acceptodds
Under review as a conference paper at ICLR 2027

FAIP-Diff: Fluctuation-Aware Information-Preserving Diffusion for Time-Series Forecasting

Abstract

Conventional probabilistic forecasting methods often rely on predefined distributional assumptions to characterize multiple possible future outcomes. Diffusion models have become a promising alternative, as they can flexibly learn data distributions without restrictive parametric assumptions. However, existing diffusion-based time-series forecasting methods have not fully solved the problem of temporal pattern distortion that occurs during the diffusion process of non-stationary time series. Moreover, we found that temporal pattern distortion manifests in temporal fluctuation suppression during forward noising and denoising trajectory drift during reverse sampling. To address these questions, we propose FAIP-Diff, a fluctuation-aware information-preserving diffusion framework for probabilistic time series forecasting. FAIP-Diff estimates the conditional mean and volatility as forecasting priors. To alleviate temporal fluctuation suppression during forward noising, FAIP-Diff introduces an adaptive noise remodulation strategy that adjusts the corruption process based on position-dependent forecasting difficulty, enhancing learning in fluctuation-sensitive regions. Second, to mitigate denoising trajectory drift during reverse sampling, FAIP-Diff develops a constrained-space guided sampling mechanism that softly rectifies the predicted noise at each denoising step, steering generated trajectories toward realistic temporal patterns while preserving predictive diversity. Additionally, we introduce a volatility-aware weighted training objective that further alleviates temporal pattern distortion through adaptive loss weighting. Experiments across diverse real-world benchmarks demonstrate that FAIP-Diff delivers more accurate and reliable forecasts, particularly under complex non-stationary dynamics.

Primary area:
learning on time series and dynamical systems
Keywords:
Probabilistic time-series forecasting, diffusion models, non-stationary time series, adaptive noise remodulation, constrained-space sampling.

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_WzNDOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

FAIP-Diff proposes a probabilistic forecasting framework to address temporal pattern distortion during forward and reverse diffusion. The authors introduce three mechanisms: adaptive noise remodulation, which increases forward perturbation according to volatility-normalized forecasting residuals; constrained-space guided sampling, which modifies predicted denoising noise using a PCA subspace constructed from training future windows; and a volatility-aware weighted objective, which emphasizes high-residual positions during volatility estimation.

Critical strengths

1. Empirical gains are good compared to diffusion baselines (Table 1).

2. Useful component-wise and targeted empirical analyses (Table 2).

Critical weaknesses

1. The adaptive forward diffusion process is not sufficiently justified, raising concerns about the soundness of the proposed formulation.

Firstly, the explanation in lines 257–269 (Section 4.2) is unclear. Equation (8) introduces the base noise variance, (σbase,k2)(\sigma_{\mathrm{base},k}^2), without properly motivating its formulation. Although the authors briefly mention (σy2)(\sigma_y^2) as the true local volatility, they do not clearly explain how it is calculated or why (σx2)(\sigma_x^2) and (σy2)(\sigma_y^2) are combined using these specific coefficients. While Appendix A.1 provides an algebraic expansion of the coefficients, it does not sufficiently explain the reasoning behind this particular variance formulation.

Furthermore, Equation (11) modifies the forward diffusion process by adding noise based on the ground-truth prediction error. However, Section 4.3 describes the reverse denoising process without clearly explaining how it accounts for this modified noise. It is unclear whether the reverse process is properly aligned with the adaptive forward diffusion process introduced in Section 4.2. Since the forward and reverse processes are closely related in diffusion models, the authors should provide a clearer explanation of how these two processes work together and justify the proposed formulation. It is very difficult to follow the paper in current form

2. The PCA-based guidance in Section 4.3 is not clearly justified.

In Section 4.3, the authors use PCA to obtain the main patterns from the ground-truth future sequences in the training data. These patterns are then used to correct the predicted noise during reverse diffusion. However, it is unclear why the PCA components obtained from the future sequences are suitable for correcting the predicted noise. The authors do not clearly explain the relationship between these two. Also, PCA mainly captures the dominant patterns in the training data, so it may miss some rare but important fluctuations. Since the main goal of the paper is to preserve temporal patterns, this raises a concern about whether the proposed guidance can preserve such fluctuations. Although the results show improvements, the paper does not clearly demonstrate how the PCA-based correction helps reduce denoising trajectory drift.

Questions for authors

Q1. In Equation (8), could the authors provide the exact calculation of (σy2)(\sigma_y^2) and further justify why the predicted and target variances are combined using these particular coefficients?

Q2. How is the reverse denoising process mathematically consistent with the adaptive forward diffusion process in Equation (11)?

Q3. Although the results show improvements, could the authors further explain why increasing noise in regions with larger prediction errors helps preserve temporal information instead of suppressing it further?

Q4. In Section 4.3, what justifies using PCA components obtained from clean future sequences to correct predicted noise, given that these represent different quantities?

Q5. Could the authors provide a direct quantitative analysis showing that the proposed PCA-based guidance reduces denoising trajectory drift, beyond improvements in general forecasting metrics?

Q6. Could the authors clarify whether all baseline methods were evaluated using the same preprocessing, number of generated samples, and evaluation settings, and explain how fair comparisons were ensured?

8 comments
AuthorsOpenReview ↗

Response to Weakness 1 and Questions 1–2

Weakness 1a / Q1. The calculation of the target variance is unclear, and Equation (8) does not sufficiently justify why the predicted and target variances are combined using the stated coefficients.

Weakness 1b / Q2. It is unclear whether the reverse denoising process is mathematically consistent with the adaptive forward process in Equation (11).

Response. We thank the reviewer for identifying that the manuscript omitted important derivation details.

Weakness 1a and Q1

1. Exact calculation of the target local variance

Let the historical input and observed training future be concatenated as z=[x1:L,y1:H]z=[x_{1:L},y_{1:H}], where LL and HH denote the input and forecasting lengths. For each future position ii and variable nn, the target local variance is computed using a trailing window of length WW. Specifically, the local mean is μi,n(W)=1W∑j=i−Wi−1zj,n\mu_{i,n}^{(W)}=\frac{1}{W}\sum_{j=i-W}^{i-1}z_{j,n}, and the corresponding local variance is σy,i,n2=1W∑j=i−Wi−1(zj,n−μi,n(W))2+ε\sigma_{y,i,n}^{2}=\frac{1}{W}\sum_{j=i-W}^{i-1}(z_{j,n}-\mu_{i,n}^{(W)})^2+\varepsilon.

The implementation uses the population variance (unbiased=False). The first forecasting position is estimated from the most recent historical observations, while later positions additionally use preceding ground-truth future values. Consequently, σy2\sigma_y^2 is available only during training and is never computed from the unknown test future at inference.

2. Derivation of the base forward-noise variance

At diffusion step kk, the clean future is attenuated by αˉk\sqrt{\bar\alpha_k}, so the total perturbation mass is βˉk=1−αˉk\bar\beta_k=1-\bar\alpha_k. The base conditional process decomposes this perturbation into two non-negative components, giving σbase,k2=(βˉk−β~k)σx2+β~kσy2\sigma_{\mathrm{base},k}^{2}=(\bar\beta_k-\tilde\beta_k)\sigma_x^2+\tilde\beta_k\sigma_y^2.

Here, β~k\tilde\beta_k is determined entirely by the diffusion schedule, and the implementation verifies that 0≤β~k≤βˉk0\leq\tilde\beta_k\leq\bar\beta_k. Therefore, both coefficients are non-negative and satisfy (βˉk−β~k)+β~k=βˉk(\bar\beta_k-\tilde\beta_k)+\tilde\beta_k=\bar\beta_k.

The formula is consequently a schedule-controlled variance decomposition rather than an arbitrary interpolation. The term (βˉk−β~k)σx2(\bar\beta_k-\tilde\beta_k)\sigma_x^2 represents the portion supported by the history-conditioned variance prior, whereas β~kσy2\tilde\beta_k\sigma_y^2 injects the realized local variability available during training. Importantly, the coefficients preserve the total perturbation coefficient βˉk\bar\beta_k. If σx2=σy2=σ2\sigma_x^2=\sigma_y^2=\sigma^2, the expression reduces to σbase,k2=βˉkσ2\sigma_{\mathrm{base},k}^{2}=\bar\beta_k\sigma^2, which is the corresponding homoscedastic conditional-diffusion form.

AuthorsOpenReview ↗

Response to Weakness 1 and Questions 1–2(Ⅱ)

Weakness 1b and Q2

Training-time corruption and objective. During training, the observed future y0y_0 is available. Its residual relative to the conditional mean μx\mu_x determines the score ss and the elementwise modulation factor h=1+λsh=1+\lambda s. The forward noise variance is therefore σadapt,k2=h⊙σbase,k2\sigma_{\mathrm{adapt},k}^{2}=h\odot\sigma_{\mathrm{base},k}^{2}.

The denoiser receives states constructed using this adaptive variance and is trained to predict the injected standard Gaussian variable ϵ\epsilon, together with the reverse-transition variance. Its noise-prediction target is ϵ\epsilon, rather than the scaled perturbation σadapt,k2⊙ϵ\sqrt{\sigma_{\mathrm{adapt},k}^{2}}\odot\epsilon. Thus, residual-aware modulation affects both the training states and the learned denoising function.

Inference-time calculation. At inference, y0y_0 is unknown. Therefore, the true residual-dependent factor hh cannot be evaluated. The sampler retains the diffusion schedule, conditional priors, learned noise prediction ϵ^k\hat\epsilon_k, and predicted reverse-transition variance. Using the inferred base-process noise variance σ^ϵ,k2\hat\sigma_{\epsilon,k}^{2}, it calculates the clean-future estimate as y~0(k)=yk−(1−αˉk)μx−σ^ϵ,k2⊙ϵ^kαˉk\tilde y_0^{(k)}=\frac{y_k-(1-\sqrt{\bar\alpha_k})\mu_x-\sqrt{\hat\sigma_{\epsilon,k}^{2}}\odot\hat\epsilon_k}{\sqrt{\bar\alpha_k}}.

We therefore do not claim that this update is the exact analytical reverse of the residual-modulated forward process: the unavailable factor hh is not reinstated in the clean-state reconstruction. Nor do we assume that its effect is necessarily negligible. Eqs. (14)–(15) describe an inference-time learned conditional sampling calculation, whose practical sensitivity to the choice of factor must be assessed empirically.

Scope of related precedents. Previous diffusion studies illustrate that sampling design can be considered separately from the formulas used to construct training states. DDIM derives non-Markovian processes that share a noise-prediction training objective with DDPM [1]. DPM-Solver samples from pretrained diffusion models using a dedicated diffusion-ODE solver [2]. EDM studies sampling choices separately, including their application to pretrained score networks [3]. Each of these methods has its own sampling rationale. We cite them as precedents for separating training and sampling design, rather than as proof that omitting our residual-dependent factor hh yields an exact reverse process. We evaluate this specific design choice through the following paired ablation.

Paired sampling-factor ablation

We compared the current sampler (h=1h=1) with a training-set mean hh and a history-predicted hh, changing only the factor used in clean-state reconstruction. The checkpoint, test windows, random numbers, and CSG setting were fixed. The h=1h=1 implementation reproduced the original sampler with a maximum output difference below 10−510^{-5}. For each dataset, we used three training seeds and generated 100 trajectories per test window. The table reports the three-seed mean percentage change relative to the current sampler; negative values indicate lower and therefore better CRPS or MAE.

DatasetTest windows per seedTraining-mean hh: CRPSTraining-mean hh: MAEHistory-predicted hh: CRPSHistory-predicted hh: MAE
ETTh111+0.033%+0.056%+0.033%+0.054%
ETTh211+0.078%+0.099%+0.080%+0.101%
ETTm144+0.326%+0.157%+0.262%+0.123%
ETTm244+0.267%+0.358%+0.282%+0.377%
ExchangeRate6−0.013%−0.017%−0.015%−0.020%
ILI1−0.085%−0.082%−0.087%−0.085%

Neither implementable factor approximation provides a consistent improvement, and every reported mean change has a magnitude below 0.4%0.4\%. Under the tested checkpoints, windows, and metrics, omitting an explicit residual factor from the sampler has a limited observed effect on final forecasting performance relative to the two feasible proxy-factor samplers. These results support the narrower practical claim that the present non-exact but learned conditional sampling procedure is feasible in the evaluated settings.

References

[1] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising Diffusion Implicit Models. In International Conference on Learning Representations, pp. 1–20, 2021.

[2] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. DPM-Solver: A Fast ODE Solver for Diffusion Probabilistic Model Sampling in Around 10 Steps. In Advances in Neural Information Processing Systems, volume 35, pp. 5775–5787, 2022.

[3] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the Design Space of Diffusion-Based Generative Models. In Advances in Neural Information Processing Systems, volume 35, pp. 26565–26577, 2022.

AuthorsOpenReview ↗

Response to Weakness 2 and Questions 4–5

Weakness 2. In Section 4.3, the authors use PCA to obtain the main patterns from the ground-truth future sequences in the training data. These patterns are then used to correct the predicted noise during reverse diffusion. However, it is unclear why the PCA components obtained from the future sequences are suitable for correcting the predicted noise. The authors do not clearly explain the relationship between these two. Also, PCA mainly captures the dominant patterns in the training data, so it may miss some rare but important fluctuations. Since the main goal of the paper is to preserve temporal patterns, this raises a concern about whether the proposed guidance can preserve such fluctuations. Although the results show improvements, the paper does not clearly demonstrate how the PCA-based correction helps reduce denoising trajectory drift.

Q4. In Section 4.3, what justifies using PCA components obtained from clean future sequences to correct predicted noise, given that these represent different quantities?

Q5. Could the authors provide a direct quantitative analysis showing that the proposed PCA-based guidance reduces denoising trajectory drift, beyond improvements in general forecasting metrics?

Response to Weakness 2a and Q4

Thank you for pointing out that the connection between the training-derived PCA space and predicted-noise correction was insufficiently explained. CSG is designed to mitigate denoising trajectory drift, rather than to use the principal components as direct predictions of the future.

Multivariate future windows are high-dimensional, but their dominant trends, periodic patterns, scale changes, and cross-variable dependencies are often concentrated in a lower-dimensional correlated structure. The PCA basis B\mathbf B extracted from the training futures therefore provides a data-driven coordinate system: its span S\mathcal S represents directions supported by the dominant training structure, while directions orthogonal to S\mathcal S represent deviations from that structure.

This structural coordinate system can guide noise because the predicted noise ϵ^k\hat{\boldsymbol\epsilon}_k determines the direction of each reverse update and therefore affects the shape, scale, and cross-variable relationships of the generated trajectory. CSG does not assume that the noise itself follows the future distribution. Instead, it uses S\mathcal S to identify and attenuate the component of the predicted noise that drives the reverse trajectory away from the training-supported structure. Let P=B⊤B\mathbf P=\mathbf B^{\top}\mathbf B denote the projection onto S\mathcal S. The corrected noise satisfies ∥(I−P)ϵk′∥2=∣1−2ρ∣∥(I−P)ϵ^k∥2.\left\|(\mathbf I-\mathbf P)\boldsymbol\epsilon'_k\right\|_2=|1-2\rho|\left\|(\mathbf I-\mathbf P)\hat{\boldsymbol\epsilon}_k\right\|_2.

With ρ=0.6\rho=0.6, the orthogonal component is reversed and reduced to 20% of its original magnitude. Because ϵk′\epsilon_{k}^{\prime} is used to calculate the next state yk−1\mathbf{y}_{k-1}, this correction limits the accumulation of off-subspace deviations over subsequent reverse steps. In short, PCA supplies the structural reference, while noise correction acts on the direction that actually drives reverse sampling.

Response to Weakness 2b

Preserving all such details is not the specific objective of CSG. The role of CSG is limited to mitigating denoising trajectory drift during reverse sampling by reducing directions that are weakly supported by the dominant training-future structure.

Response to Weakness 2c and Q5

We added a matched-sample structural diagnostic comparing the same trained model with CSG disabled and enabled. The initial and stepwise sampling random numbers are paired. Each dataset uses 30 test windows and 50 predictive samples per window. Lower values indicate smaller structural deviation.

Each cell reports without CSG -> with CSG (reduction).

DatasetOut-of-subspace errorScale discrepancyACF errorSpectral errorCross-variable correlation error
ETTh20.433409→0.420508 (2.977%)0.438193→0.407700 (6.959%)0.268660→0.256018 (4.706%)0.011426→0.010904 (4.567%)0.353027→0.306317 (13.231%)
ETTm10.784229→0.779955 (0.545%)0.796002→0.791456 (0.571%)0.099580→0.099348 (0.232%)0.006224→0.006207 (0.269%)0.526262→0.522342 (0.745%)
Weather0.235438→0.233135 (0.978%)1.641348→1.632744 (0.524%)0.391509→0.385824 (1.452%)0.095826→0.094987 (0.875%)0.762517→0.757847 (0.612%)

The largest reductions are observed on ETTh2, where trajectory-scale discrepancy, autocorrelation error, spectral-shape error, and cross-variable correlation error decrease by 6.96%, 4.71%, 4.57%, and 13.23%, respectively. These results quantitatively support that CSG can reduce structural deviation in the resulting forecast trajectories.

Reviewer_F1fqOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes FAIP-Diff for probabilistic time series forecasting under non-stationary fluctuations. Building on conditional mean and variance predictors, it introduces three components: residual-dependent modulation of the forward noise variance, a sampling correction based on a PCA subspace fitted to training future windows, and a volatility loss that assigns larger weights to high-residual positions. The PCA correction modifies predicted noise through a blend with its reflection about the learned subspace. Experiments compare the full model with forecasting baselines and component ablations; the appendix adds horizon sensitivity, high-volatility analyses, and matched-seed statistical comparisons with NsDiff.

Critical strengths
  1. The evaluation is comprehensive and the results are good.
  2. The method is widely applicable, including deterministic and probabilistic forecasting using diffusion model.
Critical weaknesses

1 Equation (9) multiplies forward variance by 1+λs1+\lambda s, while Eqs. (14)–(15) reconstruct the clean signal using a variance expression without this factor. Appendix Eq. (70) supervises the standard Gaussian noise used in the adaptively perturbed state. If both the noise and variance heads have the unmodulated meanings suggested by these equations, the reconstruction would not generally invert Eq. (11).

Questions for authors
  1. What are the actual training schedule, terminal αˉK\bar\alpha_K, and sampling timesteps? If the 20 reported steps are respaced from a longer process, please provide that schedule and the mapping used by the sampler. How is the resulting terminal training distribution matched to Eq. (12)?
  2. Is the correction intended primarily as a heuristic regularizer, or is the noise-space contraction claimed to imply improved signal reconstruction? If the latter, what additional assumption is needed? An existing-sampler diagnostic of clean-prediction error before and after correction would help clarify this interpretation.
6 comments
AuthorsOpenReview ↗

Response to Weakness and Question 1

Weakness. Equation (9) multiplies the forward variance by 1+λs1+\lambda s, while Eqs. (14)–(15) reconstruct the clean signal using a variance expression without this factor. Appendix Eq. (70) supervises the standard Gaussian noise used in the adaptively perturbed state. If both the noise and variance heads have the unmodulated meanings suggested by these equations, the reconstruction would not generally invert Eq. (11).

Question 1. What are the actual training schedule, terminal αˉK\bar\alpha_K, and sampling timesteps? If the 20 reported steps are respaced from a longer process, please provide that schedule and the mapping used by the sampler. How is the resulting terminal training distribution matched to Eq. (12)?

Response. Thank you for pointing out that the manuscript does not adequately distinguish among the adaptive training corruption, the learned training objective, and the implementable inference sampler. We address the variance meanings, schedule, and terminal initialization separately below.

1. What is learned under the adaptive forward process?

Let y0y_0 be the observed future in a training pair, let μx\mu_x and σx2\sigma_x^2 be the conditional mean and variance predicted from the historical input, and let σy,k2\sigma_{y,k}^2 denote the future local variance estimated from the training sequence. At diffusion step kk, the base forward variance is σbase,k2=(βˉk−β~k)σx2+β~kσy,k2\sigma_{\mathrm{base},k}^{2}=(\bar\beta_k-\tilde\beta_k)\sigma_x^2+\tilde\beta_k\sigma_{y,k}^{2}. The residual-derived score and modulation factor are defined as s=sigmoid⁡(∣y0−μx∣σx2+ε−1)s=\operatorname{sigmoid}\left(\frac{|y_0-\mu_x|}{\sqrt{\sigma_x^2}+\varepsilon}-1\right) and h=1+λsh=1+\lambda s, respectively. The adaptive forward variance is then calculated as σadapt,k2=h⊙σbase,k2\sigma_{\mathrm{adapt},k}^{2}=h\odot\sigma_{\mathrm{base},k}^{2}.

The actual noisy training state is constructed as yk=αˉk y0+(1−αˉk)μx+σadapt,k2⊙ϵy_k=\sqrt{\bar\alpha_k}\,y_0+(1-\sqrt{\bar\alpha_k})\mu_x+\sqrt{\sigma_{\mathrm{adapt},k}^{2}}\odot\epsilon, where ϵ∼N(0,I)\epsilon\sim\mathcal N(0,I).

The denoising network receives this adaptively perturbed state and produces (ϵ^k,σθ,k2)=Dθ(yk,μx,σx2,k)(\hat\epsilon_k,\sigma_{\theta,k}^{2})=D_\theta(y_k,\mu_x,\sigma_x^2,k). Here, ϵ^k\hat\epsilon_k predicts the sampled standard Gaussian variable ϵ\epsilon, rather than the already scaled perturbation σadapt,k2⊙ϵ\sqrt{\sigma_{\mathrm{adapt},k}^{2}}\odot\epsilon. Therefore, using ϵ\epsilon as the noise-prediction target in Appendix Eq. (70) remains meaningful even though the input state is adaptively perturbed.

The variance head directly outputs σθ,k2\sigma_{\theta,k}^{2}, which is the predicted reverse-transition variance. It does not directly output σ^ϵ,k2\hat\sigma_{\epsilon,k}^{2}, the forward-noise variance estimate used during inference. Instead, σ^ϵ,k2\hat\sigma_{\epsilon,k}^{2} is calculated from σx2\sigma_x^2, the diffusion-schedule coefficients, and an inferred future variance.

The implemented training objective consists of three components and is defined as L=Lmean+Lvar+Ldiff\mathcal L=\mathcal L_{\mathrm{mean}}+\mathcal L_{\mathrm{var}}+\mathcal L_{\mathrm{diff}}. The mean-prediction loss is Lmean=Mean⁡[(μx−y0)2]\mathcal L_{\mathrm{mean}}=\operatorname{Mean}\left[(\mu_x-y_0)^2\right], while the variance loss is Lvar=Mean⁡[w⊙(σx2−σy,k2)2]\mathcal L_{\mathrm{var}}=\operatorname{Mean}\left[w\odot\left(\sqrt{\sigma_x^2}-\sqrt{\sigma_{y,k}^2}\right)^2\right]. The diffusion loss is defined as Ldiff=Mean⁡[(ϵ−ϵ^k)2+σy,k2σθ,k2−log⁡(σy,k2σθ,k2)]\mathcal L_{\mathrm{diff}}=\operatorname{Mean}\left[(\epsilon-\hat\epsilon_k)^2+\frac{\sigma^2_{y,k}}{\sigma_{\theta,k}^2}-\log\left(\frac{\sigma^2_{y,k}}{\sigma_{\theta,k}^2}\right)\right].

In these expressions, ww is the clipped residual-based weight, and σy,k2\sigma^2_{y,k} is the target reverse-posterior variance derived from the base conditional process. The noise-prediction term is trained on adaptively perturbed states, but the variance-matching term does not incorporate hh. Accordingly, Ldiff\mathcal L_{\mathrm{diff}} is a KL-inspired surrogate objective, rather than the exact Gaussian reverse KL divergence for the residual-modulated forward process. As stated in Eq. (70) of the paper, “Inspired by the Gaussian reverse KL divergence, we adopt the following simplified surrogate objective.”

AuthorsOpenReview ↗

Response to Weakness and Question 1(Ⅱ)

2. Relationship between adaptive training corruption and inference sampling

Training-time corruption and objective. During training, the observed future y0y_0 is available. Its residual relative to the conditional mean μx\mu_x determines the score ss and the elementwise modulation factor h=1+λsh=1+\lambda s. The forward noise variance is therefore σadapt,k2=h⊙σbase,k2\sigma_{\mathrm{adapt},k}^{2}=h\odot\sigma_{\mathrm{base},k}^{2}.

The denoiser receives states constructed using this adaptive variance and is trained to predict the injected standard Gaussian variable ϵ\epsilon, together with the reverse-transition variance. Its noise-prediction target is ϵ\epsilon, rather than the scaled perturbation σadapt,k2⊙ϵ\sqrt{\sigma_{\mathrm{adapt},k}^{2}}\odot\epsilon. Thus, residual-aware modulation affects both the training states and the learned denoising function.

Inference-time calculation. At inference, y0y_0 is unknown. Therefore, the true residual-dependent factor hh cannot be evaluated. The sampler retains the diffusion schedule, conditional priors, learned noise prediction ϵ^k\hat\epsilon_k, and predicted reverse-transition variance. Using the inferred base-process noise variance σ^ϵ,k2\hat\sigma_{\epsilon,k}^{2}, it calculates the clean-future estimate as y~0(k)=yk−(1−αˉk)μx−σ^ϵ,k2⊙ϵ^kαˉk\tilde y_0^{(k)}=\frac{y_k-(1-\sqrt{\bar\alpha_k})\mu_x-\sqrt{\hat\sigma_{\epsilon,k}^{2}}\odot\hat\epsilon_k}{\sqrt{\bar\alpha_k}}.

We therefore do not claim that this update is the exact analytical reverse of the residual-modulated forward process: the unavailable factor hh is not reinstated in the clean-state reconstruction. Nor do we assume that its effect is necessarily negligible. Eqs. (14)–(15) describe an inference-time learned conditional sampling calculation, whose practical sensitivity to the choice of factor must be assessed empirically.

Scope of related precedents. Previous diffusion studies illustrate that sampling design can be considered separately from the formulas used to construct training states. DDIM derives non-Markovian processes that share a noise-prediction training objective with DDPM [1]. DPM-Solver samples from pretrained diffusion models using a dedicated diffusion-ODE solver [2]. EDM studies sampling choices separately, including their application to pretrained score networks [3]. Each of these methods has its own sampling rationale. We cite them as precedents for separating training and sampling design, rather than as proof that omitting our residual-dependent factor hh yields an exact reverse process. We evaluate this specific design choice through the following paired ablation.

Paired sampling-factor ablation

We compared the current sampler (h=1h=1) with a training-set mean hh and a history-predicted hh, changing only the factor used in clean-state reconstruction. The checkpoint, test windows, random numbers, and CSG setting were fixed. The h=1h=1 implementation reproduced the original sampler with a maximum output difference below 10−510^{-5}. For each dataset, we used three training seeds and generated 100 trajectories per test window. The table reports the three-seed mean percentage change relative to the current sampler; negative values indicate lower and therefore better CRPS or MAE.

DatasetTest windows per seedTraining-mean hh: CRPSTraining-mean hh: MAEHistory-predicted hh: CRPSHistory-predicted hh: MAE
ETTh111+0.033%+0.056%+0.033%+0.054%
ETTh211+0.078%+0.099%+0.080%+0.101%
ETTm144+0.326%+0.157%+0.262%+0.123%
ETTm244+0.267%+0.358%+0.282%+0.377%
ExchangeRate6−0.013%−0.017%−0.015%−0.020%
ILI1−0.085%−0.082%−0.087%−0.085%

Neither implementable factor approximation provides a consistent improvement, and every reported mean change has a magnitude below 0.4%0.4\%. Under the tested checkpoints, windows, and metrics, omitting an explicit residual factor from the sampler has a limited observed effect on final forecasting performance relative to the two feasible proxy-factor samplers. These results support the narrower practical claim that the present non-exact but learned conditional sampling procedure is feasible in the evaluated settings.

References

[1] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising Diffusion Implicit Models. In International Conference on Learning Representations, pp. 1–20, 2021.

[2] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. DPM-Solver: A Fast ODE Solver for Diffusion Probabilistic Model Sampling in Around 10 Steps. In Advances in Neural Information Processing Systems, volume 35, pp. 5775–5787, 2022.

[3] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the Design Space of Diffusion-Based Generative Models. In Advances in Neural Information Processing Systems, volume 35, pp. 26565–26577, 2022.

AuthorsOpenReview ↗

Response to Weakness and Question 1(Ⅲ)

3. Actual training schedule and sampler mapping

The reported 20 steps constitute the complete training diffusion schedule; they are not respaced from a longer process. The schedule is linear and is defined by K=20K=20, βk=10−4+k−119(10−2−10−4)\beta_k=10^{-4}+\frac{k-1}{19}(10^{-2}-10^{-4}) for k=1,…,20k=1,\ldots,20, and αk=1−βk\alpha_k=1-\beta_k.

Consequently, the terminal cumulative signal coefficient is αˉK=∏k=120(1−βk)≈0.90361935\bar\alpha_K=\prod_{k=1}^{20}(1-\beta_k)\approx0.90361935, with 1−αˉK≈0.096380651-\bar\alpha_K\approx0.09638065.

The implementation uses zero-based indices t=0,…,19t=0,\ldots,19, corresponding to the manuscript indices k=t+1k=t+1. During training, it samples these indices to construct noisy states. During inference, it initializes the terminal state and applies the reverse sampler at t=19,18,…,1t=19,18,\ldots,1, followed by the final update at t=0t=0. Thus, no longer-process schedule, respacing rule, or nontrivial timestep mapping is used.

4. Terminal-distribution matching

Equation (12) defines the conditional prior used to initialize reverse sampling during validation and testing. It is not used to generate noisy training states and is not an identity derived from the adaptive terminal distribution used during training.

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

FAIP-Diff aims to improve probabilistic time-series forecasting, particularly around sharp changes and peryoniods of high volatility. Building on a diffusion model with predicted mean and volatility, it introduces three components. Adaptive Noise Remodulation increases training noise at positions where the future is difficult to predict. Constrained Space Guided Sampling adjusts predicted noise using a PCA subspace constructed from training sequences. A volatility-aware loss gives more weight to positions with large prediction residuals. The paper evaluates the method on eight forecasting datasets using accuracy and calibration metrics. It also includes component ablations, sensitivity studies, analyses of volatile and shifted windows, and comparisons with NsDiff across matched random seeds.

Critical strengths

The paper addresses a meaningful forecasting problem. Sharp changes and local volatility are important in many applications, and average forecast errors can hide poor performance around these events. The paper makes a useful effort to study them explicitly.

The method contains interesting and practical ideas. Residual-dependent noise modulation and training-derived subspace guidance are relatively simple additions to an existing forecasting model. They provide concrete ways to modify both learning and sampling without requiring a substantially larger backbone.

Critical weaknesses

For ANR, the noise variance increases even when the prediction residual is zero. It is therefore unclear how much of the gain comes from targeting difficult positions and how much comes from increasing noise overall. A uniform-noise baseline with the same average variance, together with shuffled residual scores, would help answer this. For the sampling module, the qualitative random-subspace comparison is useful, but quantitative comparisons with random subspaces and ordinary projection are still needed.

The training and inference equations need clarification. ANR changes the forward noise variance using observed future residuals, but its consequences for the reverse-process equations and training objective are not clearly traced. The computation of σy,k2\sigma_{y,k}^{2} in Eq. 14 also remains unclear.

Questions for authors

Does ANR benefit from adapting noise to difficult positions, or simply from adding more noise? Could you compare it with a uniform increase matched to the same average noise variance, and with a version that shuffles the residual-based scores across positions? These controls would help isolate the contribution of the adaptive rule.

What does the proposed subspace correction gain over alternative methods? A global PCA subspace may not preserve temporal structure and may also underrepresent unusual patterns. Please provide quantitative evidence shows that it improves the temporal structure of generated forecasts.

The evaluation does not directly establish preservation of fluctuations or temporal structure. CRPS, QICE, MAE, and MSE provide useful evidence of forecasting quality, but they do not sufficiently test the paper’s central claims about peak recovery and trajectory structure. A model can improve these aggregate scores while still predicting peaks at the wrong time or failing to reproduce dependencies across time and variables.

7 comments
AuthorsOpenReview ↗

Response to Weakness 1 and Question 1–2

Weakness 1. ANR increases noise variance even at zero residual, leaving the contributions of residual-based adaptation and overall noise inflation unclear. Mean-matched uniform-noise and shuffled-score controls are needed, along with quantitative sampling comparisons against random subspaces and ordinary PCA projection.

Question 1. Does ANR benefit from targeting difficult positions rather than simply increasing noise? Please evaluate mean-matched uniform-noise and shuffled-residual controls.

Question 2. What does the proposed subspace correction gain over alternative methods? Please quantify its effect on temporal structure and address whether global PCA underrepresents unusual patterns.

Response to Weakness 1a and Question 1

Thank you for suggesting these controls. We added three controls: temporally uniform variance, a uniform multiplicative gain on NsDiff’s time-varying variance, and shuffled residual scores across forecast positions. Each control matches ANR’s mean variance over forecast positions for each sample and variable; shuffled-score variances are renormalized to account for the nonuniform base variance. All variants share the same architecture, loss definitions, and sampling procedure.

MethodCRPS↓QICE(%)↓MAE↓MSE↓RMSE↓
Uniform variance, mean-matched0.2433 ± 0.02904.9606 ± 1.22190.3075 ± 0.03720.1723 ± 0.03690.4135 ± 0.0451
Uniform gain on NsDiff variance, mean-matched0.2742 ± 0.03195.6383 ± 0.87370.3433 ± 0.04310.2279 ± 0.05310.4751 ± 0.0570
Shuffled residual scores, mean-matched0.2533 ± 0.02125.2002 ± 1.13480.3219 ± 0.02480.1862 ± 0.02490.4309 ± 0.0285
ANR (ours)0.2202 ± 0.00904.1568 ± 1.02930.2791 ± 0.00860.1414 ± 0.00870.3759 ± 0.0115

ANR achieves the lowest mean CRPS, QICE, MAE, MSE, and RMSE among the four variants. These results provide evidence that, on the evaluated ExchangeRate windows, residual-dependent allocation offers benefits beyond a uniform increase in noise magnitude, and that correspondence between residual scores and temporal positions can contribute to forecasting performance.

Response to Weakness 1b

Thank you for this suggestion. We added quantitative comparisons with no correction, random-subspace reflection, and ordinary PCA projection on ExchangeRate. All methods use the same checkpoints and paired sampling randomness, with three training checkpoints and five sampling seeds per checkpoint.

Sampling methodCRPS↓QICE(%)↓MAE↓MSE↓
No correction0.210444 ± 0.0052023.510273 ± 0.3522540.269543 ± 0.0067100.133252 ± 0.013925
Random-subspace reflection0.210904 ± 0.0056573.528938 ± 0.3625050.270187 ± 0.0069610.133768 ± 0.014444
Ordinary PCA projection0.210440 ± 0.0052073.508733 ± 0.3508980.269548 ± 0.0067100.133250 ± 0.013923
CSG (ours)0.210439 ± 0.0052083.508441 ± 0.3506670.269549 ± 0.0067110.133250 ± 0.013923

As shown in the table, CSG improves all four metrics over random-subspace reflection and achieves the lowest mean CRPS, QICE, and MSE among the four methods. Its performance is close to ordinary PCA projection, with small improvements in these three metrics, while no correction achieves the lowest MAE. These results support the usefulness of CSG in this setting, particularly compared with random-subspace guidance.

Response to Question 2

Thank you for raising this concern. To additionally assess temporal structure, we compared the same trained model with CSG disabled and enabled, using paired initial and stepwise sampling random numbers.

Each cell reports without CSG -> with CSG (reduction).

DatasetOut-of-subspace errorScale discrepancyACF errorSpectral errorCross-variable correlation error
ETTh20.433409→0.420508 (2.977%)0.438193→0.407700 (6.959%)0.268660→0.256018 (4.706%)0.011426→0.010904 (4.567%)0.353027→0.306317 (13.231%)
ETTm10.784229→0.779955 (0.545%)0.796002→0.791456 (0.571%)0.099580→0.099348 (0.232%)0.006224→0.006207 (0.269%)0.526262→0.522342 (0.745%)
Weather0.235438→0.233135 (0.978%)1.641348→1.632744 (0.524%)0.391509→0.385824 (1.452%)0.095826→0.094987 (0.875%)0.762517→0.757847 (0.612%)

The largest reductions are observed on ETTh2, where trajectory-scale discrepancy, autocorrelation error, spectral-shape error, and cross-variable correlation error decrease by 6.96%, 4.71%, 4.57%, and 13.23%, respectively. These results quantitatively support that CSG can reduce structural deviation in the resulting forecast trajectories.

AuthorsOpenReview ↗

Response to Weakness 2

Weakness 2. The training and inference equations need clarification. ANR changes the forward noise variance using observed future residuals, but its consequences for the reverse-process equations and training objective are not clearly traced. The computation of the local target variance in Eq. (14) also remains unclear.

Response: Thank you for pointing out that the manuscript does not adequately distinguish among the adaptive training corruption, the learned training objective, and the implementable inference sampler.

1. Computation of the inference-time future-variance estimate

At reverse step kk, let yky_k be the current noisy state, μx\mu_x the conditional mean predicted from historical observations, and σx2\sigma_x^2 the corresponding conditional variance. The denoising network takes the concatenation of (yk,μx,σx2)(y_k,\mu_x,\sigma_x^2) together with a learned embedding of step kk. Its noise and variance heads produce (ϵ^k,σθ,k2)=Dθ(yk,μx,σx2,k)(\hat\epsilon_k,\sigma_{\theta,k}^2)=D_\theta(y_k,\mu_x,\sigma_x^2,k). The variance head uses a Softplus output, so σθ,k2\sigma_{\theta,k}^2 is the predicted reverse-transition variance.

The unknown future variance used in the reverse calculation, denoted by σy,k2\sigma^2_{y,k}, is then inferred from σθ,k2\sigma_{\theta,k}^2 using the posterior-variance relationship of the base conditional process. To show the computation, we suppress the future-time and variable indices and define a=αka=\alpha_k, b=1−αkb=1-\alpha_k, A=β~k−1A=\tilde\beta_{k-1}, and B=βˉk−1−β~k−1B=\bar\beta_{k-1}-\tilde\beta_{k-1}.

Setting v=σy,k2v=\sigma^2_{y,k} gives the quadratic equation λ0,kv2+λ1,kv+λ2,k=0\lambda_{0,k}v^2+\lambda_{1,k}v+\lambda_{2,k}=0, where λ0,k=abA\lambda_{0,k}=abA, λ1,k=(b2A+abB)σx2−σθ,k2(aA+ab)\lambda_{1,k}=(b^2A+abB)\sigma_x^2-\sigma_{\theta,k}^2(aA+ab), and λ2,k=b2B(σx2)2−σθ,k2σx2(aB+b2)\lambda_{2,k}=b^2B(\sigma_x^2)^2-\sigma_{\theta,k}^2\sigma_x^2(aB+b^2).

The implementation selects the positive root, given by σy,k2=−λ1,k+λ1,k 2−4λ0,kλ2,k2λ0,k\sigma^2_{y,k}=\frac{-\lambda_{1,k}+\sqrt{\lambda_{1,k}^{\,2}-4\lambda_{0,k}\lambda_{2,k}}}{2\lambda_{0,k}}. This stepwise estimate is subsequently substituted into the base noise-variance expression σ^ϵ,k2=(βˉk−β~k)σx2+β~kσy,k2\hat\sigma^2_{\epsilon,k}=(\bar\beta_k-\tilde\beta_k)\sigma_x^2+\tilde\beta_k\sigma^2_{y,k}, which is used in clean-future reconstruction and the reverse update.

Therefore, inference uses the current noisy state and the network-predicted reverse variance to estimate σy,k2\sigma^2_{y,k}; it does not use the observed test future or a precomputed training-set data-distribution variance.

2. Effect of residual adaptation on training

The implemented training objective consists of three components and is defined as L=Lmean+Lvar+Ldiff\mathcal L=\mathcal L_{\mathrm{mean}}+\mathcal L_{\mathrm{var}}+\mathcal L_{\mathrm{diff}}. The mean-prediction loss is Lmean=Mean⁡[(μx−y0)2]\mathcal L_{\mathrm{mean}}=\operatorname{Mean}\left[(\mu_x-y_0)^2\right], while the variance loss is Lvar=Mean⁡[w⊙(σx2−σy,k2)2]\mathcal L_{\mathrm{var}}=\operatorname{Mean}\left[w\odot\left(\sqrt{\sigma_x^2}-\sqrt{\sigma_{y,k}^2}\right)^2\right]. The diffusion loss is defined as Ldiff=Mean⁡[(ϵ−ϵ^k)2+σy,k2σθ,k2−log⁡(σy,k2σθ,k2)]\mathcal L_{\mathrm{diff}}=\operatorname{Mean}\left[(\epsilon-\hat\epsilon_k)^2+\frac{\sigma^2_{y,k}}{\sigma_{\theta,k}^2}-\log\left(\frac{\sigma^2_{y,k}}{\sigma_{\theta,k}^2}\right)\right].

In these expressions, ww is the clipped residual-based weight, and σy,k2\sigma^2_{y,k} is the target reverse-posterior variance derived from the base conditional process. The noise-prediction term is trained on adaptively corrupted states, but the variance-matching term does not incorporate hh. Accordingly, Ldiff\mathcal L_{\mathrm{diff}} is a KL-inspired surrogate objective, rather than the exact Gaussian reverse KL divergence for the residual-modulated forward process. As stated in Eq. (70) of the paper, “Inspired by the Gaussian reverse KL divergence, we adopt the following simplified surrogate objective.”

AuthorsOpenReview ↗

Response to Weakness 2(Ⅱ)

3. Relationship between adaptive training corruption and inference sampling

Training-time corruption and objective. During training, the observed future y0y_0 is available. Its residual relative to the conditional mean μx\mu_x determines the score ss and the elementwise modulation factor h=1+λsh=1+\lambda s. The resulting forward noise variance is σadapt,k2=h⊙σbase,k2\sigma_{\mathrm{adapt},k}^{2}=h\odot\sigma_{\mathrm{base},k}^{2}.

The denoiser receives states constructed using this adaptive variance and is trained to predict the injected standard Gaussian variable ϵ\epsilon, together with the reverse-transition variance. Its noise-prediction target is ϵ\epsilon, rather than the scaled perturbation σadapt,k2⊙ϵ\sqrt{\sigma_{\mathrm{adapt},k}^{2}}\odot\epsilon. Thus, residual-aware modulation affects the training states and the learned denoising function.

Inference-time calculation. At inference, y0y_0 is unknown. Therefore, the true residual-dependent factor hh cannot be evaluated. The sampler retains the diffusion schedule, conditional priors, learned noise prediction ϵ^k\hat\epsilon_k, and predicted reverse-transition variance. Using the inferred base-process noise variance σ^ϵ,k2\hat\sigma_{\epsilon,k}^{2}, it calculates the clean-future estimate as y~0(k)=yk−(1−αˉk)μx−σ^ϵ,k2⊙ϵ^kαˉk\tilde y_0^{(k)}=\frac{y_k-(1-\sqrt{\bar\alpha_k})\mu_x-\sqrt{\hat\sigma_{\epsilon,k}^{2}}\odot\hat\epsilon_k}{\sqrt{\bar\alpha_k}}.

We therefore do not claim that this update is the exact analytical reverse of the residual-modulated forward process: the unavailable factor hh is not reinstated in the clean-state reconstruction. Nor do we assume that its effect is necessarily negligible. Eqs. (14)–(15) describe an inference-time learned conditional sampling calculation, whose practical sensitivity to the choice of factor must be assessed empirically.

Scope of related precedents. Previous diffusion studies illustrate that sampling design can be considered separately from the formulas used to construct training states. DDIM derives non-Markovian processes that share a noise-prediction training objective with DDPM [1]. DPM-Solver samples from pretrained diffusion models using a dedicated diffusion-ODE solver [2]. EDM studies sampling choices separately, including their application to pretrained score networks [3]. Each of these methods has its own sampling rationale. We cite them as precedents for separating training and sampling design, rather than as proof that omitting our residual-dependent factor hh yields an exact reverse process. We evaluate this specific design choice through the following paired ablation.

Paired sampling-factor ablation

We compared the current sampler (h=1h=1) with a training-set mean hh and a history-predicted hh, changing only the factor used in clean-state reconstruction. The checkpoint, test windows, random numbers, and CSG setting were fixed. The h=1h=1 implementation reproduced the original sampler with a maximum output difference below 10−510^{-5}. For each dataset, we used three training seeds and generated 100 trajectories per test window. The table reports the three-seed mean percentage change relative to the current sampler; negative values indicate lower and therefore better CRPS or MAE.

DatasetTest windows per seedTraining-mean hh: CRPSTraining-mean hh: MAEHistory-predicted hh: CRPSHistory-predicted hh: MAE
ETTh111+0.033%+0.056%+0.033%+0.054%
ETTh211+0.078%+0.099%+0.080%+0.101%
ETTm144+0.326%+0.157%+0.262%+0.123%
ETTm244+0.267%+0.358%+0.282%+0.377%
ExchangeRate6−0.013%−0.017%−0.015%−0.020%
ILI1−0.085%−0.082%−0.087%−0.085%

Neither implementable factor approximation provides a consistent improvement, and every reported mean change has a magnitude below 0.4%0.4\%. Under the tested checkpoints, windows, and metrics, omitting an explicit residual factor from the sampler has a limited observed effect on final forecasting performance relative to the two feasible proxy-factor samplers. These results support the narrower practical claim that the present non-exact but learned conditional sampling procedure is feasible in the evaluated settings.

References

[1] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising Diffusion Implicit Models. In International Conference on Learning Representations, pp. 1–20, 2021.

[2] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. DPM-Solver: A Fast ODE Solver for Diffusion Probabilistic Model Sampling in Around 10 Steps. In Advances in Neural Information Processing Systems, volume 35, pp. 5775–5787, 2022.

[3] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the Design Space of Diffusion-Based Generative Models. In Advances in Neural Information Processing Systems, volume 35, pp. 26565–26577, 2022.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to Reviewers, ACs, SACs, and PCs

We thank the reviewers, ACs, SACs, and PCs for their careful consideration. We summarize below the main clarifications and additional experiments. The revised appendix provides the detailed derivations, protocols, and results.

  • ANR mechanism (Reviewers WzND, DfKp). We clarify that ANR applies bounded residual-dependent modulation to the time-varying NsDiff noise variance. Mean-matched uniform-noise and shuffled-residual controls assess whether improvements arise from overall noise magnitude or its allocation across positions. The results support benefits from residual-dependent allocation in several settings, while showing dataset dependence. We have added the mechanism analysis, experimental setup, and complete results to Appendix A.1 and Appendix B.1.

  • Training and sampling (Reviewers WzND, DfKp, F1fq). We clarify the target local variance, three-part training objective, reverse-variance calculation, diffusion schedule, and terminal initialization. The test-future residual factor is unavailable at inference, so the sampler is a learned conditional procedure, not an exact analytical reverse of the adaptive forward process. Replacing its h=1h=1 factor with either of two inference-feasible proxies produces only small mean changes in CRPS and MAE. We have added the derivations, paired experimental protocol, and full results to Appendix A.2--A.3.

  • CSG mechanism and sampling controls (Reviewers WzND, DfKp, F1fq). We explain how training-derived PCA directions guide predicted noise without assuming that noise and clean futures share a distribution. Fixed-checkpoint comparisons with no correction, random-subspace reflection, and ordinary PCA projection isolate the sampling module's contribution. These controls show dataset- and metric-dependent effects rather than universal superiority. We have added the projection analysis, experimental protocol, and complete results to Appendix A.4 and Appendix B.2, together with step-resolved visualizations of peak and volatility fidelity and clean-estimate error in Appendix B.2.1 (Figures 5 and 6)

  • Framework-level temporal-structure evaluation (Reviewer F1fq). We separately compare the complete FAIP-Diff framework with the original NsDiff baseline on matched test windows, evaluating local volatility, first-difference correlation, peak recovery and timing, and cross-variable dependencies. We have added the metric definitions, evaluation protocol, and full results to Appendix C.

  • Evaluation protocol and limitations (Reviewer WzND). We have added dataset preprocessing, forecast and sample counts, metric definitions, and the limits of small or overlapping test-window sets to Appendix D.

The point-by-point replies provide the complete evidence and the precise scope of each conclusion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.