acceptodds
Under review as a conference paper at ICLR 2027

Spectral SNR-t Bias Correction for Diffusion Models in Non-Stationary Time Series Forecasting

Abstract

Diffusion probabilistic models have been increasingly applied to probabilistic time series forecasting. However, the inference-time signal-to-noise ratio–timestep (SNR-) bias previously observed in visual diffusion models remains insufficiently explored in this setting. SNR- bias denotes an inference-time mismatch between the empirical SNR of an intermediate denoising sample and the expected SNR implied by its nominal diffusion timestep. While training establishes a fixed correspondence between timesteps and noise statistics through the forward perturbation process, accumulated reverse-process errors can break this correspondence during inference. In this work, we systematically examine SNR- bias in diffusion-based time series forecasting across multiple models and benchmark datasets. We find that the bias exhibits a distinctive spectral signature: the estimated noise residual develops frequency-dependent power distortions along the reverse trajectory, a phenomenon we term **spectral SNR- bias**. This spectral distortion can introduce spurious trends, periodic artifacts, and miscalibrated prediction intervals. Building on this diagnosis, we propose **Spectral-Shape SNR Correction (SSSC)**, a training-free, plug-and-play inference procedure that adaptively reshapes intermediate samples toward the spectral profile expected at their nominal timesteps. SSSC uses spectral energy to modulate frequency-wise correction strengths calibrated on held-out data and abstains from correction when reliable calibration is not possible. Experiments across seven diffusion forecasters show that SSSC is particularly effective for over-dispersed forecasts, reducing CRPS by up to relative to uncorrected baselines while also improving point forecast accuracy and probabilistic calibration.

Primary area:
learning on time series and dynamical systems
Keywords:
Machine Learning, Time‑Series, Test‑Time Compute

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_x8dSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies spectral distortions associated with inference-time SNR–t mismatch in diffusion-based forecasting and proposes Spectral-Shape SNR Correction (SSSC). As defined in the method and algorithm, SSSC estimates average target and forecast power spectra, constructs candidate frequency-dependent gains, and selects a correction using calibration-side CRPS and an abstention rule. The selected gain is applied once to final forecast samples, without retraining the backbone. The paper includes diagnostics of noise-prediction spectra, paired evaluations across diffusion forecasters, component ablations, scalar correction controls, and a comparison with repeated in-loop correction. The full results report six backbones across thirteen datasets.

Critical strengths
  1. A simple correction with broad practical applicability. SSSC operates on forecast samples from frozen models and does not require modifying their training. This breadth provides useful evidence that output spectral correction can be beneficial beyond a single model.

  2. Several relevant controls and transparent reporting of mixed outcomes. The paper includes scalar scaling and dispersion controls, as well as a comparison with stepwise correction. It also reports settings where CRPS and QICE move in different directions and explicitly limits what the theoretical analysis guarantees. These are helpful features for assessing a calibration method rather than relying only on favorable examples.

Critical weaknesses
  1. The SNR–t diagnosis is not sufficiently connected to the final-output correction. The method uses average target and forecast spectra to select a gain applied after sampling. It does not directly measure or restore the SNR of intermediate states, a limitation acknowledged in the appendix. Appendix A gives a sufficient mechanism under an orthogonality assumption, rather than establishing that this mechanism explains the errors in the evaluated models. General output-scale or spectral errors could also benefit from the same filter. The distinction matters because the claimed contribution includes a diffusion-specific diagnosis, not only an effective post-processing rule.

  2. The abstract describes held-out calibration, while the method and algorithm describe spectral estimation and selection using training data. It remains unclear whether these windows are disjoint from those used to train the backbone. This is consequential because in-sample forecast spectra and errors need not represent deployment behavior.

  3. Section 4.3 reports global scaling and mean-preserving dispersion scaling for CSDI and TMDM on Electricity, which is a useful control. The stepwise comparison is also limited to two backbones on ETTh1. These results do not yet establish how consistently the frequency-dependent method adds value across the wider benchmark.

Questions for authors
  1. Which empirical result distinguishes correction of SNR–t bias from generic output spectral calibration? Is there evidence that settings with a larger diagnosed mismatch are the ones that benefit from SSSC, or should the mechanistic claim be limited to a sufficient explanation rather than an established cause?

  2. Are the spectral-estimation and selection windows disjoint from backbone training? Please give the temporal partitions and calibration-origin count. If training windows were reused, how do the results change under an independent temporal calibration block in a few representative settings?

  3. Why does the selection rule accept the correction in the ETTm2–TimeDiff failure case? What were the calibration-side gain and abstention statistic there, and how did they change at test time? How should users interpret cases where CRPS improves but QICE deteriorates substantially?

  4. Can the scalar calibration controls be evaluated on one additional dataset and a relatively strong baseline using the same calibration data and selection budget? Please also summarize the typical relative gain and its uncertainty across test origins, so that the practical conclusion does not depend primarily on the maximum improvement.

  5. Please provide the actual gmin⁡g_{\min}, gmax⁡g_{\max}, δ\delta, τ\tau, ϵ\epsilon, low-frequency cutoff, and spectral-band boundaries used in each setting, together with how they were chosen. Which TimeDiff implementation and configuration produced the reported forecasts, and are the exact selected corrections recoverable from the released configuration?

1 comment
AuthorsOpenReview ↗

Response to Reviewer x8dS

Thank you for the detailed review. We added temporal separation, scalar controls, and failure analysis.

1. Mechanistic attribution

Real-model evidence cannot uniquely exclude generic output spectral calibration. The appendix provides a sufficient mechanism under explicit assumptions, not an established cause of all improvements. TMDM scalar controls support extra CRPS benefit from frequency information, but not unique attribution to SNR–t bias. We will clarify that SSSC acts once on final samples; diffusion diagnostics motivate it but do not measure or restore intermediate SNR.

2. Calibration independence

Original experiments reused training windows; the abstract's held-out claim requires correction. New TMDM estimation/selection/test counts are 64/64/128 (ETTh1) and 32/32/64 (Electricity), with 30 samples per origin. Electricity target periods:

StageStartEnd
Estimation2018-08-07 06:002018-09-27 13:00
Selection2018-09-27 14:002018-11-24 21:00
Test2018-11-24 22:002019-07-02 01:00

Target windows crossing boundaries are excluded; inputs precede forecast origins. Separated Electricity CRPS is 0.449240 → 0.289827: 35.49% reduction [34.59%,36.59%]. Validation excludes optimization labels but contributed to early stopping. A stricter unused ETTh1 tail uses 64/64/128 origins: estimation 2018-02-21 to 2018-03-24 10:00; selection 2018-03-24 11:00 to 2018-04-24 21:00; evaluation 2018-04-24 22:00 to 2018-06-26 19:00. TMDM selects identity (0.587323 → 0.587323).

3. Original TimeDiff/ETTm2 failure

Calibration CRPS 0.378853 → 0.274413 improves 27.57%, exceeding the 3% acceptance margin. Reconstructed concentration statistic ρ=0.00118282\rho=0.00118282 exceeds the source-default rejection threshold 7.5×10−57.5\times10^{-5}, so that rule would accept it. The original JSON does not record the threshold; we will identify its source.

DC gain is approximately 0.559; mean low/middle/high gains are 2.949/2.808/2.706. Test CRPS worsens 0.200843 → 0.557726 (+177.69%), while QICE×100 improves 13.995 → 11.811. Calibration acceptance cannot guarantee test benefit. Paired low-frequency protection suggests mean/low-frequency changes contribute, but missing full gains/origin indices prevent precise recovery of original test ρ\rho or error attribution. We retain the failure. CRPS, point errors, QICE, and coverage should be reported together; one improved score does not establish overall calibration.

4. Scalar controls and uncertainty

We extended controls to TMDM/ETTh1 and retained Electricity. SSSC, global scaling, and mean-preserving dispersion scaling share selection windows, cached samples, nine candidates including identity, and acceptance margin; computation times are not matched.

Electricity results (SSSC uses separated calibration):

MethodCRPSReduction vs BaseQICE×100
Base0.449240—8.969
SSSC0.28982735.49%5.462
Global scaling0.37141417.32%1.374
Dispersion scaling0.31135630.69%3.793

SSSC reduces CRPS relative to global/dispersion scaling by 21.97% [21.00%,23.11%]/6.91% [6.49%,7.32%], using each scalar score as denominator. Scalars have better QICE: this supports extra CRPS benefit, not superiority on all metrics. All methods select identity on ETTh1 (CRPS 0.483879). Median per-origin relative improvement is 36.47%/0% on Electricity/ETTh1. Intervals average per-origin scores over three inference seeds before 2000 moving-block bootstrap replicates. Frozen checkpoints exclude retraining variability.

5. Parameters and recoverability

Defaults: gain bounds 0.5/3.0; η∈{0.5,1.0}\eta\in\{0.5,1.0\}; protection cutoff fraction 0.25; acceptance margin 0.03; rejection threshold 7.5×10−57.5\times10^{-5}. For H=192, rFFT has 97 bins: protection [0,24), non-DC bands [1,24), [24,48), [48,97). For H=24: [0,3), [1,3), [3,6), [6,13). Right endpoints are excluded. These empirical defaults are retained after sensitivity analysis. Rejection uses DC and band-average powers, not summed high-frequency energy; we will give the formula.

TimeDiff uses the local CNN, x0x_0 prediction, cosine schedule, 100 training steps and 20 native DPM sampling steps; recorded nfe=100 is not the sampling count. New runs preserve checkpoint checksums, configurations, candidate CRPS, selections, abstention reasons, and complete gains. Historical band-only records have limited recoverability.

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

The paper investigates inference-time signal-to-noise ratio–timestep (SNR-t) mismatch in diffusion-based time series forecasting and proposes Spectral-Shape SNR Correction (SSSC). SSSC is a training-free, plug-and-play inference procedure that adaptively reshapes intermediate samples of diffusion models toward the spectral profile expected at their nominal timesteps. SSSC has three components. Spectral Profile Estimation (SPE) estimates the power spectra of training targets and forecasts from the frozen model to identify frequency dependent energy discrepancies. Spectral Correction (SC) constructs bounded gains through spectral matching and energy gated shrinkage, with optional low frequency protection. Adaptive Selection (AS) selects a configuration using the continuous ranked probability score (CRPS) on training data, retaining the original forecasts when the required improvement is not achieved orthespectral stand down rule is triggered.

Critical strengths

1. Training free improvement across different diffusion settings

2. Ablation study shows empirical advantage of three components proposed in the paper

Critical weaknesses

1. Insufficient empirical evidence connecting spectral SNR-t bias to the proposed correction mechanism

Section 4.1 demonstrates sensitivity to SNR timestep mismatch and frequency dependent differences between forward and reverse noise-prediction spectra, it does not directly measure spectral SNR-t bias in intermediate reverse states. Furthermore, SSSC constructs correction gains from the power spectra of ground-truth targets and completed forecasts, without explicitly using intermediate state SNR information. Therefore, the extent to which the observed improvements specifically result from correcting spectral SNR-t bias, rather than general output spectral discrepancies, remains unclear.

2. Limited robustness of the correction under non-stationarity

SSSC estimates a single frequency-dependent gain by averaging spectral statistics across training windows and variables, then applies it uniformly to test forecasts. This may limit its effectiveness when spectral characteristics change across time or forecasting

Questions for authors

Q1: Although Figure 3(c) demonstrates frequency-dependent differences in noise-prediction spectra between forward and reverse states, can the authors directly quantify the spectral SNR-t bias of intermediate reverse states to further validate the proposed phenomenon.

Q2: Is there a quantitative relationship between the frequency dependent forward reverse mismatch observed in Figure 3(c) and the spectral correction gains constructed by SSSC ?

Q3: Since SSSC uses the same training side data to estimate spectral correction gains and select the best configuration based on CRPS, could the authors investigate whether this procedure leads to overfitting? In particular, how does performance change when gain estimation and configuration selection are performed on separate chronological subsets?

1 comment
AuthorsOpenReview ↗

Response to Reviewer NCxU

Thank you for recognizing the method's applicability without retraining and the component ablations. We address the mechanism, calibration reuse, and temporal transfer below.

Q1. Direct measurement of intermediate spectral SNR–t bias

We agree that forward/reverse noise-prediction spectra do not directly measure intermediate-state SNR. Real reverse samples lack an independently observable decomposition into reference signal, noise, and model error; predicted residuals also contain estimation error.

Our controlled experiment uses known signals, an oracle posterior reference, and paired reverse noise to analyze signal/error power and verify a sufficient mechanism under explicit assumptions. It does not provide direct SNR measurements in real TMDM states. We will describe real-model results as spectral diagnostics and clarify that SSSC corrects completed forecasts once, using gains fixed before testing.

Q2. Quantitative relationship between mismatch and gains

The matched branch relates gains to final-output spectra. With positive powers and inactive numerical flooring/clipping,

S_{corr}(f)=P(f)^\eta S(f)^{1-\eta},$$ where $P,S$ are calibration target/forecast power spectra. This contracts the calibration log-spectral discrepancy. It does not establish a universal relationship between Figure 3(c)'s intermediate noise-prediction mismatch and selected gains, or guarantee test-score improvement. Current TMDM evidence cannot establish that larger intermediate mismatch yields greater forecasting benefit. Gains are constructed and selected from final spectra and calibration CRPS, not intermediate SNR. We will make this distinction explicit. **Q3. Calibration reuse and chronological separation** The original experiments reused training windows. We added training reuse, validation reuse, and chronological separation of validation subsets for spectrum estimation and selection. Target windows crossing subset boundaries are excluded. Estimation/selection/test counts are 64/64/128 for ETTh1 and 32/32/64 for Electricity, with 30 samples per origin. Each row uses the same paired Base; positive CRPS reduction indicates improvement. | Dataset | Training-reuse reduction | Separated Base → SSSC | Separated reduction [95% CI] | |---|---:|---:|---:| | Electricity | 35.59% | 0.449240 → 0.289827 | 35.49% [34.59%,36.59%] | | ETTh1 | 0.00% | 0.483879 → 0.483879 | 0.00%, identity | Electricity's improvement persists after separation; ETTh1 retains identity. Thus the Electricity benefit does not depend on reusing identical windows for estimation and selection. These validation windows exclude backbone optimization labels but previously contributed to early stopping. We therefore also split the previously unused ETTh1 tail, after the original train/validation/test periods, into estimation, selection, and evaluation blocks. TMDM again selects identity: CRPS 0.587323 → 0.587323. This establishes selection behavior on independent temporal blocks, not an additional improvement. Checkpoints are frozen; three seeds vary inference randomness, not training. Intervals average per-origin scores across seeds before 2000 moving-block bootstrap replicates, accounting for temporal dependence but excluding retraining variability. **Nonstationarity** The gain is fixed before testing and shared across windows and variables. These experiments support limited temporal transfer on Electricity, not effectiveness under arbitrary spectral drift. Calibration abstention cannot guarantee future benefit. We will state this limitation explicitly.
Reviewer_STXQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The SNR–t bias between the predicted sample and the input timestep during inference has been demonstrated in image diffusion models. However, this bias remains underexplored in time series forecasting. This paper identifies a distinctive spectral characteristic of the SNR–t mismatch in time series: along the reverse diffusion trajectory, the estimated noise residuals exhibit frequency-dependent power distortions. Based on this observation, the paper proposes Spectral-Shape SNR Correction (SSSC), a training-free, plug-and-play method that adaptively corrects predicted samples. Specifically, SSSC adjusts the correction strength across different frequencies in the spectral domain and refrains from applying corrections when their reliability is insufficient. Extensive experiments demonstrate the effectiveness, synergy, and generalizability of SSSC.

Critical strengths

(1) The problem is meaningful and worth investigating. Time series data differ fundamentally from images and videos, making it worthwhile to investigate the SNR–t mismatch specifically in the context of time series forecasting.

(2) The proposed method is simple and effective. It requires no retraining, can be seamlessly integrated into various diffusion models, and introduces negligible computational overhead.

(3) The experiments are comprehensive. The evaluation covers multiple diffusion models and datasets. The paper also provides targeted ablation studies and visualizations to support its findings.

Critical weaknesses

(1) Lack of direct evidence connecting the core motivation to the proposed method. The paper is primarily motivated by the mismatch between the empirical SNR of intermediate states during reverse denoising and their nominal timesteps. However, SSSC actually performs a one-time frequency-domain correction on the final predicted samples after the sampling process is complete. While the existing experiments demonstrate that such post-processing can improve forecasting performance in certain settings, they do not establish that these improvements result from mitigating the SNR–t mismatch during inference. The paper also acknowledges that its trajectory experiments do not directly measure the SNR of intermediate reverse diffusion states. Furthermore, the abstract claims that the method reshapes intermediate samples, which appears inconsistent with the actual post-sampling correction mechanism. This discrepancy should be clarified or corrected.

(2) Insufficient discussion and comparison with relevant methods. Although the paper cites related work [1-3] on inference-time bias and frequency-domain correction, it does not adequately discuss the mechanistic differences between these approaches and SSSC or their respective applicability conditions. Moreover, the experiments only consider simple scaling operations as comparison methods, which is insufficient to substantiate the claimed superiority of the proposed approach. I also noticed that some of the SSSC results in Tables 1 and 4 are identical to those of the baselines, which further weakens the evidence supporting the method's contribution.

(3) Lack of robustness analysis for the method's hyperparameters. The paper does not provide sensitivity analyses for several key hyperparameters, including the gain bounds (gmin,gmax)(g_{min}, g_{max}), the low-frequency protection cutoff klok_{\mathrm{lo}}, the minimum improvement threshold δ\delta, the spectral rejection threshold τ\tau, and the number of training samples used for power spectral density estimation. Consequently, it remains unclear whether the observed performance gains depend on careful hyperparameter tuning or whether the proposed method is robust across different parameter settings.

[1] Frequency Regulation for Exposure Bias Mitigation in Diffusion Models, ACM MM 2025

[2] Elucidating the SNR-t Bias of Diffusion Probabilistic Models, CVPR 2026

[3] Spectral Prior for Reducing Exposure Bias in Diffusion Models, ECCV 2026

Questions for authors

See the Weaknesses section.

1 comment
AuthorsOpenReview ↗

Response to Reviewer STXQ

Thank you for recognizing the relevance, simplicity, and experimental coverage. We have added comparisons and sensitivity analyses and will clarify the mechanistic claims.

1. Motivation and correction location

We agree that the abstract misstates the intervention. SSSC applies a fixed frequency-dependent gain once to completed forecast samples, selected before testing using calibration data. We will align the abstract, method, and algorithm.

Forward/reverse noise-prediction spectra are diagnostics, not direct measurements of intermediate-state SNR. The appendix gives a sufficient mechanism under assumptions including orthogonality of error to reference signal and noise. A controlled experiment with known clean signals, shared reverse noise, and injected spectral errors shows that reverse errors can produce output spectral distortion and that final correction can reduce it. This does not establish the cause of all real-model improvements.

We will distinguish diffusion-process diagnostics, conditional mechanistic explanations, and paired forecasting results. The directly supported contribution is output spectral correction without retraining.

2. Related methods and identity results

We added DCW, SPA, and DPM-FR adaptations on TMDM across four datasets. Comparisons share checkpoints, test origins, and paired sampling randomness. Three seeds vary inference randomness, not training. ETT/PEMS04 use 32 origins and Electricity 16, with 10 samples per origin.

SSSC uses completed target/forecast spectra and acts after sampling. DCW/DPM-FR adjust frequencies during reverse updates; SPA fits a timestep-dependent spectral prior for sampling guidance. FFT/wavelet operations act along the forecast horizon, with variables as separate channels.

CRPS (lower is better): SPA uses paper DDPM parameters eta=0.01, a=1; DPM-FR uses the public example's high-frequency gain 1.087.

DatasetBaseSSSCDCW 0.1SPADPM-FR
ETTh10.4994690.4994690.8863090.50125736.037375
ETTm20.3243370.3243370.5584680.32731822.824927
Electricity0.5128880.3269730.6276210.512535120.343597
PEMS040.5055080.3001810.5874120.505184120.193502

SSSC reduces CRPS by 36.25%/40.62% on Electricity/PEMS04 and selects identity on ETT. DCW 0.1 increases CRPS by 77.45%, 72.19%, 22.37%, and 16.20%, respectively.

DCW 0.1 is a sensitivity setting; its public example uses 0.005. SPA uses 128 prior-estimation windows, fewer than the image paper. These methods lack official configurations for these forecasting checkpoints. The results characterize the tested adaptations, not universal superiority. All configurations and positive/negative results will be retained in supplementary material. Related methods apply fixed settings while SSSC retains its selector, so tuning budgets are unequal.

Identical Base/SSSC scores mean identity was selected, with no correction benefit. We will label selections and abstention reasons and separate identity from improvements.

3. Hyperparameter sensitivity

On TMDM/ETTh1 and Electricity, three inference seeds cover one-factor scans of gain lower bounds {0.25,0.5,0.75}, upper bounds {1.5,3,6}, protection cutoff fractions {0.125,0.25,0.375}, acceptance thresholds 0–10%, rejection thresholds 0–3×10−43\times10^{-4}, and estimation origins 8–32 (Electricity)/8–64 (ETT).

Electricity results; reduction is 100(1−CRPScorr/CRPSBase)100(1-\mathrm{CRPS}_{corr}/\mathrm{CRPS}_{Base}):

ParameterValuesCRPS reduction
Gain lower bound0.25 / 0.5 / 0.7545.10% / 35.49% / 19.50%
Estimation origins8 / 16 / 3235.57% / 35.52% / 35.49%

Improvement persists in these scans, but its magnitude depends substantially on the lower bound. Origin count has little effect here. We will report full scans and explain flat curves when the selected operation is unchanged. Defaults are retained; test sensitivity results do not select new defaults.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.