472.71
GCGNet0.6940.6683.75
Sdwpfh2MarCO0.7770.7394.89
DAG0.7840.7395.74
GCGNet0.8450.7975.68
Avg.MarCO0.5510.5264.35
DAG0.5640.5384.56
GCGNet0.5810.5554.30

Across backbones and datasets, models with future covariate enhancement consistently outperform their direct-forecasting counterparts by 4.3%–4.6% on average, confirming the effectiveness of explicitly forecasting future covariates under limited real-world data. Moreover, MarCO achieves the lowest overall MSE both without (0.551) and with (0.526) this mechanism.

We have updated Section 1 to clarify that target forecasting is our primary goal, and added Appendix A.5 (Table 9, referenced in Section 5.3) detailing the direct-variant ablation setups and cross-backbone MSE analysis (changes in blue).

[1] DAG: A Dual Correlation Network for Time Series Forecasting with Exogenous Variables. ICML 2026.

[2] GCGNet: Graph-Consistent Generative Network for Time Series Forecasting with Exogenous Variables. ICLR 2026.

AuthorsOpenReview ↗

W2: Probabilistic Target Forecasting Evaluation (CRPS) Across All 12 Datasets

W2: Probabilistic Target Forecasting Evaluation (CRPS) Across All 12 Datasets

Thank you for highlighting this important issue and helping us improve our manuscript.

To comprehensively evaluate probabilistic forecasting quality, we extended our experiments to the test sets of all 12 datasets using CRPS. As clarified in our response to W1 & Q1, because our primary objective is forecasting the endogenous target variable—while modeling the joint distribution of future covariates serves strictly as an internal mechanism to propagate uncertainty—we focus our probabilistic evaluation directly on target forecasting quality via CRPS.

1. Baselines and Full-Benchmark CRPS Results

We compare MarCO against three categories of baselines:

  1. KITE [1], an end-to-end probabilistic covariate-aware forecasting method.
  2. K²VAE [2] and TSFlow [3], channel-independent end-to-end probabilistic forecasting methods.
  3. Chronos-2 [4] and TimesFM-3 [5], covariate-aware time series foundation models evaluated in both zero-shot (ZS) and full-shot fine-tuned (FFT) settings.

The CRPS results are summarized below (bold: best, italic: second-best; lower is better):

DatasetMarCO (Ours)KITEK²VAETSFlowChronos-2 ZSTimesFM-3 ZSChronos-2 FFTTimesFM-3 FFT
NP0.1920.1980.2350.2910.1960.1900.1940.188
PJM0.1200.1240.1470.1710.1340.1290.1220.125
BE0.1710.1980.2080.2570.1840.1800.1810.180
FR0.1470.1500.1670.1950.1500.1470.1490.146
DE0.3010.3050.3390.4290.3230.3040.3020.303
Energy0.1650.1940.2110.2700.1880.1810.1870.179
Sdwpfm10.3430.3770.4140.6490.3720.3530.3650.347
Sdwpfm20.3730.4150.4270.6880.4030.3820.4000.379
Sdwpfh10.4090.4320.4520.7490.4210.4110.4160.410
Sdwpfh20.4380.4810.4830.7620.4650.4440.4630.447
Colbun0.0670.0770.2440.1450.0680.0650.0660.064
Rapel0.1670.1920.3090.2990.1730.1670.1730.165
Average0.2410.2620.3030.4090.2560.2460.2520.244

2. Key Takeaways

MarCO achieves the lowest average CRPS (0.241), ranking first on 8 of 12 datasets and second on 2, outperforming both fine-tuned foundation models and end-to-end probabilistic baselines (including lower CRPS than KITE on all 12 datasets, an 8.0% reduction on average). These results confirm that explicitly propagating future covariate uncertainty yields superior target predictive distributions.

We have added this probabilistic comparison as Table 2 in Section 5.2, with the evaluation protocol detailed in Appendix A.4 (changes in blue).

[1] KITE: Knowledge-Guided Probabilistic Modeling for Time Series Forecasting with Exogenous Variables. ICML 2026.

[2] K²VAE: A Koopman-Kalman Enhanced Variational AutoEncoder for Probabilistic Time Series Forecasting. ICML 2025.

[3] Flow Matching with Gaussian Process Priors for Probabilistic Time Series Forecasting. ICLR 2025.

[4] Chronos-2: From Univariate to Universal Forecasting. arXiv preprint arXiv:2510.15821, 2025.

[5] TimesFM-3: A Zero-Shot Foundation Model for Multivariate Forecasting. Google Research Blog, 2026.

AuthorsOpenReview ↗

Q2: Probabilistic and Deterministic Comparisons with Time Series Foundation Models

Q2: Probabilistic and Deterministic Comparisons with Time Series Foundation Models

Thank you for this valuable suggestion. We extended our evaluation to include Chronos-2 and TimesFM-3 under both zero-shot (ZS) and full-shot fine-tuned (FFT) settings, covering both probabilistic and deterministic forecasting performance across all 12 datasets.

1. Probabilistic Forecasting

Please refer to our response to W2 above, where MarCO achieves the lowest average CRPS (0.241) ahead of both TimesFM-3 FFT (0.244) and Chronos-2 FFT (0.252).

2. Deterministic Forecasting

We further compare deterministic point-forecast MSE under the same evaluation protocol (bold: best, italic: second-best; lower is better):

DatasetMarCO (Ours)Chronos-2 ZSTimesFM-3 ZSChronos-2 FFTTimesFM-3 FFT
NP0.2150.2540.2390.2030.225
PJM0.0680.0850.0830.0700.078
BE0.3470.3440.3520.3190.337
FR0.3370.3540.3560.3420.339
DE0.3970.4990.4310.4160.428
Energy0.1090.1180.1090.1040.104
Sdwpfm10.4850.6410.5800.5860.568
Sdwpfm20.5650.7370.6600.6510.645
Sdwpfh10.6320.7820.7360.7070.721
Sdwpfh20.7390.9330.8390.8300.821
Colbun0.0600.0730.0720.0610.066
Rapel0.1760.2360.2360.2140.225
Average0.3440.4210.3910.3750.380

MarCO achieves the lowest average MSE (0.344), ranking first on 9 of 12 datasets and second on 2. Notably, it outperforms both zero-shot and full-shot fine-tuned Chronos-2 and TimesFM-3—reducing average MSE by 8.3% and 9.5% over Chronos-2 FFT and TimesFM-3 FFT, respectively—with less than 4% of their parameters (≤4.2M\le 4.2\text{M} vs. 120M and 331M), demonstrating the effectiveness and efficiency of explicitly propagating future covariate uncertainty.

We have added these zero-shot and full-shot fine-tuned foundation-model comparisons as Table 8 in Appendix A.4 (changes in blue).


We thank Reviewer NBgB again for the valuable feedback. We hope our motivation clarification and new experiments (cross-backbone ablations, 12-dataset CRPS, and foundation-model comparisons) resolve your concerns, and we would be grateful if you could reconsider your assessment and rating. We welcome any further discussion.

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

The paper proposes MarCO for time-series forecasting with unknown future covariates. Instead of predicting a single future covariate trajectory, MarCO models future covariates probabilistically. It predicts their marginal distributions, uses Sklar’s theorem and a low-rank Gaussian copula to construct a joint distribution, samples coherent covariate scenarios, and uses them to modulate the target forecast. The final prediction is obtained by mixing the scenario-conditioned target distributions. Experiments on 12 real-world datasets show strong performance.

Critical strengths
  1. Clear problem formulation and strong motivation. The paper identifies an important limitation of existing methods: future covariates are often unknown, while prior approaches either ignore their future evolution or reduce it to a single trajectory. Modeling them probabilistically and propagating their uncertainty to the target is well motivated.
  2. Well-structured method design. The paper uses Sklar’s theorem to separate future-covariate marginals from their dependence structure, and uses scenario-conditioned location-scale modulation to propagate covariate uncertainty to the target forecast.
  3. Strong empirical performance and informative ablations. MarCO performs strongly on 12 real-world datasets, and the ablations clearly show the contributions of future-covariate modeling, scenario sampling, copula coupling, and mixture-level training.
Critical weaknesses
  1. Limited statistical robustness. All reported results use a single random seed (2021). Although the appendix specifies separate random streams for the copula factors, idiosyncratic noise, and within-scenario draws, it does not report independent retraining across multiple seeds. Since some gains over strong baselines are relatively small, multi-seed experiments with mean and standard deviation would provide stronger evidence for the stability of the reported improvements.
  2. Limited evaluation of probabilistic forecasting quality. The paper is primarily motivated by modeling and propagating future-covariate uncertainty, while the full 12-dataset benchmark is mainly evaluated with MSE and MAE. The paper does report CRPS and Energy Score in the five-dataset ablation, which is useful, but broader probabilistic evaluation across the full benchmark would better support the central uncertainty-modeling claim, especially in terms of calibration and distributional quality.
  3. Limited analysis of inference efficiency and sampling sensitivity. MarCO relies on Monte Carlo scenario sampling at inference time, using S=128 covariate scenarios and K=64 draws per scenario according to the appendix. However, the paper does not provide a detailed analysis of inference latency or the trade-off between the number of samples and forecasting quality. Since scenario sampling is a central component of MarCO, such an analysis would help clarify the practical cost of the proposed probabilistic formulation.
Questions for authors
  1. Could the authors clarify what randomness is controlled by the reported single seed (2021) and provide multi-seed results?
  2. Could the authors provide more comprehensive probabilistic evaluation on the full benchmark, such as CRPS or calibration-related metrics?
  3. Why is the median of the scenario mixture used as the point forecast for MSE? Have the authors compared it with the mixture mean?
  4. How sensitive is MarCO to the number of scenarios (S) and target draws (K) at inference time?
6 comments
AuthorsOpenReview ↗

W1 & Q1: Random Seed Scope and Multi-Seed Retraining Results

We sincerely thank Reviewer GexC for the positive assessment and constructive suggestions. Following the order of the review, we respond to W1 & Q1, W2 & Q2, W3 & Q4, and Q3 below.

W1 & Q1: Random Seed Scope and Multi-Seed Retraining Results

Thank you for highlighting this important issue and helping us improve our manuscript.

1. What the Random Seed Controls

We set the random seed to 2021 for Python, NumPy, and PyTorch. It therefore controls both training (parameter initialization, mini-batch shuffling, dropout, and scenario sampling) and inference (covariate-scenario and target sampling).

2. Independent Retraining Across Multiple Seeds

To evaluate the stability of our improvements over the strongest baselines, we independently retrained MarCO, DAG, and KITE from scratch across five random seeds, focusing on the five electricity-price datasets due to the limited rebuttal period. The table below reports the mean ± standard deviation of test MSE and MAE across the five runs (bold: best, italic: second-best):

ModelMetricNPPJMBEFRDEAvg
DAGMSE0.238±0.0010.072±0.0010.379±0.0030.350±0.0040.433±0.0090.294±0.003
MAE0.260±0.0030.164±0.0020.251±0.0130.189±0.0020.412±0.0040.255±0.002
KITEMSE0.230±0.0010.070±0.0010.371±0.0020.333±0.0050.443±0.0130.289±0.002
MAE0.255±0.0020.162±0.0010.260±0.0020.200±0.0020.420±0.0080.259±0.002
MarCOMSE0.217±0.0020.069±0.0010.348±0.0010.333±0.0030.398±0.0020.273±0.002
MAE0.256±0.0010.163±0.0010.239±0.0010.196±0.0020.401±0.0020.251±0.002

MarCO achieves the lowest average MSE (0.273±0.002) and MAE (0.251±0.002), reducing average MSE by 7.1% and 5.5% relative to DAG and KITE, respectively, with non-overlapping mean ± standard deviation intervals. Moreover, MarCO exhibits low variability across seeds (standard deviation ≤0.003\le 0.003 on all datasets, compared with up to 0.0130.013 for DAG and KITE). It ranks first in MSE on all 5 datasets and first or second across all 10 dataset–metric pairs, supporting the robustness of its improvements across random seeds.

We have clarified the random seed configuration in Appendix A.2 and Appendix A.7, and added these multi-seed results as Table 12 in Appendix A.7 (changes in blue).

AuthorsOpenReview ↗

W2 & Q2: Probabilistic Evaluation (CRPS) on the Full 12-Dataset Benchmark

W2 & Q2: Probabilistic Evaluation (CRPS) on the Full 12-Dataset Benchmark

Thank you for this valuable suggestion. To comprehensively evaluate the probabilistic forecasting quality of MarCO across all application domains, we extended our evaluation to the test sets of all 12 datasets using CRPS, a strictly proper scoring rule that jointly assesses distributional calibration and sharpness.

1. Baselines and Full-Benchmark CRPS Results

We compare MarCO against three categories of baselines:

  1. KITE [1], a covariate-aware end-to-end probabilistic forecasting method.
  2. K²VAE [2] and TSFlow [3], channel-independent end-to-end probabilistic forecasting methods.
  3. Chronos-2 [4] and TimesFM-3 [5], covariate-aware time series foundation models evaluated in both zero-shot (ZS) and full-shot fine-tuned (FFT) settings.

The test CRPS results are summarized below (bold: best, italic: second-best; lower is better):

DatasetMarCO (Ours)KITEK²VAETSFlowChronos-2 ZSTimesFM-3 ZSChronos-2 FFTTimesFM-3 FFT
NP0.1920.1980.2350.2910.1960.1900.1940.188
PJM0.1200.1240.1470.1710.1340.1290.1220.125
BE0.1710.1980.2080.2570.1840.1800.1810.180
FR0.1470.1500.1670.1950.1500.1470.1490.146
DE0.3010.3050.3390.4290.3230.3040.3020.303
Energy0.1650.1940.2110.2700.1880.1810.1870.179
Sdwpfm10.3430.3770.4140.6490.3720.3530.3650.347
Sdwpfm20.3730.4150.4270.6880.4030.3820.4000.379
Sdwpfh10.4090.4320.4520.7490.4210.4110.4160.410
Sdwpfh20.4380.4810.4830.7620.4650.4440.4630.447
Colbun0.0670.0770.2440.1450.0680.0650.0660.064
Rapel0.1670.1920.3090.2990.1730.1670.1730.165
Average0.2410.2620.3030.4090.2560.2460.2520.244

2. Key Takeaways

MarCO achieves the lowest average CRPS (0.241), ranking first on 8 of 12 datasets and second on 2. It outperforms both fine-tuned foundation models and end-to-end probabilistic baselines in average CRPS, achieving lower CRPS than KITE on all 12 datasets (8.0% reduction in average CRPS). These results demonstrate the strong probabilistic forecasting performance of MarCO across domains.

We have added this CRPS comparison as Table 2 in Section 5.2, with the evaluation protocol detailed in Appendix A.4 (changes in blue).

[1] KITE: Knowledge-Guided Probabilistic Modeling for Time Series Forecasting with Exogenous Variables. ICML 2026.

[2] K²VAE: A Koopman-Kalman Enhanced Variational AutoEncoder for Probabilistic Time Series Forecasting. ICML 2025.

[3] Flow Matching with Gaussian Process Priors for Probabilistic Time Series Forecasting. ICLR 2025.

[4] Chronos-2: From Univariate to Universal Forecasting. arXiv preprint arXiv:2510.15821, 2025.

[5] TimesFM-3: A Zero-Shot Foundation Model for Multivariate Forecasting. Google Research Blog, 2026.

AuthorsOpenReview ↗

W3 & Q4: Sampling Sensitivity (S, K) and Inference Latency

W3 & Q4: Sampling Sensitivity (S, K) and Inference Latency

Thank you for highlighting this important issue. To analyze the trade-off between sample size, forecasting quality, and inference cost, we varied S∈{2,4,…,256}S \in \{2, 4, \dots, 256\} and K∈{1,4,16,64}K \in \{1, 4, 16, 64\} at inference time and evaluated both test MSE and latency across all 12 datasets.

1. Forecasting Quality

The table below reports test MSE averaged over all 12 datasets for different numbers of covariate scenarios SS and target draws KK (the default setting S=128,K=64S=128, K=64 is highlighted in bold):

S∖KS \setminus K141664
20.46410.38400.35690.3500
40.41200.36420.35040.3471
80.38100.35410.34710.3455
160.36260.34890.34550.3447
320.35330.34640.34470.3443
640.34850.34510.34430.3441
1280.34620.34450.34410.3440
2560.34500.34420.34400.3440

Accuracy improves with both SS and KK and saturates quickly: MSE is clearly worse only with very few samples (e.g., 0.4641 at S=2,K=1S=2, K=1), whereas every setting with S≥64S \ge 64 and K≥4K \ge 4 stays within about 0.3% of the default (0.3440), and doubling SS beyond 128 brings no further gain at K=64K=64.

2. Inference Latency

We further report the corresponding inference latency (ms per batch of 64 windows on a single NVIDIA RTX3090 GPU, averaged over all 12 datasets) for the same S×KS \times K grid:

S∖KS \setminus K141664
23.873.913.923.91
43.863.903.903.90
83.873.913.933.95
163.883.923.924.45
323.893.913.945.47
643.873.894.437.63
1283.883.935.4813.38
2564.064.818.0223.90

Latency primarily depends on the total sample count S×KS \times K: it remains around 3.9–4.1 ms when S×K≤512S \times K \le 512 and grows roughly linearly beyond that, reaching 13.38 ms at the default setting and 23.90 ms at S=256,K=64S=256, K=64. Overall, settings with S∈[64,128]S \in [64, 128] and K∈[4,16]K \in [4, 16] keep MSE within about 0.3% of the default while providing a 2.4–3.4× speedup (3.89–5.48 ms vs. 13.38 ms); the default setting is therefore a conservative choice rather than a requirement.

We have added the numerical grids (Tables 10 and 11) and corresponding heatmaps (Figure 3) to Appendix A.6 of the revised manuscript (changes in blue).

Reviewer_xzq8OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MarCO, a probabilistic time series forecasting framework that explicitly models and propagates the uncertainty of future exogenous variables into target forecasting. The paper addresses a practical setting where only historical target and covariate observations are available, and future covariates are unknown at prediction time. Unlike existing approaches that either use historical covariates directly or rely on a single deterministic forecast of future covariates, MarCO models their future values probabilistically and incorporates multiple plausible scenarios into target forecasting.

Specifically, MarCO first employs a channel-wise Transformer with a shared quantile prediction head to estimate the marginal distributions of future exogenous variables and the target. A low-rank Gaussian copula then captures dependencies across covariates and forecasting horizons, enabling the generation of coherent future covariate scenarios. These scenarios are subsequently used to modulate the location and scale of the target's base quantile function. The final predictive distribution is obtained by mixing the scenario-conditioned target distributions. The framework is trained end-to-end using quantile regression losses, copula likelihood, and CRPS.

Experiments on 12 real-world datasets demonstrate competitive forecasting performance, achieving the best or tied-best MSE on 11 datasets. Additional ablation studies investigate the contributions of future covariate forecasting, probabilistic scenario sampling, copula-based dependence modeling, and the training objectives.

Critical strengths

S1. The paper addresses an important issue in time series forecasting with exogenous variables: future covariates are often unavailable at prediction time, and their uncertainty may substantially influence the target forecast. Rather than reducing future covariates to a single deterministic trajectory, MarCO explicitly models their predictive distributions and propagates the resulting uncertainty to the target. This provides a well-motivated perspective on probabilistic forecasting with uncertain exogenous information.

S2. A coherent and well-structured framework for propagating covariate uncertainty.

S3. Strong empirical performance across diverse application domains and informative ablation studies that support the main design choices.

Critical weaknesses

W1. Insufficient probabilistic forecasting evaluation and experimental robustness analysis.

Although MarCO is primarily motivated by propagating future covariate uncertainty into target forecasting, the main experiments on 12 datasets focus on deterministic metrics (MSE and MAE). Probabilistic forecasting metrics, including CRPS for the target and Energy Score for the covariate scenarios, are mainly reported in the ablation studies on five electricity price datasets using the validation set. Consequently, the experimental evidence for probabilistic forecasting quality across different application domains remains relatively limited. Furthermore, all reported experiments use a single random seed, making it difficult to assess the statistical reliability of the observed performance improvements, particularly when the differences between competing methods are small.

W2. Limited evidence for the additional benefits of copula-based dependence modeling.

The ablation study in Table 2 provides useful evidence that explicitly forecasting future covariates improves target forecasting, and that probabilistic scenario sampling offers additional benefits over using a single deterministic scenario. However, the improvements from Gaussian copula sampling over independent probabilistic sampling are relatively modest in the evaluated settings. Moreover, this comparison is limited to five electricity price datasets, leaving the generalizability of the copula's additional benefits to other application domains insufficiently established. Therefore, while the overall effectiveness of the proposed framework is supported, the specific contribution of joint dependence modeling warrants further empirical validation.

Questions for authors

Q1. Could the authors provide more comprehensive probabilistic forecasting evaluations and statistical reliability analyses (see W2)?

Q2. How consistent are the benefits of Gaussian copula modeling over independent scenario sampling across different application domains?

Table 2 shows that incorporating future covariate information provides substantial improvements, while the additional gains from modeling joint dependencies with a Gaussian copula are relatively modest. Could the authors further investigate whether these additional benefits are consistently observed on datasets beyond electricity prices, particularly those involving multiple correlated weather or energy-related covariates? Such comparisons would help clarify when explicitly modeling joint covariate dependencies is particularly beneficial compared with independent probabilistic scenario sampling.

5 comments
AuthorsOpenReview ↗

W1 & Q1: Full-Benchmark Probabilistic Evaluation (CRPS) and Multi-Seed Reliability

We sincerely thank Reviewer xzq8 for the positive assessment and constructive suggestions. Following the order of the review, we provide new full-benchmark CRPS evaluations, multi-seed reliability results, and cross-domain copula ablations to address W1 & Q1 and W2 & Q2 below.


W1 & Q1: Full-Benchmark Probabilistic Evaluation (CRPS) and Multi-Seed Reliability

Thank you for highlighting this important issue.

1. Probabilistic Evaluation on the Full Benchmark

To evaluate probabilistic forecasting quality across domains, we extended our evaluation from the validation sets of the five electricity-price datasets to the test sets of all 12 datasets using CRPS (bold: best, italic: second-best; lower is better), comparing MarCO against:

  1. KITE [1], a covariate-aware end-to-end probabilistic forecasting method.
  2. K²VAE [2] and TSFlow [3], channel-independent end-to-end probabilistic forecasting methods.
  3. Chronos-2 [4] and TimesFM-3 [5], covariate-aware foundation models in zero-shot (ZS) and full-shot fine-tuned (FFT) settings.
DatasetMarCO (Ours)KITEK²VAETSFlowChronos-2 ZSTimesFM-3 ZSChronos-2 FFTTimesFM-3 FFT
NP0.1920.1980.2350.2910.1960.1900.1940.188
PJM0.1200.1240.1470.1710.1340.1290.1220.125
BE0.1710.1980.2080.2570.1840.1800.1810.180
FR0.1470.1500.1670.1950.1500.1470.1490.146
DE0.3010.3050.3390.4290.3230.3040.3020.303
Energy0.1650.1940.2110.2700.1880.1810.1870.179
Sdwpfm10.3430.3770.4140.6490.3720.3530.3650.347
Sdwpfm20.3730.4150.4270.6880.4030.3820.4000.379
Sdwpfh10.4090.4320.4520.7490.4210.4110.4160.410
Sdwpfh20.4380.4810.4830.7620.4650.4440.4630.447
Colbun0.0670.0770.2440.1450.0680.0650.0660.064
Rapel0.1670.1920.3090.2990.1730.1670.1730.165
Average0.2410.2620.3030.4090.2560.2460.2520.244

MarCO achieves the lowest average CRPS (0.241), ranking first on 8 of 12 datasets and second on 2. Compared with KITE under the same forecasting setting, MarCO reduces average CRPS by 4.5%, 14.9%, 8.3%, and 13.0% across the electricity-price, Energy, wind-power, and hydropower domains.

2. Multi-Seed Retraining Reliability

To assess statistical reliability, we independently retrained MarCO, DAG, and KITE across five random seeds on the five electricity-price datasets (bold: best, italic: second-best):

ModelMetricNPPJMBEFRDEAvg
DAGMSE0.238±0.0010.072±0.0010.379±0.0030.350±0.0040.433±0.0090.294±0.003
MAE0.260±0.0030.164±0.0020.251±0.0130.189±0.0020.412±0.0040.255±0.002
KITEMSE0.230±0.0010.070±0.0010.371±0.0020.333±0.0050.443±0.0130.289±0.002
MAE0.255±0.0020.162±0.0010.260±0.0020.200±0.0020.420±0.0080.259±0.002
MarCOMSE0.217±0.0020.069±0.0010.348±0.0010.333±0.0030.398±0.0020.273±0.002
MAE0.256±0.0010.163±0.0010.239±0.0010.196±0.0020.401±0.0020.251±0.002

MarCO achieves the lowest average MSE (0.273±0.002) and MAE (0.251±0.002), lowering average MSE relative to DAG and KITE by 0.021 (7.1%) and 0.016 (5.5%) with non-overlapping mean ± standard deviation intervals. Based on unrounded means, MarCO ranks first in 7 of 10 dataset–metric pairs and first or second in all 10.

We have clarified the seed configuration in Appendix A.2 and Appendix A.7, added the multi-seed results as Table 12 in Appendix A.7, and included the CRPS comparison as Table 2 in Section 5.2 with the protocol in Appendix A.4 (changes in blue).

[1] KITE: Knowledge-Guided Probabilistic Modeling for Time Series Forecasting with Exogenous Variables. ICML 2026.

[2] K²VAE: A Koopman-Kalman Enhanced Variational AutoEncoder for Probabilistic Time Series Forecasting. ICML 2025.

[3] Flow Matching with Gaussian Process Priors for Probabilistic Time Series Forecasting. ICLR 2025.

[4] Chronos-2: From Univariate to Universal Forecasting. arXiv preprint arXiv:2510.15821, 2025.

[5] TimesFM-3: A Zero-Shot Foundation Model for Multivariate Forecasting. Google Research Blog, 2026.

AuthorsOpenReview ↗

W2 & Q2: Cross-Domain Generalizability of Gaussian Copula vs. Independent Sampling

W2 & Q2: Cross-Domain Generalizability of Gaussian Copula vs. Independent Sampling

Thank you for this valuable suggestion. To investigate whether the benefits of Gaussian copula modeling over independent sampling hold consistently beyond electricity prices, we extended the comparison between Independent (A=0A = 0, retrained from scratch) and MarCO from the five electricity-price datasets to the test sets of all 12 datasets across all four application domains, under the same per-dataset configurations. The target MSE and CRPS results are summarized below (bold: better):

DomainDataset (DD)MSE (Indep.)MSE (MarCO)Improvement (%)CRPS (Indep.)CRPS (MarCO)Improvement (%)
Electricity PriceBE (2)0.3480.3470.290.1800.1715.00
FR (2)0.3360.337-0.300.1480.1470.68
PJM (2)0.0740.0688.110.1240.1203.23
DE (2)0.4000.3970.750.3020.3010.33
NP (2)0.2240.2154.020.1960.1922.04
Avg.0.2760.2731.300.1900.1862.00
HydropowerColbun (2)0.0630.0604.760.0690.0672.90
Rapel (2)0.1810.1762.760.1740.1674.02
Avg.0.1220.1183.280.1220.1173.70
Wind PowerSdwpfm1 (6)0.4950.4852.020.3540.3433.11
Sdwpfm2 (6)0.5750.5651.740.3810.3732.10
Sdwpfh1 (6)0.6430.6321.710.4170.4091.92
Sdwpfh2 (6)0.7540.7391.990.4490.4382.45
Avg.0.6170.6051.860.4000.3912.37
Multi-Source EnergyEnergy (5)0.1150.1095.220.1700.1652.94
All 12 DatasetsOverall Avg.0.3510.3441.850.2470.2412.40

Improvement is the relative error reduction from Independent to MarCO, with aggregate improvements computed from dataset-level values before rounding the averages.

Conclusion: The benefits of Gaussian copula modeling over independent sampling hold consistently across all four application domains, achieving lower CRPS on all 12 datasets (12/12) and lower MSE on 11 of 12 datasets (reducing overall average MSE by 1.85% and CRPS by 2.40%). Moreover, as anticipated by the reviewer, the gains are more pronounced on domains driven by multiple correlated weather or energy covariates—specifically Hydropower (3.28% in MSE, 3.70% in CRPS), Multi-Source Energy (5.22% in MSE, 2.94% in CRPS), and Wind Power (1.86% in MSE, 2.37% in CRPS)—than on Electricity Price (1.30% in MSE, 2.00% in CRPS), supporting the value of modeling joint covariate dependencies.

We have added this 12-dataset cross-domain ablation study as Table 13 in Appendix A.8 (changes in blue).


We sincerely thank Reviewer xzq8 again for the positive assessment and valuable suggestions. We hope that the full-benchmark CRPS evaluation, multi-seed reliability study, and 12-dataset cross-domain copula analysis fully resolve your concerns and strengthen your support for our work. We welcome further discussion on any remaining questions.

AuthorsOpenReview ↗

Kindly Request for Reviewer's Feedback

Dear Reviewer xzq8,

Since the End of the Rebuttal comes soon, we would like to inquire if our response addresses your primary concerns. If they do, we would greatly appreciate it if you could take them into account in your final assessment. We would also be glad to discuss any further questions or suggestions. Thank you again for your positive and constructive review!

Kind Regards, Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.