acceptodds
Under review as a conference paper at ICLR 2027

Anchor-Preserving Coarse-to-Fine Diffusion for Probabilistic Time Series Forecasting

Abstract

Probabilistic time series forecasts should represent uncertainty at both individual time steps and aggregated temporal resolutions. Reliable uncertainty at individual steps, however, does not necessarily imply reliable aggregate uncertainty. Coarse-to-fine generation can use deterministic coarse forecasts to guide refinement. Yet fixing every sample's aggregate to the same forecast eliminates aggregate uncertainty, while conditioning alone does not guarantee that the aggregate ensemble mean matches it. Preliminary experiments on two datasets show that combining a point forecast with centered stochastic deviations reduces CRPS. With fixed final coarse scenarios, controlled comparisons provide limited support for centering before detail generation. We propose an anchor-preserving coarse-to-fine diffusion framework that constrains the coarse ensemble mean rather than individual aggregate outcomes. Residual diffusion generates stochastic coarse scenarios, which are centered around a deterministic anchor before conditional detail generation. A conditional detail diffusion model then refines each scenario using projected detail with zero block averages, so refinement cannot alter its coarse aggregate. The resulting forecasts satisfy sample-wise coherence and coarse ensemble-mean preservation up to numerical precision while retaining stochastic aggregate variation. The proposed method achieves the lowest mean CRPS in 15 of 24 dataset–horizon settings among methods with complete three-seed results, but substantially underperforms on Electricity and Traffic. Matched-mean evaluation of shared forecast ensembles further shows that fine-resolution gains need not translate into better aggregate forecasts.

Primary area:
learning on time series and dynamical systems
Keywords:
probabilistic time series forecasting, diffusion models, coarse-to-fine generation, forecast coherence, aggregate uncertainty

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_J4Y5OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes an anchor-preserving coarse-to-fine diffusion framework for probabilistic time series forecasting. Coarse residual samples are centered around a deterministic forecast, while projected detail generation preserves each sample’s coarse aggregates. The method achieves the lowest mean CRPS in 15 of 24 settings, but its gains over simple mean correction are limited, and performance deteriorates substantially on Electricity and Traffic.

Critical strengths

Clear formulation and structural guarantees. The paper distinguishes ensemble-mean preservation from fixing individual aggregate outcomes, retaining coarse uncertainty while ensuring sample-wise coarse–fine consistency.

Informative controlled comparisons. Fixed-coarse-scenario and matched-mean experiments help separate the effects of centering order, forecast location, and distributional modeling. Evaluating the same trajectories across resolutions is also appropriate.

Transparent reporting. The authors report results across three seeds and openly investigate failure cases, baseline reproduction issues, and methodological limitations.

Critical weaknesses

Limited additional benefit from the core design. The method combines established ensemble recentering, null-space projection, and conditional diffusion. Pre-centering improves CRPS over post-centering by only 0.0024 and 0.0007 on the two evaluated datasets. Simple mean correction substantially narrows the gap to D³U and slightly outperforms the proposed method at aggregate resolutions on Weather, leaving the necessity of the more elaborate pipeline insufficiently established.

Severe degradation on two major datasets. At horizon 96, CRPS is 0.8667 on Electricity versus 0.1725 for TM+G, and 1.1141 on Traffic versus 0.2093 for D³U. Near-100% coverage of nominal 90% intervals and substantial improvements after removing coarse stochasticity indicate severe overdispersion. The capacity analysis partially explains this behavior but does not resolve the performance deficit.

Insufficient justification for enforcing an MAE-trained anchor as the ensemble mean. MAE targets the conditional median, which can differ from the conditional expectation. Although acknowledged, this mismatch may cause centering to remove useful location corrections learned by the residual model. Comparing MAE- and MSE-trained anchors and testing sensitivity to anchor bias would strengthen the justification.

Missing configuration and computational-cost analyses. The aggregation factor is fixed at b=4b=4, without sensitivity analysis. Five-fold cross-fitting and two-stage diffusion also introduce substantial computational work, yet training time, inference latency, memory usage, and parameter counts are not reported. These measurements are needed to assess whether the gains justify the added cost.

Questions for authors

see weakness

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

This paper investigates probabilistic time series forecasting with consistency constraints across temporal resolutions. It distinguishes two requirements: preserving a prescribed deterministic coarse forecast as the mean of a stochastic ensemble, and ensuring that each fine-resolution trajectory aggregates to its corresponding coarse scenario.

To address these requirements, the authors propose an anchor-preserving coarse-to-fine diffusion framework. A deterministic TimeMixer predictor first provides a coarse forecast anchor. A residual diffusion model then generates stochastic coarse scenarios, which are recentered to preserve the anchor as their empirical ensemble mean. Subsequently, a conditional detail diffusion model generates fine-resolution variations within the null space of the temporal aggregation operator, ensuring sample-wise coherence by construction.

Experiments on six real-world datasets and four forecasting horizons show competitive fine-resolution CRPS in 15 of 24 settings. Additional analyses investigate the effects of centering order, aggregate uncertainty, and coarse diffusion capacity.

Critical strengths

[1] The paper explicitly distinguishes ensemble-level anchor preservation from sample-wise coarse-to-fine coherence. This distinction is relevant to multi-resolution probabilistic forecasting, where preserving a deterministic reference forecast should not eliminate uncertainty in individual aggregate outcomes.

[2] The combination of ensemble recentering and null-space projection provides a mathematically transparent mechanism for satisfying both constraints. In particular, the projected detail diffusion process remains in the null space of the aggregation operator, ensuring sample-wise coherence independently of denoising accuracy. These guarantees are established through straightforward and verifiable algebraic arguments.

[3] The paper goes beyond standard benchmark comparisons by examining pre-centering versus post-centering, evaluating common forecast ensembles at multiple temporal resolutions, and comparing methods after matching their coarse ensemble means. The authors also explicitly acknowledge substantial failures on Electricity and Traffic and investigate the influence of coarse diffusion capacity. These analyses provide useful insights into the distinction between structural consistency and distributional accuracy.

[4] The proposed method achieves the lowest mean fine-resolution CRPS in 15 of 24 dataset–horizon settings among methods with complete three-seed results. Its performance on the ETT datasets and Weather suggests that the proposed framework can be effective when the coarse predictive distribution is modeled adequately.

Critical weaknesses

[1] The proposed method exhibits severe overdispersion on Electricity and Traffic, leading to substantially degraded CRPS and prediction interval coverage close to 100%. The ablation studies in Table 5 demonstrate that coarse stochastic variation is the primary source of these failures, while Table 6 reveals considerable sensitivity to the capacity of the coarse denoiser. Although these diagnostics are informative, the underlying causes of such dataset-dependent behavior remain insufficiently understood. The deterministic coarse variant C^\hat{C} substantially improves CRPS but eliminates aggregate uncertainty, making it an incomplete solution for probabilistic forecasting. Consequently, the reliability and applicability of the proposed coarse uncertainty modeling mechanism remain unclear, particularly in high-dimensional settings.

[2] The framework explicitly constrains the empirical mean of the generated coarse scenarios to match a deterministic anchor trained with MAE. However, the population-optimal predictor under MAE corresponds to the conditional median, which does not necessarily coincide with the conditional mean of the target distribution. Although the authors acknowledge this distinction, the statistical implications of enforcing such an anchor as the predictive ensemble mean remain insufficiently established. In particular, this constraint may introduce systematic location bias when the conditional distribution is asymmetric or the deterministic predictor is inaccurate, potentially limiting the flexibility and accuracy of the resulting probabilistic forecasts.

[3] The proposed framework involves multiple training stages, including five-fold cross-fitting of the deterministic predictor and the training of two separate diffusion models. During inference, coarse scenarios and their corresponding fine-resolution details must be generated sequentially across the two diffusion stages. Despite this additional computational complexity, the paper provides no quantitative analysis of training time, inference latency, or computational resource consumption relative to the baselines. This omission makes it difficult to assess the practical efficiency of the framework and whether its forecasting improvements justify the additional computational cost.

Questions for authors

Q1. What causes the severe overdispersion in coarse uncertainty modeling, and how does denoiser capacity affect the relative performance of different coarse variants?

Tables 5 and 6 provide informative diagnostics, showing that coarse stochastic variation is responsible for substantial performance degradation on Electricity and Traffic, while increasing the coarse denoiser's capacity considerably improves performance on Traffic. However, the underlying causes of this behavior remain unclear. In particular, could the authors further compare the CRPS and prediction interval coverage of different coarse variants under larger denoiser capacities (e.g., hidden dimensions of 1024 or 2048)? Such a controlled comparison would help distinguish whether the observed failures primarily arise from insufficient model capacity or more fundamental difficulties in coarse uncertainty modeling. It would also provide useful guidance for selecting an appropriate coarse variant in practical applications.

Q2. What is the statistical motivation for enforcing an MAE-trained coarse anchor as the predictive ensemble mean?

As acknowledged in Section 3.2, MAE training generally targets the conditional median rather than the conditional mean. Meanwhile, the coarse diffusion model learns cross-fitted prediction residuals, which may contain systematic conditional location corrections, but Eq. (10) subsequently removes their empirical ensemble mean. Could the authors elaborate on the statistical motivation for this design and its implications for probabilistic calibration? Would using an MSE-trained anchor or otherwise accounting for conditional residual bias provide a useful alternative?

Q3. What is the additional computational cost introduced by the two diffusion models?

The proposed framework employs separate Coarse Diffusion and Fine Detail Diffusion models, which introduce additional training and sampling operations. Given that the deterministic point predictor can be regarded as a replaceable or pretrained component, our primary concern is the computational overhead associated with the two diffusion stages. Could the authors provide quantitative comparisons of diffusion training time, inference latency, and GPU memory consumption relative to existing single-diffusion forecasting methods? In particular, reporting the inference cost of generating the 100-sample ensembles used in the experiments would help clarify the practical efficiency of the proposed framework.

Q4. What additional benefits does scenario-conditioned detail generation provide compared with post-hoc mean correction?

The extended comparisons in Tables 3 and 4 are informative, particularly in demonstrating the broader applicability of mean correction to existing probabilistic forecasting methods. However, the observed differences between pre-centering and post-centering are relatively modest in the evaluated settings. Could the authors further clarify the specific advantages of conditioning Fine Detail Diffusion on already-centered coarse scenarios, beyond satisfying the same aggregate mean-preservation constraint? In particular, under what conditions would the proposed generation order be expected to provide meaningful benefits over simpler post-hoc correction?

Q5. What is the contribution of the additional anchor condition in Fine Detail Diffusion, and how sensitive is the model to the training–inference differences in its conditioning information?

In Eq. (12), the Fine Detail Denoiser receives the historical observations XX, the ground-truth coarse scenario UC∗UC^*, and the out-of-fold anchor UC~U\tilde C. Since the anchor is itself predicted from XX, its additional contribution to detail denoising is not immediately clear. Could the authors provide an ablation examining the effect of removing the anchor condition UC~U\tilde C (and UC^U\hat C at inference) while retaining the other inputs? Furthermore, training uses ground-truth coarse scenarios and out-of-fold anchors, whereas inference uses generated coarse scenarios and the final deterministic anchor. How sensitive is the detail generator to potential distributional discrepancies between these training and inference conditions?

Suggestion on Mathematical Formulation

S1. A possible simplification of the projected DDIM updates in Eq. (14).

The mathematical formulation of the projected diffusion process is interesting and provides an elegant way to guarantee sample-wise coherence. I would like to suggest a possible simplification of Eq. (14) that may make the null-space invariance more explicit.

According to Eq. (11), the forward diffusion process involves the projected noise Πξ\Pi\xi, rather than the original Gaussian noise ξ\xi. Defining the estimated effective noise as

ξ^t⊥(s):=Πξ^t(s)=Dt(s)−αˉtD^0∣t(s)1−αˉt,\hat\xi_t^{\perp(s)} := \Pi \hat\xi_t^{(s)} = \frac{D_t^{(s)}-\sqrt{\bar\alpha_t}\hat D_{0|t}^{(s)}} {\sqrt{1-\bar\alpha_t}},

the deterministic DDIM update can be written as

Dτ(s)=αˉτD^0∣t(s)+1−αˉτξ^t⊥(s).D_\tau^{(s)} = \sqrt{\bar\alpha_\tau}\hat D_{0|t}^{(s)} + \sqrt{1-\bar\alpha_\tau}\hat\xi_t^{\perp(s)}.

Since both Dt(s)D_t^{(s)} and D^0∣t(s)\hat D_{0|t}^{(s)} already lie in ker⁡(A)\ker(A), the estimated effective noise ξ^t⊥(s)\hat\xi_t^{\perp(s)} and the subsequent state Dτ(s)D_\tau^{(s)} also remain in this subspace by linearity. Therefore, the additional projections in Eq. (14) are algebraically redundant under exact arithmetic, and the simplified formulation preserves the same structural guarantee.

This is intended as a suggestion for improving mathematical clarity rather than a concern about the correctness of the proposed sampler. I would also be interested in understanding whether the repeated projections in the original formulation were primarily motivated by numerical stability or other implementation considerations.

Reviewer_acY7OpenReview ↗
Rating
3(+1)
Confidence
1
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes an anchor-preserving low-to-high-resolution diffusion framework that constrains the low-resolution ensemble mean rather than individual aggregation results. Overall, I think the performance is good and the method is clearly explained.

Critical strengths

The experiments are thorough, and the performance is good.

I appreciate that the methodology is easy to understand. The description is honest and detailed.

Critical weaknesses

First of all, this paper is extremely unfriendly to readers.

  1. D³U and TimeMixer suddenly appear. Okay, I know what they are, but introducing these abbreviated names without citations means I have to look them up individually and check whether they refer to what I think they do. There are so many instances like this that understanding the paper is genuinely difficult. What exactly are ETTh1 and ETTh2, and why do they suddenly appear in Table 1 without any explanation? Of course, some readers may already know them, but please at least provide citations. Please place the relevant citation right next to each abbreviated name when it first appears.

  2. I checked the ETT repository at https://github.com/zhouhaoyi/ETDataset, and it explicitly asks users to cite the associated paper. Why use the dataset without citing it? If this is the dataset you used, please cite it. Otherwise, please clarify which dataset you used.

  3. This may be something I have misunderstood, but it seems rather tricky. Were the other methods not tuned?

Questions for authors

N/A

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.