acceptodds
Under review as a conference paper at ICLR 2027

HyperDiff: High-Order Structure-Driven Diffusion for Multivariate Probabilistic Forecasting

Abstract

Multivariate probabilistic forecasting must model both future uncertainty and how that uncertainty is organized across interacting variables. Existing generative forecasters can represent flexible predictive distributions, but they rarely use dynamic group-wise dependence as an explicit uncertainty prior. We propose HyperDiff, a high-order structure-driven diffusion framework that treats a window-conditioned hypergraph as a data-dependent structural prior. HyperDiff learns group-wise relations from each historical window, uses them to organize history-derived volatility while preserving variable-specific heteroscedasticity, and feeds the resulting conditions into a variable-specific diffusion schedule. A deterministic history-based forecast provides the trend center, allowing reverse diffusion to focus on stochastic deviations around a plausible trajectory. Experiments on seven multivariate benchmarks and four forecasting horizons show that HyperDiff achieves lower CRPS and QICE than the evaluated probabilistic forecasting baselines. These results support high-order structural conditioning for multivariate probabilistic forecasting. Code is available at https://anonymous.4open.science/r/HyperDiff-1D2B.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Multivariate Probabilistic Forecasting, Diffusion Models, Dynamic Hypergraph Learning, High-Order Dependency Modeling

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_HvMEOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HyperDiff for multivariate probabilistic forecasting. It uses a dynamic hypergraph to organize historical volatility and combines a deterministic forecast center with adaptive diffusion scheduling to generate predictive distributions.

Critical strengths
  1. Motivation is reasonable, using dynamic relationships among features to improve uncertainty modeling.
  2. Results are generally good, with favorable CRPS and QICE performance across seven datasets and four forecasting horizons.
Critical weaknesses
  1. Novelty is somewhat limietd. Mean centering, variance priors, heteroscedastic diffusion, and reverse-variance recovery all have counterparts in NsDiff. The main additions are existing hypergraph propagation, gated fusion, and structure-conditioned scheduling. The method appears primarily to combine existing techniques.

  2. Insufficient coverage and comparison of recent related work. PPM[1] already studies generative forecasting guided by dynamic conditional priors, but this paper neither cites nor compares against it. CW-Gen [2] also uses conditional means and covariances to improve the generative process. Although CW-Gen was cited, it is not included in the analysis and experimental comparison.

  3. The method supervises its volatility branch using the rolling variance of the raw time series. However, large temporal variation does not necessarily imply low predictability; a regular periodic signal can also have high variance.

  4. Although the paper includes an ablation replacing the hypergraph with a graph, it lacks capacity-matched comparisons and controls using fixed or random structures. It is therefore difficult to determine whether the gains come from high-order relationships or additional parameters and conditioning information.

[1] Parametric Prior Mapping Framework for Non-stationary Probabilistic Time Series Forecasting

[2] Conditionally Whitened Generative Models for Probabilistic Time Series Forecasting

Questions for authors

See Weaknesses.

8 comments

W1. Methodological Novelty and Component Coupling

We appreciate the reviewer's observation. HyperDiff's novelty does not lie in individual hypergraph propagation or gating operations, but in the coupling of high-order structure, volatility modeling, and diffusion generation.

Specifically, the preliminary hypergraph regularizes history-derived volatility, volatility feedback refines the structure, and the refined structure subsequently conditions both noise scheduling and reverse denoising. This extends high-order dependencies from volatility modeling to probabilistic generation rather than merely stacking independent modules.

The ablation results support this design. On ETTh2, replacing one-step structural feedback with late fusion increases CRPS from 0.294 to 0.369, while replacing the hypergraph with a graph increases CRPS to 0.361 (Table 2). These results indicate that the gains depend not only on incorporating additional components but also on their interaction and high-order dependency modeling.

W2. Comparison with PPM and CW-Gen

We agree that PPM and CW-Gen are important related methods. We have incorporated both into our evaluation across seven datasets and four forecasting horizons.

Table RW2. Average CRPS / QICE (%) across four horizons (↓).

DatasetHyperDiff (Ours)PPMCW-Gen
ETTh10.387 / 1.3050.348 / 2.5450.477 / 4.572
ETTh20.349 / 2.2420.294 / 3.2660.438 / 4.607
ETTm10.341 / 1.7070.291 / 3.5870.387 / 3.009
ETTm20.293 / 2.1880.245 / 4.4840.325 / 3.545
Exchange Rate0.342 / 4.2300.340 / 5.1540.471 / 6.922
Traffic0.370 / 2.3490.214 / 5.7590.421 / 8.310
Electricity0.294 / 4.8870.197 / 6.5130.446 / 6.336

The results reveal different strengths rather than uniform superiority. PPM achieves lower average CRPS on all seven datasets, whereas HyperDiff achieves lower average QICE on all seven. Across the 28 dataset–horizon settings, PPM obtains the lowest CRPS in 26 cases, while HyperDiff obtains the lowest QICE in 22.

Methodologically, PPM employs dynamic parametric prior mapping, and CW-Gen uses conditional mean–covariance whitening. HyperDiff instead organizes history-derived volatility through a window-conditioned hypergraph and uses the refined structure to guide diffusion scheduling and denoising. The results suggest an advantage in quantile-interval coverage calibration, while PPM remains stronger in overall CRPS.

We have incorporated these comparisons into Table 1, Section 4.2, Appendix A, and Appendix H (Table H.1).

W3. Rolling Variance Versus Predictive Uncertainty

We agree that large temporal fluctuations do not necessarily imply high conditional predictive uncertainty. HyperDiff uses rolling variance as an auxiliary reference for local variation scale, rather than treating it as a direct estimate of irreducible predictive uncertainty. The learned high-order structure organizes this volatility information, while the conditioned diffusion process models future stochastic deviations.

To examine this distinction, we conducted a controlled intervention that amplifies predictable periodic components.

Table RW3. Effects of amplified predictable periodic components.

DatasetCRPS change90% interval width change90% coverage
ETTh1−4.33%+2.76%—
Electricity−6.09%+8.56%98.20% → 94.52%

Despite the increased interval widths, CRPS improves on both datasets. On Electricity, coverage moves closer to the nominal 90% level, although overcoverage remains.

These observations demonstrate the response of the forecasting pipeline to predictable amplitude changes, but do not establish that rolling variance equals conditional uncertainty. We therefore clarify its role as a local variation-scale reference and acknowledge the remaining calibration limitations.

Reviewer_G2gwOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

HyperDiff introduces a hypergraph into the diffusion model to better learn the conditional distribution of multivariate time series. It uses the hypergraph to capture the relationships among different variables and how these relationships change over time. The main idea is to introduce a conditional mean and variance so that the distribution at the end of the forward diffusion process becomes a normal distribution with the given mean and variance. In general, existing methods first learn the overall conditional mean and distribution, and then use regularization to guide the final distribution toward a standard normal distribution. In contrast, HyperDiff uses a conditional mean and variance that can change with the data. It first uses the information learned from the hypergraph to determine the mean and variance of the forward diffusion process, and then adds noise to the original data. During training, the model jointly learns the noise, the conditional mean, and the conditional variance. At each step, it also uses the current model output to update its estimate of the true data, which further adjusts the conditional mean. Finally, the reverse diffusion process generates samples whose distribution is close to the original data distribution.

Critical strengths
  1. By introducing a hypergraph, HyperDiff better captures the relationships among variables and their changes over time. This provides a useful way to handle variable dependencies and nonstationarity in conditional generative models for time series.
  2. HyperDiff introduces a relatively novel diffusion framework. In the forward process, it uses a conditional mean and variance to better control how noise is added to the data. During training, the model jointly learns the noise, conditional variance, hypergraph structure, and predicted values. In the reverse process, it also updates the estimated clean data at each step to adjust the denoising process.
Critical weaknesses
  1. During training, you use Eq. (C.23) to estimate the clean data. Since you already have an estimate of the clean data from Eq. (C.23), why do you still need so many complex steps to calculate the mean and variance for denoising at the current time step, and then perform denoising step by step?
  2. At each step, the model uses the currently estimated noise to update its estimate of the clean data. It then uses this updated estimate to calculate the mean and variance needed for the current denoising step. However, this process may introduce additional errors at every step. How can you ensure that these errors do not accumulate and negatively affect the final generation quality? Moreover, since the mean and variance are estimated at each step, how can you guarantee that the parameters used in the learned reverse process are consistent with those of the theoretical reverse process?
Questions for authors
  1. Since you already have an estimate of the clean data from Eq. (C.23), why do you still need so many complex steps to calculate the mean and variance for denoising at the current time step, and then perform denoising step by step?
  2. How can you ensure that these errors do not accumulate and negatively affect the final generation quality? Better offer some theorem to prove this.
  3. Since the mean and variance are estimated at each step, how can you guarantee that the parameters used in the learned reverse process are consistent with those of the theoretical reverse process?
  4. You need to compare your method with some representative methods, such as CW-Gen, D³U and Diff-MGR.
8 comments

Q1. Clean-State Reconstruction and Iterative Reverse Sampling

We thank the reviewer for this question. The distinction is between clean-state estimation and reverse-state evolution. Eq. (C.23) reconstructs a clean-state estimate from the current noisy state using the predicted noise and inferred variance scale. This estimate can be returned directly to obtain a one-pass sampler, while the subsequent equations specify how the sampling state evolves through the reverse process.

In HyperDiff, the reconstructed clean state, deterministic forecast center, and variable-specific volatility jointly determine the reverse transition. The reverse mean determines the transition’s location, and the estimated variance controls its stochastic dispersion. Each updated state then serves as the input to the next denoising step, allowing the clean-state estimate to be recomputed at successive noise levels.

To assess the contribution of these iterative updates, we compared one-pass clean-state reconstruction, few-step early stopping, and full reverse sampling on ETTh1 and ETTh2 at forecasting horizon 96, using identical evaluation settings across the three sampling strategies.

Table RQ1: Comparison of different reverse sampling strategies on ETTh1 and ETTh2.

DatasetSampling methodCRPS ↓QICE (%) ↓
ETTh1One-pass reconstruction0.3961.665
ETTh1Five-evaluation early stopping0.3931.631
ETTh1Full 20-step sampling0.3911.265
ETTh2One-pass reconstruction0.2441.630
ETTh2Five-evaluation early stopping0.2431.412
ETTh2Full 20-step sampling0.2401.375

The results presented in the table favor full reverse sampling for both CRPS and QICE, with early stopping yielding intermediate performance. This comparison supports the benefit of continuing reverse-state updates beyond the initial clean-state reconstruction for probabilistic accuracy and calibration. We have clarified this distinction after Eq. (C.23) in Appendix C.2.

Q2. Error Propagation and Its Effect on Generation Quality

We quantify error propagation at two levels: the forward variance recursion and the reverse sampling process.

(1) Forward variance stability. HyperDiff constructs its history-derived conditions through one structural refinement: H(0)⟶(Vint,Fvol)⟶H⋆H^{(0)} \longrightarrow \bigl(V_{\mathrm{int}},F^{\mathrm{vol}}\bigr) \longrightarrow H^\star and reuses c(X)=(Y^0,V,H⋆,αtt=1T)c(X)=\bigl(\widehat Y_0,V,H^\star,\\{\alpha_t\\}_{t=1}^{T}\bigr) throughout reverse sampling. For a fixed schedule, Eq. (B.21) gives νt=atV+btStar,at=Bt−B~t,bt=B~t\nu_t=a_tV+b_tS_{\mathrm{tar}},\qquad a_t=B_t-\widetilde B_t,\quad b_t=\widetilde B_t. Starting from a0=b0=0a_0=b_0=0, the coefficients satisfy at=αtat−1+(1−αt)2a_t=\alpha_ta_{t-1}+(1-\alpha_t)^2, bt=αtbt−1+αt(1−αt)b_t=\alpha_tb_{t-1}+\alpha_t(1-\alpha_t). Since 0≤αt≤10\le\alpha_t\le1, at,bt≥0,at+bt=1−∏r=1tαr≤1a_t,b_t\ge0,\qquad a_t+b_t=1-\prod_{r=1}^{t}\alpha_r\le1. For two variance references under the same schedule, ∣νt−νt′∣≤at∣V−V′∣+bt∣Star−Star′∣\lvert\nu_t-\nu_t'\rvert\le a_t\lvert V-V'\rvert+b_t\lvert S_{\mathrm{tar}}-S_{\mathrm{tar}}'\rvert. Therefore, \boxed{\lVert\nu_t-\nu_t'\rVert_\infty\le\max\\!\left\\{\lVert V-V'\rVert_\infty,\\,\lVert S_{\mathrm{tar}}-S_{\mathrm{tar}}'\rVert_\infty\right\\}}. This establishes a step-independent perturbation bound for the forward variance recursion.

(2) Finite-step reverse error propagation. Fix the historical observation X=xX=x. Let K^t\widehat K_t and Kt⋆K_t^\star denote the implemented and reference transition kernels, with P^t−1=P^tK^t,Pt−1⋆=Pt⋆Kt⋆\widehat P_{t-1}=\widehat P_t\widehat K_t,\qquad P_{t-1}^\star=P_t^\star K_t^\star. Assume finite second moments and the following transition bounds: W2!(K^t(y,⋅),K^t(y′,⋅))≤Lt∥y−y′∥W_2\\!\left(\widehat K_t(y,\cdot),\widehat K_t(y',\cdot)\right)\le L_t\lVert y-y'\rVert, sup⁡yW2!(K^t(y,⋅),Kt⋆(y,⋅))≤δt\sup_y W_2\\!\left(\widehat K_t(y,\cdot),K_t^\star(y,\cdot)\right)\le\delta_t. Define et=W2(P^t,Pt⋆)e_t=W_2(\widehat P_t,P_t^\star). The triangle inequality yields et−1≤W2!(P^tK^t,Pt⋆K^t)+W2!(Pt⋆K^t,Pt⋆Kt⋆)≤Ltet+δte_{t-1}\le W_2\\!\left(\widehat P_t\widehat K_t,P_t^\star\widehat K_t\right)+W_2\\!\left(P_t^\star\widehat K_t,P_t^\star K_t^\star\right)\le L_te_t+\delta_t. Iterating from T to 1 gives

\boxed{e_0\le\left(\prod_{j=1}^{T}L_j\right)e_T+\sum_{t=1}^{T}\left(\prod_{j=1}^{t-1}L_j\right)\delta_t}\tag{R1}

Here, eTe_T measures initialization error, δt\delta_t measures transition error, and the products of LjL_j determine subsequent amplification or attenuation. Empty products equal one, and the final deterministic reconstruction is included as a transition kernel. For contractive kernels satisfying Lt≤ρ<1,δt≤δL_t\le\rho\lt 1,\qquad\delta_t\le\delta, Eq. (R1) becomes

\boxed{e_0\le\rho^Te_T+\frac{1-\rho^T}{1-\rho}\delta\le\rho^Te_T+\frac{\delta}{1-\rho}}\tag{R2}

Thus, contractivity provides a sufficient condition for a uniform bound on accumulated perturbations.

(3) Effect on generation quality. Define the reference approximation error as bref=W2!(P0⋆,Pdata(⋅∣x))b_{\mathrm{ref}}=W_2\\!\left(P_0^\star,P_{\mathrm{data}}(\cdot\mid x)\right). Combining the triangle inequality with Eq. (R1),

\boxed{W_2\\!\left(\widehat P_0,P_{\mathrm{data}}(\cdot\mid x)\right)\le\underbrace{\left(\prod_{j=1}^{T}L_j\right)e_T}_{\text{initialization}}+\underbrace{\sum_{t=1}^{T}\left(\prod_{j=1}^{t-1}L_j\right)\delta_t}_{\text{transition errors}}+\underbrace{b_{\mathrm{ref}}}_{\text{reference approximation}}}\tag{R3}

For the standard Energy Score and a fixed observation y,

\left|\operatorname{ES}(\widehat P_0,y)-\operatorname{ES}(P_0^\star,y)\right|\le 2W_1(\widehat P_0,P_0^\star)\le 2e_0\tag{R4}

Equations (R1)–(R4) connect initialization and transition errors to the final distribution and predictive score. The general bound retains the actual propagation factors LtL_t; the contractive corollary applies under the additional condition Lt≤ρ<1L_t\le\rho\lt 1.

Q3. Consistency Between Learned and Analytical Reverse Parameters

HyperDiff promotes mean–variance agreement by supervising the reverse process against analytical posterior targets derived from the same forward process. Let m=Y^0(X)m=\widehat Y_0(X), vt=νt−1v_t=\nu_{t-1}, and Dt=αtvt+ut=νtD_t=\alpha_tv_t+u_t=\nu_t. For nondegenerate steps t>1t\gt 1, Gaussian conditioning gives s~t=utvtDt,μ~t=utmt−1+αtvt[Yt−(1−αt)m]Dt\widetilde s_t=\frac{u_tv_t}{D_t},\qquad \widetilde\mu_t=\frac{u_tm_{t-1}+\sqrt{\alpha_t}v_t[Y_t-(1-\sqrt{\alpha_t})m]}{D_t}, where mt−1=αˉt−1Y+(1−αˉt−1)mm_{t-1}=\sqrt{\bar\alpha_{t-1}}Y+(1-\sqrt{\bar\alpha_{t-1}})m. For the corresponding analytical mean parameterization with matched forward variance quantities, the mean error satisfies μθ,t−μ~t=−utαtνt(ϵ^θ−ϵ)\mu_{\theta,t}-\widetilde\mu_t=-\frac{u_t}{\sqrt{\alpha_t\nu_t}}(\widehat\epsilon_\theta-\epsilon).

This identity directly connects noise-prediction accuracy to posterior mean accuracy. We supervise noise and variance through Lϵ=E∣ϵ−ϵ^θ∣2,Lσ≡E[rt−log⁡rt−1],rt=s~tsθ,t\mathcal L_\epsilon=\mathbb E\\|\epsilon-\widehat\epsilon_\theta\\|^2,\qquad \mathcal L_\sigma\equiv\mathbb E[r_t-\log r_t-1],\quad r_t=\frac{\widetilde s_t}{s_{\theta,t}}, where numerical stabilization is omitted and ≡\equiv denotes equality up to an additive constant. The variance penalty is minimized at sθ,t=s~ts_{\theta,t}=\widetilde s_t. Joint agreement is characterized by the full Gaussian posterior KL: DKL(qt∣pθ,t)=12∑k[rt,k−log⁡rt,k−1+(μ~t,k−μθ,t,k)2sθ,t,k]D_{\mathrm{KL}}(q_t\\|p_{\theta,t})=\frac{1}{2}\sum_k\left[r_{t,k}-\log r_{t,k}-1+\frac{(\widetilde\mu_{t,k}-\mu_{\theta,t,k})^2}{s_{\theta,t,k}}\right].

For positive variances, this KL is zero exactly when both means and variances agree. Thus, the analytical posterior supplies explicit supervision targets, while the full KL provides a unified consistency criterion. The training losses promote agreement rather than guarantee exact equality after finite training; auxiliary variance recovery introduces a further approximation at inference. The expressions above use the unit-temperature reference kernel. For the reported sampling temperature τσ=1.2\tau_\sigma=1.2, the actual transition variance is τσ2sθ,t=1.44sθ,t\tau_\sigma^2 s_{\theta,t}=1.44s_{\theta,t}, which must replace sθ,ts_{\theta,t} when assessing that sampling kernel.

Reviewer_pRo7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HyperDiff for multivariate probabilistic forecasting. It learns a sparse hypergraph from each historical window as a structural prior, regularizes history derived volatility through a one step hypergraph correction while preserving variable specific heteroscedasticity, and uses the refined structure to generate variable level adaptive noise schedules for heteroscedastic diffusion centered on a deterministic forecast. It achieves better CRPS and QICE on seven benchmarks.

Critical strengths

The motivation is clear, and the paper distinguishes between using high order structure to organize uncertainty and homogenizing uncertainty across variables. The method forms a complete pipeline, with the window conditioned hypergraph, structure regularized volatility, and heteroscedastic diffusion each playing a well defined role, and the appendix gives a self consistent derivation of the variance recursion and the reverse variance. Inference relies only on historical information, which avoids target leakage. Experiments cover seven datasets and four horizons, show consistent gains in CRPS and QICE, and include ablation, sensitivity, efficiency, and robustness analyses.

Critical weaknesses

The novelty boundary of the diffusion backbone needs clarification. The closed form forward variance recursion, the reverse variance target, and the quadratic recovery of the auxiliary variance closely follow the NsDiff framework built on the location scale noise model, and the recent DiffPTS has already unified several diffusion paradigms under this model into a joint training objective. We suggest explicitly marking the inherited parts and highlighting the increment brought by the hypergraph.

The effect of structure on joint uncertainty is indirect. The hypergraph only adjusts marginal variance and the noise schedule, while diffusion noise remains diagonal and independent, so cross variable correlation enters only implicitly through the denoising condition, whereas recent conditional whitening methods explicitly use conditional means and covariances as priors. Evidence on the joint distribution also comes from only one dataset and one horizon.

The increment of the adaptive schedule is relatively limited. Beyond NsDiff, StaTS already learns a data adaptive schedule and modulates denoising strength across diffusion steps and variables, and this paper mainly adds a bounded variable level multiplicative factor. In the ablation, removing the scheduler has a small effect and even improves results on Exchange, so its necessity should be argued more thoroughly.

Questions for authors

Please refer to weaknesses.

3 comments

W1. Clarifying Contributions Beyond NsDiff and DiffPTS

We thank the reviewer for encouraging us to clarify our methodological contributions. NsDiff models time-varying uncertainty through location–scale diffusion, whereas DiffPTS focuses on unifying the diffusion training objective. However, in real-world multivariate time series, variables may exhibit time-varying higher-order dependencies in their volatility while retaining distinct volatility magnitudes and temporal patterns. Neither model explicitly addresses how to account for both properties in uncertainty modeling. We therefore propose HyperDiff to organize volatility information through dynamic higher-order dependencies and guide probabilistic generation, while preserving variable-specific heteroscedasticity.

To this end, HyperDiff introduces a hypergraph conditioned on the historical time window, with three core innovations:

  1. Structure-regularized volatility modeling. Structural regularization organizes volatility information extracted from historical observations without requiring structurally related variables to have identical variances.
  2. Volatility-guided structure refinement. Volatility feedback refines the higher-order structure through a single update.
  3. Structure-conditioned diffusion. The refined structure is incorporated into both variable-specific noise schedules and the reverse denoising process.

Together, these mechanisms extend the role of higher-order dependencies beyond feature enhancement, allowing them to directly shape uncertainty modeling and stochastic generation.

The ablation results in Table 2 support these design choices. On ETTh2, removing the hypergraph learning module increases CRPS from 0.294 to 0.353. Replacing the single-step feedback mechanism with late fusion increases CRPS to 0.369. We have expanded the Introduction and Appendices B–C to distinguish the standard diffusion formulations from the methodological contributions introduced by HyperDiff.

W2. High-Order Dependency Modeling and Joint Distribution Quality

We thank the reviewer for this valuable comment. To clarify how HyperDiff captures cross-variable dependencies despite using diagonal diffusion variance, we further elaborate on the role of high-order structural conditioning and extend our evaluations on the joint distribution, as summarized in the table below.

(1) High-Order Dependency Modeling: Although HyperDiff employs diagonal variance at each reverse diffusion step, this does not imply that the final multivariate predictive distribution is independent across variables. Specifically, HyperDiff learns a window-conditioned dynamic hypergraph and incorporates the refined high-order structure into both variable-specific noise scheduling and reverse denoising. This enables cross-variable structural information to influence the reverse generation process and the resulting joint predictive trajectories, without explicitly parameterizing a full conditional covariance matrix.

(2) Extended Joint-Distribution Evaluation: We extend the Energy Score (ES) and Variogram Score (VS) evaluations to ETTh1 and ETTh2 across four forecasting horizons (96, 192, 336, and 720). Averaged across these horizons, HyperDiff achieves the lowest ES on both ETTh1 (1.397) and ETTh2 (0.905), ranking first at all four horizons on ETTh1 and three of four horizons on ETTh2. For VS, HyperDiff achieves the best average performance on ETTh2 (1.757), while NsDiff performs better on ETTh1 (3.178 vs. 3.591). HyperDiff also achieves the lowest VS at horizon 96 on ETTh1 and horizon 720 on ETTh2. These results demonstrate competitive joint-distribution quality, while indicating that improvements in cross-variable dependence modeling vary across forecasting scenarios.

Manuscript Revision: We have refined the joint-distribution analysis in Section 4.2 and added the complete ES and VS comparisons in Appendix F (Figure F.1).

Table RW2: Average ES and VS results across four forecasting horizons (96, 192, 336, and 720) on the ETTh1 and ETTh2 datasets

ModelETTh1 ES ↓ETTh1 VS ↓ETTh2 ES ↓ETTh2 VS ↓
HyperDiff1.397 ± 0.02883.591 ± 0.13540.905 ± 0.01361.757 ± 0.0527
NsDiff1.448 ± 0.03203.178 ± 0.15820.933 ± 0.01401.803 ± 0.0541
KOONPRO3.690 ± 0.04646.747 ± 0.12642.591 ± 0.03892.727 ± 0.0818
TMDM1.549 ± 0.02573.308 ± 0.11161.213 ± 0.01822.827 ± 0.0848
DiffusionTS2.270 ± 0.03007.678 ± 0.21121.433 ± 0.02152.698 ± 0.0809
FlowTS2.307 ± 0.05886.596 ± 0.25342.132 ± 0.03206.773 ± 0.2032
TimeDiff1.903 ± 0.02475.130 ± 0.15981.096 ± 0.01652.062 ± 0.0619

W3. Distinction and Necessity of the Adaptive Scheduler

We thank the reviewer for this suggestion. HyperDiff's adaptive scheduler differs in leveraging high-order structure and volatility feedback to guide noise allocation, rather than merely introducing a bounded multiplicative factor.

(1) Methodological Distinction: While StaTS primarily uses spectral information for adaptive scheduling, HyperDiff employs volatility-refined high-order structure to guide both variable-specific noise scheduling and reverse denoising.

(2) Ablation Evidence: The scheduler's contribution is not uniformly marginal. Table 2 shows CRPS reductions of 18.3% on ETTh2 and 10.8% on ETTm1, with improvements in both CRPS and QICE on three of the four datasets.

(3) Dataset-Dependent Effects: On Exchange Rate, removing the scheduler reduces CRPS from 0.200 to 0.188, indicating that its benefits vary across datasets. Figure 4 further illustrates the scheduler's responses to volatility states and structural roles.

Manuscript Revision: We have expanded the comparison with StaTS in Appendix A and clarified the dataset-dependent ablation results in Section 4.3.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.