- 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
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:
- KITE [1], an end-to-end probabilistic covariate-aware forecasting method.
- K²VAE [2] and TSFlow [3], channel-independent end-to-end probabilistic forecasting methods.
- 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):
| Dataset | MarCO (Ours) | KITE | K²VAE | TSFlow | Chronos-2 ZS | TimesFM-3 ZS | Chronos-2 FFT | TimesFM-3 FFT |
|---|---|---|---|---|---|---|---|---|
| NP | 0.192 | 0.198 | 0.235 | 0.291 | 0.196 | 0.190 | 0.194 | 0.188 |
| PJM | 0.120 | 0.124 | 0.147 | 0.171 | 0.134 | 0.129 | 0.122 | 0.125 |
| BE | 0.171 | 0.198 | 0.208 | 0.257 | 0.184 | 0.180 | 0.181 | 0.180 |
| FR | 0.147 | 0.150 | 0.167 | 0.195 | 0.150 | 0.147 | 0.149 | 0.146 |
| DE | 0.301 | 0.305 | 0.339 | 0.429 | 0.323 | 0.304 | 0.302 | 0.303 |
| Energy | 0.165 | 0.194 | 0.211 | 0.270 | 0.188 | 0.181 | 0.187 | 0.179 |
| Sdwpfm1 | 0.343 | 0.377 | 0.414 | 0.649 | 0.372 | 0.353 | 0.365 | 0.347 |
| Sdwpfm2 | 0.373 | 0.415 | 0.427 | 0.688 | 0.403 | 0.382 | 0.400 | 0.379 |
| Sdwpfh1 | 0.409 | 0.432 | 0.452 | 0.749 | 0.421 | 0.411 | 0.416 | 0.410 |
| Sdwpfh2 | 0.438 | 0.481 | 0.483 | 0.762 | 0.465 | 0.444 | 0.463 | 0.447 |
| Colbun | 0.067 | 0.077 | 0.244 | 0.145 | 0.068 | 0.065 | 0.066 | 0.064 |
| Rapel | 0.167 | 0.192 | 0.309 | 0.299 | 0.173 | 0.167 | 0.173 | 0.165 |
| Average | 0.241 | 0.262 | 0.303 | 0.409 | 0.256 | 0.246 | 0.252 | 0.244 |
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.
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):
| Dataset | MarCO (Ours) | Chronos-2 ZS | TimesFM-3 ZS | Chronos-2 FFT | TimesFM-3 FFT |
|---|---|---|---|---|---|
| NP | 0.215 | 0.254 | 0.239 | 0.203 | 0.225 |
| PJM | 0.068 | 0.085 | 0.083 | 0.070 | 0.078 |
| BE | 0.347 | 0.344 | 0.352 | 0.319 | 0.337 |
| FR | 0.337 | 0.354 | 0.356 | 0.342 | 0.339 |
| DE | 0.397 | 0.499 | 0.431 | 0.416 | 0.428 |
| Energy | 0.109 | 0.118 | 0.109 | 0.104 | 0.104 |
| Sdwpfm1 | 0.485 | 0.641 | 0.580 | 0.586 | 0.568 |
| Sdwpfm2 | 0.565 | 0.737 | 0.660 | 0.651 | 0.645 |
| Sdwpfh1 | 0.632 | 0.782 | 0.736 | 0.707 | 0.721 |
| Sdwpfh2 | 0.739 | 0.933 | 0.839 | 0.830 | 0.821 |
| Colbun | 0.060 | 0.073 | 0.072 | 0.061 | 0.066 |
| Rapel | 0.176 | 0.236 | 0.236 | 0.214 | 0.225 |
| Average | 0.344 | 0.421 | 0.391 | 0.375 | 0.380 |
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 ( 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.
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:
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):
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.