acceptodds
Under review as a conference paper at ICLR 2027

AmpTS: Amplitude-Guided Progressive Restoration Method for Time-Series Forecasting

Abstract

Time-series forecasting (TSF) plays an important role in a wide range of real-world applications. Existing TSF methods typically either generate the complete future sequence through a single direct mapping or formulate forecasting as a noise-conditioned generative task that progressively recovers the future through multiple denoising stages. Although the latter provides a progressive forecasting perspective, diffusion-based approaches generally rely on stochastic perturbations to construct intermediate states, with their progressive processes constrained by predefined noise-evolution paths. To address this limitation, we propose AmpTS, an amplitude-guided progressive restoration method that uses structure-preserving reference amplitude levels to organize iterative transitions among self-generated restoration states. Starting from a zero state, AmpTS repeatedly re-estimates the complete future sequence at each restoration stage by conditioning on historical observations and the current state, enabling multi-stage progressive forecasting without relying on stochastic perturbations to construct restoration states. Extensive experiments on real-world datasets show that AmpTS achieves competitive and consistent performance against representative direct forecasting and diffusion-based forecasting methods, demonstrating the effectiveness of amplitude-guided progressive restoration for time-series point forecasting.

Primary area:
learning on time series and dynamical systems
Keywords:
time-series forecasting, progressive restoration, amplitude-guided, self-generated restoration states

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_WPMSOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper reframes deterministic point forecasting as a multi-stage restoration process. Instead of corrupting the target with Gaussian noise, it defines reference states by rescaling the ground-truth horizon, Yt=1−t/T Y0Y_t=\sqrt{1-t/T}\,Y_0, and trains a step-conditioned network Rθ(H,Y^t,t)R_\theta(H,\hat Y_t,t) that re-estimates the whole horizon at every stage, starting from Y^T=0\hat Y_T=0. Transitions follow a deterministic DDIM-style rule (Eqs. 6–7) with a per-step learnable coefficient γt\gamma_t in Eq. 5. Experiments cover eight datasets against eleven non-diffusion and seven diffusion baselines.

Critical strengths

The question is a fair one: must the entire horizon be produced in a single mapping? Reporting a negative result on noise-based state construction is useful, and the empirical section is more careful than most. Appendix B.2 gives bootstrap intervals per restoration step, B.4 reports three-seed variance and Wilcoxon tests, and Fig. 5 together with the encoder-substitution study in Table 4 shows the authors thought about cost and about generality. The model itself is small and fully deterministic, which I think is the right instinct for point forecasting.

Critical weaknesses
  1. Eq. 1 defines YtY_ t, but YtY_ t appears nowhere in Algorithm 1, Algorithm 2, or the loss (Eq. 8 is a final-step MSE). Only the scalars αˉt\bar\alpha_ t and βˉt\bar\beta_t survive into computation. Combining Eqs. 6 and 7 makes the consequence explicit:
c_2(t)=\sqrt{\tfrac{t-1}{t}},\qquad c_1(t)=\sqrt{\tfrac{T-t+1}{T}}-\sqrt{\tfrac{(t-1)(T-t)}{Tt}},$$ where both coefficients depend on $(t,T)$ alone. At $T=15$, $c_2\approx0.89$–$0.97$ and $c_1\approx0.09$–$0.13$; at $t=1$, $c_1=1$ and $c_2=0$, so the output is just the last estimate; at $t=T$ the state is zero. The method is therefore a one-shot forecast followed by $T-1$ fixed-coefficient refinement steps. The ablation agrees: Generic Iterative Refinement ties on ETTh2 (0.378 vs 0.373), Exchange (0.347 vs 0.347) and Stock (0.051 vs 0.048). The claim that states need not follow "a predefined intermediate-state path" also sits awkwardly with Eqs. 1–2. 2. Comparing with[1], their Eq. 5, $\hat X^0=\frac{W X^t+(1-bW)D}{(1+cW)^d}$, is Eq. 5 here at $(b,c,d)=(2,-1,\tfrac12)$. Their Eq. 6 simplifies to $\frac{X^t-\sqrt{\bar\alpha_t}\hat X^0}{\sqrt{1-\bar\alpha_t}}$, which is Eq. 6 here; their Eq. 9 with $\sigma_t=0$ is Eq. 7 here. ARMD likewise re-estimates the full horizon at every step and samples deterministically. The net deviation is the state construction (scalar multiples of $Y_0$ instead of sliding windows) plus end-to-end training from zero — the latter forced by the former, since self-generated states have no ground-truth target, and it leaves no restoration level individually constrained (cf. §4.4, where train/val loss falls with $T$ but test error does not). ARMD appears in no table. 3. Against the strongest competitor per dataset in Table 1, relative MSE is $-0.23\%$ (ETTh1), $-0.27\%$ (ETTh2), $-3.61\%$ (Exchange), $-0.83\%$ (Weather), $-1.75\%$ (Energy) and $-2.04\%$ (Stock), but $+5.57\%$ (ETTm1) and $+3.26\%$ (ETTm2), where it is worse; median $-0.55\%$. Section 4.2 calls this "a clear overall advantage", while B.4's own Wilcoxon test gives $p=0.69$ for MSE. Several baselines also look broken — P-sLSTM at Stock MSE $1.832$ is $31.6\times$ TimesNet and $10.3\times$ DLinear in the same table, and SSOL ($8.441$), CSDI ($1.067$) and CN-Diff ($2.144$) are comparable — so "11 of 16 best" is measured against a weakened field. *Minor issues:* - Tables 1 and 9 disagree under the same stated protocol (Energy $0.112$ vs $0.113\pm0.0001$; TQNet on Weather $0.242$ vs $0.250\pm0.0064$). Table 1 reads as a best run, and its typical margin ($0.001$–$0.002$) is below the baselines' own spread ($0.006$–$0.014$). - $T\in\{3,4,5,8,15,18\}$ and the encoder variant are chosen per dataset with no stated protocol; $T=3$ on Exchange barely counts as progressive, and $T=15/18$ on the ETTm datasets conflicts with Fig. 3. - No schedule-shape ablation (linear / cosine / constant), which is the cheapest test of whether the levels matter. - Fig. 4's monotone gain is largely an artifact of the zero start; the net gain over a single pass is $4\%$ ($0.414\to0.398$). - No argument for why iterating should help a target the history already determines; four-horizon averages hide per-horizon behaviour; Fig. 5 omits parameter counts; there is no limitations section. - N-BEATS already performs multi-stage, complete-horizon residual refinement and is neither cited nor compared; Cold Diffusion is cited as motivation but never used as a baseline. [1] Auto-Regressive Moving Diffusion Models for Time Series Forecasting. AAAI
Questions for authors
  1. Where in Algorithms 1–2 is YtY_t used, and if nowhere, what remains of "amplitude guidance" beyond a fixed scalar schedule?

  2. Are the Table 1 numbers three-seed means, and how do they square with Table 9?

  3. What is new in Eqs. 5–7 relative to ARMD's Eqs. 5, 6 and 9, and why is ARMD absent from the comparison?

  4. How were TT and "History Mode" selected, and how does T=15/18T=15/18 on the ETTm datasets square with Fig. 3?

1 comment

Dear Reviewer WPMS,

As the rebuttal period is coming to an end today, we would like to kindly ask whether you have had a chance to review our latest response.

To directly address your remaining concern, we have further clarified the amplitude-guided state transition and provided new controlled experimental results comparing AmpTS with a simpler state-mixing rule.

We sincerely hope you could take a moment to review these additional results and share your thoughts before the deadline. We would greatly appreciate your further feedback and consideration.

Thank you again for your time and constructive suggestions!

Reviewer_Bx9vOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a multistep method for time-series point forecasting. Starting from a zero state, the paper explores an alternative to stochastic diffusion for point forecasting. It evaluates eight datasets, compares against direct and diffusion-based methods, and includes useful ablations of the iterative process and state update.

Critical strengths

Competitive results across several datasets. AmpTS performs well on a broad set of forecasting benchmarks. The method produces a full-horizon forecast from a zero initial state without stochastic sampling. This is an appealing design for applications that require point predictions.

Critical weaknesses
  1. The methodological novelty is not sufficiently established. Equations (6)–(7) are algebraically equivalent to a deterministic DDIM update expressed through a predicted clean sequence and residual coordinate. AmpTS differs in its zero initialization, forecasting network, and final-forecast training objective, but the paper does not isolate the contribution of these choices.
  2. The central comparisons are missing. There is no explicit, compute-matched DDIM-style baseline or controlled test of zero initialization in a diffusion forecasting model.
  3. The notation makes the algorithm difficult to follow, for example Y0Y_0 is the ground truth, Y^t\hat Y_t is a generated state, and Y^0(t)\hat Y_0^{(t)} is a full-horizon estimate at step tt.
Questions for authors
  1. The methodological novelty is not sufficiently established. Equations (6)–(7) are algebraically equivalent to a deterministic DDIM update expressed through a predicted clean sequence and residual coordinate. AmpTS differs in its zero initialization, forecasting network, and final-forecast training objective, but the paper does not isolate the contribution of these choices.
  2. The central comparisons are missing. There is no explicit, compute-matched DDIM-style baseline or controlled test of zero initialization in a diffusion forecasting model. The “Gaussian Diffusion” ablation does not specify enough about its initialization, objective, schedule, and sampling procedure to answer whether these choices explain the performance difference.
  3. The notation makes the algorithm difficult to follow, for example Y0Y_0 is the ground truth, Y^t\hat Y_t is a generated state, and Y^0(t)\hat Y_0^{(t)} is a full-horizon estimate at step tt.
1 comment

Dear Reviewer Bx9v,

As the rebuttal period is coming to an end today, we would like to kindly ask whether you have had a chance to review our previous responses, additional experiments, and manuscript revisions. We sincerely look forward to your further feedback and would greatly appreciate it if you could provide a response when time permits, so that we can clarify any remaining concerns in a timely manner.

Thank you again for your time and valuable feedback!

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

This paper proposes AmpTS, an amplitude-guided progressive restoration method for deterministic time-series point forecasting, addressing the limitation that existing diffusion-based forecasting methods rely on stochastic noise to construct intermediate states, which can degrade temporal structures. Instead of noise perturbation, AmpTS defines structure-preserving reference amplitude levels to organize iterative transitions among self-generated restoration states, starting from a zero state and repeatedly re-estimating the full future sequence at each stage via a step-conditioned restoration network that fuses historical priors and current states with adaptive step-wise modulation and residual-coordinate transitions. Extensive experiments on eight real-world datasets show AmpTS achieves competitive or superior performance against both representative direct forecasting and diffusion-based baselines, with ablation and compatibility studies validating the effectiveness of its core components and its favorable accuracy-efficiency trade-off.

Critical strengths
  1. The paper introduces a novel amplitude-guided deterministic progressive restoration paradigm (Section 3) that eliminates reliance on stochastic noise perturbations in multi-stage forecasting, preserving temporal structures across restoration stages and providing a principled alternative to diffusion-based time-series forecasting methods.
  2. The experimental evaluation (Sections 4.2–4.5 and Appendix B) is comprehensive and rigorous, covering eight diverse real-world datasets, over ten representative direct and diffusion-based baselines, detailed ablation studies, cross-architecture compatibility tests, computational efficiency analysis, and statistical significance tests, providing solid and multifaceted empirical evidence.
  3. The proposed AmpTS framework features a lightweight design and strong generalizability, as it consistently improves forecasting performance when integrated with different backbone encoders (DLinear, iTransformer) while maintaining low training/inference overhead and memory usage, making it a flexible, deployable enhancement for existing forecasting pipelines.
Critical weaknesses
  1. The amplitude-guided progressive restoration framework presented in Section 3 relies entirely on empirically motivated heuristic designs for the state transition rule, residual coordinate formulation, and step-adaptive fusion mechanism, with no formal theoretical analysis of error accumulation across restoration stages or convergence guarantees, which is a fixable but notable gap that limits the methodological rigor of the work.
  2. The related work discussion in Section 2 fails to clearly delineate the novelty of AmpTS relative to deterministic diffusion variants and existing iterative refinement methods, as the core structure-preserving deterministic state transition idea shares substantial conceptual overlap with simplified deterministic diffusion schedules, weakening the paper’s claimed paradigm-level contribution.
  3. The experimental evaluation in Section 4 is constrained to standard offline multivariate forecasting benchmarks with fixed prediction horizons, and the optimal number of restoration steps varies drastically across datasets (from 3 to 18) without generalizable tuning guidance, which is a fixable practical limitation that reduces confidence in the method’s robustness for diverse real-world deployment scenarios.
Questions for authors
  1. Could you provide any theoretical analysis of how prediction errors accumulate across successive restoration stages, and whether the progressive process has guaranteed monotonic error reduction or convergence properties?
  2. How does the proposed amplitude-guided deterministic state transition fundamentally differ from existing deterministic diffusion sampling schedules (e.g., DDIM) applied to time-series forecasting, and what is the unique novelty beyond replacing Gaussian noise with amplitude scaling?
  3. Is there a principled mechanism to automatically select the optimal number of restoration steps T for a given dataset or task, rather than requiring per-dataset grid search as shown in the hyperparameter table?
1 comment

Dear Reviewer ReYK,

As the rebuttal period is coming to an end today, we would like to kindly ask whether you have had a chance to read our detailed responses and revised manuscript. We highly value your feedback and sincerely hope you could provide further comments when convenient, so that we can address any remaining concerns before the deadline.

Thank you very much for your time and assistance!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.