acceptodds
Under review as a conference paper at ICLR 2027

TimeGDC: Towards Efficient and Robust Joint Distribution Learning for Long-Term Probabilistic Time Series Forecasting

Abstract

Traditional point-based time series forecasting provides limited information for risk-aware decision-making. Long-term probabilistic forecasting methods, however, often suffer from low computational efficiency and limited robustness, especially when dealing with complex distributions and potentially contaminated observations over extended forecast horizons. To address these challenges, we propose TimeGDC, a contamination-aware framework for efficient joint probabilistic forecasting over long horizons. TimeGDC jointly models the predictive distribution over the entire forecasting horizon using a Gaussian mixture distribution that captures multimodal uncertainty. Each mixture component is parameterized with a Dynamically smoothed low-rank covariance structure to efficiently represent temporal dependencies among future time steps. Furthermore, we construct a Contamination-aware likelihood for joint forecasting by combining the predictive distribution with a fixed heavy-tailed reference distribution, thereby reducing the influence of abnormal observations on distribution learning. Extensive experiments on six widely used multivariate time-series benchmark datasets show that TimeGDC consistently outperforms four competitive probabilistic forecasting baselines, achieving average relative reductions of 16.6%, 7.7%, and 13.4% in CRPS, NMAE, and Energy Score, respectively.

Primary area:
learning on time series and dynamical systems
Keywords:
probabilistic time series forecasting, joint distribution learning, Gaussian mixture model, structured low-rank covariance, contamination-aware likelihood

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_FHooOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

TimeGDC targets long-term probabilistic multivariate time series forecasting by modeling the full forecasting horizon jointly per channel using a Gaussian mixture. Each mixture component uses a diagonal + low-rank covariance with dynamic temporal smoothing. To improve robustness, training uses a contamination-aware likelihood that mixes the model predictive density with a fixed Student-t reference using a fixed contamination rate. Experiments on six benchmarks report lower CRPS, NMAE, and Energy Score, plus an additional spike-perturbation robustness test.

Critical strengths
  1. The paper clearly frames the joint long-horizon dependency modeling and contamination robustness, and provides a coherent TimeGDC to address them.
  2. The use of diagonal + low-rank covariance to avoid full T×T covariance operations directly alleviates computational burdens in the training loop.
  3. The ablation table reports gains attributable to GMM, the low-rank covariance and temporal smoothing, contamination-aware loss. This effectively analyses the mechanism of TimeGDC.
Critical weaknesses
  1. Novelty is unclear, then leading to time GDC likely a composition of existing ideas without sufficient differentiation. GMM-based multimodality, low-rank covariance factorization, temporal smoothness regularization, and robust/heavy-tailed likelihood mixtures are each plausible individually, but the paper does not convincingly argue what is genuinely new beyond combining known components. This makes the contribution is incremental.

  2. The mathematical definition is not clear: 1. The core definition of CA in equation 11 lacks the definition of the function Γ\Gamma and the mechanism of pobs(y∣x)p_{obs}(y|x) in equation 12; 2. The definitions of E1 and E2 in the overall training objective are lacking.

  3. For the experimental setting, as stated in line 377, the experiments follow the same training setting for all models. This is not a fair comparison; some baselines may achieve near-optimal performance with fewer epochs, and training them for a larger number of epochs could lead to overfitting and degraded results. The setting should follow the official code.

  4. For experimental results, the results are lacking in Traffic, which is the standard dataset in the probabilistic time series forecasting task. This leads me to be concerned about how TimeGDC handles high-dimensional data. Reported results for NSDiff and K2VAE differ noticeably from the numbers reported in their original paper.

  5. The inference time comparison is not fully fair. SimDiff/NSDiff/D2VAE require iterative prediction at each time point, while TimeGDC is single-pass. What’s more, as in W.1, these baselines do not require the same number of training epochs as TimeGDC to achieve the best results.

  6. Although the ablation study explains the mechanism of TimGDC, it is still partial: 1. Too many hyperparameters need to be ablated, especially those related to CA and GMM. 2. The position embedding/Transformer blocks in the left panel of Figure 1 need to be ablated to highlight the effectiveness of the core proposition.

  7. The presentation is hard to follow, making the manuscript look like a simple addition of each mechanism. The references are not in brackets, further increasing the reading difficulty.

Questions for authors

see weakness

Reviewer_C5bQOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

TimeGDC addresses efficient and robust long-term probabilistic time-series forecasting. It models each channel's entire forecast horizon using a Gaussian mixture, capturing temporal dependencies with a low-rank covariance structure and temporal smoothing. A patch-based Transformer predicts the mixture weights, means, and covariance parameters in a single forward pass. Each component's covariance is parameterized as Σ=Diag(d)+UU⊤\Sigma=\mathrm{Diag}(d)+UU^\top, with a regularizer using learnable persistence coefficients between adjacent temporal factors. During training, a contamination-aware likelihood combines the predictive density with a fixed heavy-tailed Student-t reference density, reducing the gradient contribution of potentially contaminated observations. Experiments cover Electricity, four ETT datasets, and Weather, with forecast horizons of 96, 192, 336, and 720 steps. The authors report average relative reductions of 16.6%, 7.7%, and 13.4% in CRPS, NMAE, and Energy Score, respectively, against the comparison baselines, and the shortest inference times across all settings in Table 2. When approximately 10% of training target positions are perturbed with artificial spikes, the contamination-aware objective reduces CRPS by approximately 4.8% on average relative to standard NLL training.

Critical strengths
  1. An important practical problem. The paper addresses computational efficiency and robustness in probabilistic time-series forecasting.

  2. A reasonable methodological design. Combining Gaussian mixtures, structured low-rank covariance modeling, temporal smoothing, and a contamination-aware likelihood provides a coherent approach to the stated challenges.

Critical weaknesses
  1. Limited novelty in the problem formulation. Although the challenges are important, efficiency and robustness in probabilistic time-series forecasting are already well recognized. The problem identification itself therefore contributes limited novelty.

  2. Limited novelty in the core methodological ingredients. Gaussian-mixture-based forecasting has already been explored in time-series research, including the dynamic Gaussian mixture model of Wu et al. [1]. Likewise, low-rank-plus-diagonal covariance parameterizations have been used for efficient probabilistic forecasting by Salinas et al. [2]. These works differ from TimeGDC in their dependency structures and forecasting procedures, so they do not establish that the complete method is identical to prior work. Nevertheless, the use of GMMs and low-rank covariance alone should not be regarded as novel.

  3. The component-level gains and analysis do not sufficiently explain the overall improvement. Combining established techniques can constitute a valuable contribution, but this requires convincing performance gains and a careful account of where those gains originate. While Table 1 reports substantial improvements over external baselines, the individual gains in Table 3 are comparatively small. Additionally, more extensive qualitative and quantitative analyses are needed to verify that the components behave as intended. For example, the authors could examine whether the GMM adequately captures the range of plausible future outcomes and how the loss curvature changes with and without the contamination-aware objective.

References

[1] Wu et al., Dynamic Gaussian Mixture based Deep Generative Model For Robust Forecasting on Sparse Multivariate Time Series, AAAI 2021.

[2] Salinas et al., High-dimensional Multivariate Forecasting with Low-rank Gaussian Copula Processes, NeurIPS 2019.

Questions for authors

see weaknesses

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

TimeGDC addresses long-term probabilistic forecasting using a channel-independent Patch Transformer that predicts a Gaussian mixture over each channel’s complete forecast horizon. Each component has diagonal-plus-low-rank covariance. DTSLC regularizes adjacent loading rows using a learned persistence coefficient. Training initially uses ordinary negative log-likelihood and subsequently introduces a fixed Student-t contamination component; inference samples only the Gaussian mixture. Thus, joint modeling concerns future time steps within a channel, rather than the complete multivariate trajectory. Experiments cover six datasets and four horizons, with reported average improvements of approximately 16.6% in CRPS, 7.7% in NMAE, and 13.4% in Energy Score. The paper also reports runtime, component ablations, training-target contamination, and ACF analyses.

Critical strengths
  1. Meaningful empirical quality and efficiency. Table 1 shows substantial average improvements over the selected external baselines across six datasets. Table 2 supplies measured sampling costs: on Electricity with horizon 96, generating 100 trajectories per test window takes 5.09 seconds for the complete test set, versus 42 seconds for K²VAE.
  2. Relevant controls beyond the main table. Component ablations, matched contaminated-target training, Energy Score, and ACF-MSE provide useful initial evidence. The contamination experiment keeps validation and test targets clean, and CA improves CRPS and Energy Score on all six datasets in that setting.
Critical weaknesses
  1. Limited novelty is not compensated by sufficiently isolated incremental gains. Equations (4) and (7) describe a conditional mixture-of-factor-analyzers head, while the matrix acceleration and contamination responsibilities follow established constructions. The distinctive contribution therefore rests on the temporal regularization, contamination-aware training, and their combination. Table 3 already includes the full model, the smoothing-only variant (w/o CA), and the CA-only variant (w/o Smooth). The missing control is the same conditional mixture-of-factor-analyzers model with both smoothing and CA disabled. Without this fourth configuration under matched conditions, the results do not fully isolate the gains over the underlying explicit-density model or the interaction between the two additions. External baseline comparisons cannot resolve this attribution question, particularly without fuller documentation of channel-independent adaptations and tuning budgets.

  2. The evidence does not adequately identify the claimed mechanisms or their stability. CA assigns a responsibility to an entire forecast window, whereas Section 4.7 perturbs approximately 10% of target positions. The resulting weighting behavior for clean, contaminated, and legitimate extreme windows is not reported. Energy Score and ACF improvements provide useful evidence, but do not by themselves isolate improved conditional dependence from improvements in marginal forecasts. Appendix D defines ACF-MSE and the maximum lag, but it remains unclear whether predictive ACF is computed from individual sampled trajectories, the ensemble mean, or another point prediction, and how the results are aggregated. Finally, all experiments use one fixed seed; Table 3’s CRPS differences are approximately 0.29%–1.74%, without uncertainty estimates. These limitations are particularly consequential for assessing the incremental contributions, rather than the larger gaps to external baselines.

Questions for authors
  1. Additional evidence needed for the current contribution claims: Table 3 already reports the full model, w/o CA, and w/o Smooth. Could you add the missing configuration with both smoothing and CA disabled, while retaining the same encoder and Gaussian-mixture covariance parameterization? Comparing these four configurations under matched training schedules and tuning budgets, with multiple training seeds, would help isolate the contributions of smoothing, CA, and their interaction. Please report forecast quality and sampling cost for this matched comparison.

  2. Clarification: What exactly is the corruption sampling unit in Section 4.7, how are overlapping windows handled, and which horizon or horizon aggregation produces Table 4? These details determine how corruption of target positions translates into contamination of complete forecast windows and therefore how the whole-window responsibilities should be interpreted.

  3. Clarification and additional evidence needed for the current mechanism claims: Could you report responsibility distributions for clean, contaminated, and legitimate extreme windows? To assess whether the learned temporal dependence contributes beyond the marginal forecasts, a useful control would independently permute predictive-sample indices at each forecast time within each test window and channel, preserving the empirical per-time marginals while disrupting cross-time sample pairing, and then recompute Energy Score. Separately, Appendix D already defines ACF-MSE and the maximum lag; please clarify whether predictive ACF is computed from individual sampled trajectories, the ensemble mean, or another point prediction, and how it is aggregated across samples, windows, and channels.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.