Pith. sign in

REVIEW 3 major objections 5 minor 75 references

DiffDiff claims that a diffusion forecaster improves by making the forward corruption trajectory predictability-aware, progressively transforming the target toward its second-order differenced structure so that generative effort concentrate

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 11:28 UTC pith:MVJMLBOI

load-bearing objection A simple, general forward-process trick for diffusion forecasting that works well on seven benchmarks, but the claimed mechanism is only partially supported. the 3 major comments →

arxiv 2607.22599 v1 pith:MVJMLBOI submitted 2026-06-12 cs.AI

Differencing the Diffusion Trajectory toward Uncertain Components for Time Series Forecasting

Chen Su , Yuanhe Tian , Yan Song This is my paper
classification cs.AI
keywords probabilistic time series forecastingdiffusion modelsforward process designsecond-order differencingpredictability asymmetryfrequency-selective corruptionnon-isotropic diffusionconditional denoising
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Probabilistic forecasting with diffusion models usually treats every part of the future window as equally hard to generate, but in real series the smooth, slowly-moving part is largely determined by the observed history while the rapidly varying part carries most residual uncertainty. The paper tries to embed this asymmetry into the diffusion trajectory itself: a step-dependent forward operator gradually reshapes the noisy intermediate state from the target toward its second-order differenced structure, attenuating low-frequency (history-anchored) content and preserving high-frequency (under-determined) content. A stage-adaptive denoiser then re-supplies value and difference information from the history when the noisy target no longer carries it, while the endpoint still approaches a standard Gaussian so existing samplers apply. If correct, the mechanism that carries the improvement is the reshaping of the corruption path, not the conditioning network, and the diffusion's capacity is spent on the components the history cannot already supply.

Core claim

The central claim is that the forward corruption process of a diffusion forecaster can be redesigned so that the model's generative capacity is reallocated toward the components of the target that the observed history under-determines. DiffDiff defines the forward marginal q(xt|x0) = N(sqrt(alpha_t) At x0, (1 - alpha_t) I) with At = (1 - lambda_t) I + lambda_t D2, where D2 is the second-order differencing operator and lambda_t grows with the diffusion step. Early steps stay close to the clean target; later steps progressively shift the signal toward its differenced structure, suppressing level and trend (low-frequency) content while preserving local fluctuations (high-frequency content). The

What carries the argument

The load-bearing object is the step-dependent transition matrix At = (1 - lambda_t) I + lambda_t D2, a convex interpolation between the identity map and the second-order differencing operator D2 (whose interior entries compute zs - 2 z_{s-1} + z_{s-2}). As lambda_t increases, At behaves like a high-pass filter: at DC the squared frequency response is (1 - lambda_t)^2, and at Nyquist it is (1 + 3 lambda_t)^2, so low-frequency components are attenuated earlier and high-frequency components remain salient longer. Because the noise covariance stays isotropic, this reshapes the per-frequency signal-to-noise ratio along the chain, concentrating the diffusion's representational budget on the compon

Load-bearing premise

The load-bearing premise is that, in the target window, what the history can anchor is mostly low-frequency content and what it cannot is mostly high-frequency content; if a dataset violates this spectral alignment, the fixed second-order differencing operator suppresses precisely the components the denoiser would need.

What would settle it

Construct a synthetic forecasting task with a deterministic high-frequency oscillation and a noisy low-frequency drift, so the predictable component is high-frequency and the uncertain component is low-frequency. Train DiffDiff and its isotropic-forward ablation and compare the R^2-versus-aggressiveness correlation at t/T = 0.5; if the correlation does not reverse or vanish, the operator is doing something other than predictability alignment, and if DiffDiff's edge over the ablation disappears, the assumed spectral alignment is what the method relies on.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • The standard Gaussian endpoint is preserved, so the reshaped forward operator can be used with existing DDIM-style sub-sampling and accelerated samplers without retraining the sampling loop.
  • The forward process, not the conditioning network, is the primary source of improvement: in the four-way ablation the differencing forward alone wins 25 settings while the improved denoiser alone wins none.
  • At intermediate diffusion progress the corruption becomes predictability-aligned: the correlation between history-predictability (R^2) and corruption aggressiveness rises from about zero to +0.76 on wind and +0.58 on exchange rate.
  • The scheme is cheap: in the measured setting it keeps training memory and per-step time near the lightest diffusion baseline and gives the fastest serial inference among the six diffusion methods.
  • The gains concentrate where the spectrum matters most: on long horizons and datasets with periodic or regime-shifting structure, where the forward differencing attenuates exactly the components that would otherwise consume reconstruction capacity.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The marginal-consistency theorem holds for any sequence of operators, so a natural next step—which the paper's limitations section concedes—is to learn At from data rather than fixing D2; the same proof framework would carry over, and the open question is whether a learned operator generalizes across datasets.
  • Because D2's frequency response is closed-form, one could score candidate differencing orders or lambda-t schedules against a dataset's spectrum before training: the paper's own order ablation (D1/D2/D3) suggests the optimal order is dataset-dependent, so a pre-training spectral diagnostic could make the choice data-driven.
  • If the asymmetry thesis is right, it reframes hybrid deterministic-plus-diffusion forecasters: instead of freezing a deterministic/stochastic boundary before generation, the boundary can emerge from the corruption path, avoiding the failure mode where errors in a pre-fitted point predictor leak into the residual the diffusion must model.
  • A focused test would apply DiffDiff to series whose predictable component is deliberately high-frequency (e.g., a deterministic fast oscillation plus noisy slow drift); the fixed D2 should then suppress exactly the wrong components, which would delimit the class of series for which trajectory differencing helps.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes DIFFDIFF, a conditional diffusion framework for probabilistic time series forecasting in which the forward operator is step-dependent: the noisy state interpolates between the target itself and its second-order differenced representation via A_t = (1-λ_t)I + λ_t D_2. A stage-adaptive gate fuses value-domain and differential history encodings in the denoiser, and the target includes a short label window to preserve cross-boundary structure. The authors prove marginal consistency of the reverse update for a perfect denoiser (Theorem 1), characterize the spectral reshaping of the forward operator (Corollary 1, Proposition 3), and provide extensive experiments on seven benchmarks showing wins over six diffusion baselines and ablations attributing most gains to the differencing forward path.

Significance. If the claimed mechanism is correct, the paper makes a useful conceptual contribution: instead of decoupling deterministic and uncertain content externally, the diffusion trajectory itself is shaped so that history-anchored low-frequency content is suppressed while under-determined high-frequency content remains salient. The empirical work is substantial and carefully controlled: seven benchmarks, four horizons, ablations isolating the forward process and the denoiser, sensitivity analyses, significance tests, and a limitations section that honestly states the main assumption. The theoretical results in the appendix (spectral characterization, marginal consistency, terminal convergence) are formal and correct under the stated interior approximation. The code is promised publicly. The main weaknesses are that the training objective is not derived from a forward-process likelihood or valid divergence for the non-Markovian family, and that the core predictability-frequency alignment is demonstrated on only two datasets with a linear proxy, while the paper's own limitations section concedes this is not universally guaranteed.

major comments (3)
  1. [§3.4, Eq. (12); Appendix B.3] The training objective L_recon is a heuristic: the forward family defined by Eq. (2) is non-Markovian, and Theorem 1 guarantees marginal consistency only when the denoiser is perfect. No ELBO, score-matching identity, or other divergence is derived that would justify minimizing the value-domain MSE for this non-isotropic, non-Markovian corruption. Appendix B.3 explicitly says the simplification is 'justified empirically.' Since the paper's central claim is improved probabilistic forecasting, this gap is load-bearing. Please either derive a variational or score-matching objective for the generalized forward family, or provide a formal or synthetic validation that minimizing Eq. (12) controls the distance to the true conditional distribution.
  2. [§3.1, Eqs. (3)–(4); §5.3; Appendix K] The mechanism claim that D_2 'reallocates capacity toward history-underdetermined components' rests on the assumption that history-anchored content is mostly low-frequency and residual uncertainty mostly high-frequency. The paper's own Appendix K concedes that datasets violating this alignment would need a data-aware A_t, and Figure 3 provides evidence on only two datasets (wind, exchange rate), using linear-regression R² as the predictability measure. This is not a fatal flaw, but it means the generality of the frequency-predictability alignment is not established. I request either (a) a per-dataset analysis of the spectral-predictability correspondence across all seven benchmarks, or (b) a synthetic experiment with controlled spectral-uncertainty alignment, or (c) a clear weakening of the mechanism claim in the abstract and conclusion.
  3. [Appendix B.4, Eq. (42)] The upper bound in Proposition 2 depends on the auxiliary variance predictor ν̂ being positive and bounded, but the text notes that the linear projection is unconstrained and 'relies on the auxiliary MSE loss' to drive ν̂ toward the true ν > 0. During training this may hold approximately, but at inference there is no architectural guarantee that ν̂ is positive, and a negative variance estimate would make the de-normalization ill-defined. This issue is not load-bearing for the empirical results, but it should be addressed either by clamping ν̂ or by discussing why the loss is sufficient in practice.
minor comments (5)
  1. [§2.2] The related work on non-isotropic diffusion is accurate, but the distinction between DIFFDIFF and Whitened Score Diffusion could be sharper: both consider arbitrary Gaussian forward processes, and the paper should state more explicitly that the contribution is the forecasting-specific choice of A_t and the conditioning design, not the general framework.
  2. [§3.2, Eq. (7)] The reverse update is written for arbitrary step pairs (t,s), but the text then says 'we set σ_t = 0 in all experiments.' Please state next to Eq. (7) that the stochastic term is included only for generality and all reported results use deterministic DDIM sampling.
  3. [§5.3, Figure 3] The R²-versus-aggressiveness correlation is reported for t/T=0.5 only, and the caption does not mention how many windows/components constitute the points. Reporting the number of components and the variability across windows would strengthen the claim.
  4. [Table 10] Several Wilcoxon p-values are reported as '<10^{-300}', which is beyond numerical precision of standard double-precision computation; please report the smallest representable value or use a tie-breaking statement.
  5. [Appendix C, Table 3] Dataset statistics are useful, but the number of channels per multivariate dataset is not listed. Please add the channel count, since it affects the per-channel application of D_2.

Circularity Check

0 steps flagged

No circularity found: the derivation is self-contained and the empirical claims are benchmark-earned.

full rationale

The derivation chain is not circular. The forward operator At=(1−λt)I+λtD2 (Eq. 4) is a fixed, hand-specified construction with λt increasing linearly to λmax=0.5; λmax, Ll=48, and wfut=5 are chosen a priori and only sensitivity-tested in Appendix G, never fit to the target and then reported as predictions. The main theoretical results are general: Theorem 1 (Appendix B.2) proves marginal consistency for any sequence {At}, Proposition 4 (Appendix B.6) gives terminal Gaussianity from αT→0 and boundedness of At, and the spectral analysis in Proposition 3 merely characterizes the chosen D2. None of these results assumes the empirical conclusion that D2 reallocates capacity to under-determined components. Section 5.3's R²-versus-aggressiveness correlation is a post-hoc diagnostic on two datasets, not a fitted parameter or a prediction target; the paper itself concedes in Appendix K that a data-aware At would be needed if the low-frequency/high-frequency alignment fails, which is a scope limitation rather than a circular step. Self-citations exist (e.g., [17] in the introductory premise), but that premise is also supported by independent prior work [15,16], and the benchmarks compare against external baselines; no uniqueness theorem or ansatz is imported from the authors' own prior papers in a load-bearing way. The wins in Tables 1–2 are earned by external evaluation, so no prediction reduces to its input by construction.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central method introduces no new physical or ontological entities: D2 is a standard finite-difference matrix, and the gate and MLP conditioning are ordinary network components. The hand-set hyperparameters λmax, Ll, and wfut are the main fitted-by-hand degrees of freedom, with sensitivity analyses provided.

free parameters (4)
  • λ_max (differencing strength) = 0.5
    Controls how aggressively the forward operator shifts toward D2. Chosen by hand; sensitivity analysis (Table 6) shows stable behavior over 0.1–0.9, with 0.5 used for all reported results.
  • L_l (label window length) = 48
    Length of the observed-history overlap appended to the target; required for cross-boundary second-order differences. Chosen by hand; sensitivity analysis in Table 7.
  • w_fut (future-position loss weight) = 5
    Weights future positions in the reconstruction loss relative to label positions. Chosen by hand; sensitivity analysis in Table 8.
  • DDIM sampling steps = 10
    Number of reverse steps at inference. Chosen by hand; Table 9 shows notable non-monotonicity, e.g., Exchange H=336 CRPS is 0.388 at 10 steps but 0.569 at 20 steps.
axioms (4)
  • domain assumption Minimizing the x0-reconstruction loss (Eq. 12) on marginals of the step-dependent forward process yields a denoiser that correctly inverts the generalized corruption.
    The paper motivates this via perfect-denoiser marginal consistency (Theorem 1), but does not derive a training bound or score-matching equivalence for the non-Markovian forward family. The reverse sampler's validity depends on this assumption.
  • domain assumption Predictability aligns with frequency: history anchors low-frequency content, while residual uncertainty concentrates in high-frequency, differenced content.
    The entire design of D2 and λ_t rests on this. Appendix K explicitly concedes that datasets violating this alignment would need a data-aware A_t.
  • standard math Boundary rows of the non-circulant D2 are negligible for S≥48, so the interior frequency response characterizes the forward process.
    Corollary 1 and Proposition 3 use the interior symbol; rows s=1,2 behave differently, and the paper bounds their fraction as at most 4.2% for the shortest target.
  • domain assumption Predicted instance statistics (μ̂, ν̂) remain accurate enough to de-normalize generated samples.
    Appendix B.4 gives an upper bound that depends on ν̂ being positive; the auxiliary MSE loss is relied on to enforce this, but no explicit positivity constraint is imposed.

pith-pipeline@v1.3.0-alltime-deepseek · 28525 in / 15907 out tokens · 165125 ms · 2026-08-02T11:28:13.859667+00:00 · methodology

0 comments
read the original abstract

Diffusion models have become a widely used framework for probabilistic time series forecasting, modeling the distribution of future values given an observed history. In time series forecasting, however, the future continues the observed history, creating an asymmetry the standard diffusion process leaves unaddressed, with slowly-varying content largely determined by the observed continuity while higher-frequency dynamics carry most of the residual uncertainty. Existing diffusion-based forecasters decouple this asymmetry through an external rule before generation, leaving the corruption trajectory blind to which parts of the target the history can already anchor. We propose DiffDiff, a diffusion framework that embeds this predictability asymmetry into the diffusion trajectory itself, so that a single end-to-end diffusion process becomes aware of which parts of the target the history can already anchor. DiffDiff makes the forward operator step-dependent so that the noisy intermediate state progressively shifts from the target itself toward its second-order differenced structure, while a conditioning pathway supplies the denoiser with both value-domain and differential history information balanced by a stage-adaptive gate at each diffusion step. The terminal distribution approaches a standard Gaussian, preserving compatibility with existing samplers. On seven benchmarks across four prediction horizons, DiffDiff outperforms six diffusion baselines, and our analysis confirms that DiffDiff concentrates the diffusion's generative effort on the most uncertain components of the target while relieving it from rebuilding the history-anchored content.

Figures

Figures reproduced from arXiv: 2607.22599 by Chen Su, Yan Song, Yuanhe Tian.

Figure 1
Figure 1. Figure 1: Overall architecture of DIFFDIFF. Top: the differencing forward process gradually shifts the clean target x0 toward its differenced structure under the step-dependent operator At, and the deterministic reverse step recovers xs from the denoiser prediction xˆ0. Bottom: the stage-adaptive conditional denoiser encodes both the value-domain history xhist and its first-order differences ∆xhist, fuses them throu… view at source ↗
Figure 2
Figure 2. Figure 2: Reshaping of the corruption trajectory by the differencing forward process. [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Component-level corruption alignment and gate dynamics. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Per-window 95% prediction interval of DIFFDIFF and three diffusion baselines on il￾lustrative windows from electricity (top) and exchange rate (bottom) at H=96, each selected as a top-CRPS-improvement window of DIFFDIFF over the strongest baseline on that dataset. The gray line shows the history, the black line the ground truth, the colored line the median forecast, and the shaded band the 95% prediction i… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

75 extracted references · 13 linked inside Pith

  1. [1]

    Recent advances in electricity price forecasting: A review of probabilistic forecasting.Renewable and Sustainable Energy Reviews, 81:1548–1568, 2018

    Jakub Nowotarski and Rafał Weron. Recent advances in electricity price forecasting: A review of probabilistic forecasting.Renewable and Sustainable Energy Reviews, 81:1548–1568, 2018

    2018

  2. [2]

    Forecasting energy consumption time series using machine learning techniques based on usage patterns of residential householders.Energy, 165:709–726, 2018

    Jui-Sheng Chou and Duc-Son Tran. Forecasting energy consumption time series using machine learning techniques based on usage patterns of residential householders.Energy, 165:709–726, 2018

    2018

  3. [3]

    Financial time series forecasting model based on ceemdan and lstm.Physica A: Statistical mechanics and its applications, 519:127–139, 2019

    Jian Cao, Zhi Li, and Jian Li. Financial time series forecasting model based on ceemdan and lstm.Physica A: Statistical mechanics and its applications, 519:127–139, 2019

    2019

  4. [4]

    Financial time series forecasting-a machine learning approach.Machine Learning and Applications: An International Journal, 4(1/2):3, 2017

    Alexiei Dingli and Karl Sant Fournier. Financial time series forecasting-a machine learning approach.Machine Learning and Applications: An International Journal, 4(1/2):3, 2017

    2017

  5. [5]

    Temporal convolutional neural (tcn) network for an effective weather forecasting using time-series data from the local weather station: P

    Pradeep Hewage, Ardhendu Behera, Marcello Trovati, Ella Pereira, Morteza Ghahremani, Francesco Palmieri, and Yonghuai Liu. Temporal convolutional neural (tcn) network for an effective weather forecasting using time-series data from the local weather station: P. hewage et al.Soft Computing, 24(21):16453–16482, 2020

    2020

  6. [6]

    Transductive lstm for time-series prediction: An application to weather forecasting.Neural Networks, 125:1–9, 2020

    Zahra Karevan and Johan AK Suykens. Transductive lstm for time-series prediction: An application to weather forecasting.Neural Networks, 125:1–9, 2020

    2020

  7. [7]

    Deep unsuper- vised learning using nonequilibrium thermodynamics

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsuper- vised learning using nonequilibrium thermodynamics. InInternational conference on machine learning, pages 2256–2265. pmlr, 2015

    2015

  8. [8]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. volume 33, pages 6840–6851, 2020

    2020

  9. [9]

    Score-based generative modeling through stochastic differential equations

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. 2020

    2020

  10. [10]

    High- resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High- resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022

    2022

  11. [11]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF international conference on computer vision, pages 4195–4205, 2023

    2023

  12. [12]

    Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps

    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. volume 35, pages 5775–5787, 2022

    2022

  13. [13]

    idesigner: a high-resolution and complex-prompt following text-to-image diffusion model for interior design.arXiv preprint arXiv:2312.04326, 2023

    Ruyi Gan, Xiaojun Wu, Junyu Lu, Yuanhe Tian, Dixiang Zhang, Ziwei Wu, Renliang Sun, Chang Liu, Jiaxing Zhang, Pingjian Zhang, et al. idesigner: a high-resolution and complex-prompt following text-to-image diffusion model for interior design.arXiv preprint arXiv:2312.04326, 2023

    Pith/arXiv arXiv 2023

  14. [14]

    Taiyi-diffusion-xl: advancing bilingual text-to-image generation with large vision-language model support.arXiv preprint arXiv:2401.14688, 2024

    Xiaojun Wu, Dixiang Zhang, Ruyi Gan, Junyu Lu, Ziwei Wu, Renliang Sun, Jiaxing Zhang, Pingjian Zhang, and Yan Song. Taiyi-diffusion-xl: advancing bilingual text-to-image generation with large vision-language model support.arXiv preprint arXiv:2401.14688, 2024

    Pith/arXiv arXiv 2024

  15. [15]

    Diffusion-based decoupled deterministic and uncertain framework for probabilistic multivariate time series forecasting

    Qi Li, Zhenyu Zhang, Lei Yao, Zhaoxia Li, Tianyi Zhong, and Yong Zhang. Diffusion-based decoupled deterministic and uncertain framework for probabilistic multivariate time series forecasting. InThe Thirteenth International Conference on Learning Representations, 2025

    2025

  16. [16]

    Effective probabilistic time series forecasting with fourier adaptive noise-separated diffusion.arXiv preprint arXiv:2505.11306, 2025

    Xinyan Wang, Rui Dai, Kaikui Liu, and Xiangxiang Chu. Effective probabilistic time series forecasting with fourier adaptive noise-separated diffusion.arXiv preprint arXiv:2505.11306, 2025. 10

    Pith/arXiv arXiv 2025

  17. [17]

    Diffusion models for time series forecasting: A survey.arXiv preprint arXiv:2507.14507, 2025

    Chen Su, Zhengzhou Cai, Yuanhe Tian, Zhuochao Chang, Zihong Zheng, and Yan Song. Diffusion models for time series forecasting: A survey.arXiv preprint arXiv:2507.14507, 2025

    Pith/arXiv arXiv 2025

  18. [18]

    Conditional denois- ing meets polynomial modeling: A flexible decoupled framework for time series forecasting

    Jintao Zhang, Mingyue Cheng, Xiaoyu Tao, Zhiding Liu, and Daoyu Wang. Conditional denois- ing meets polynomial modeling: A flexible decoupled framework for time series forecasting. arXiv preprint arXiv:2410.13253, 2024

    arXiv 2024

  19. [19]

    Transformer- modulated diffusion models for probabilistic multivariate time series forecasting

    Yuxin Li, Wenchao Chen, Xinyue Hu, Bo Chen, Baolin Sun, and Mingyuan Zhou. Transformer- modulated diffusion models for probabilistic multivariate time series forecasting. InThe Twelfth International Conference on Learning Representations, 2024

    2024

  20. [20]

    Diffusion networks with task-specific noise control for radiology report generation

    Yuanhe Tian, Fei Xia, and Yan Song. Diffusion networks with task-specific noise control for radiology report generation. InProceedings of the 32nd ACM International Conference on Multimedia, pages 1771–1780, 2024

    2024

  21. [21]

    Conditional diffusion model with nonlinear data transformation for time series forecasting

    J Rishi, GVS Mothish, and Deepak Subramani. Conditional diffusion model with nonlinear data transformation for time series forecasting. InForty-second International Conference on Machine Learning, 2025

    2025

  22. [22]

    Non-stationary diffusion for probabilistic time series forecasting

    Weiwei Ye, Zhuopeng Xu, and Ning Gui. Non-stationary diffusion for probabilistic time series forecasting. 2025

    2025

  23. [23]

    A non-isotropic time series diffusion model with moving average transitions

    Chenxi Wang, Linxiao Yang, Zhixian Wang, Liang Sun, and Yi Wang. A non-isotropic time series diffusion model with moving average transitions. InForty-second International Conference on Machine Learning, 2025

    2025

  24. [24]

    Autoregressive denois- ing diffusion models for multivariate probabilistic time series forecasting

    Kashif Rasul, Calvin Seward, Ingmar Schuster, and Roland V ollgraf. Autoregressive denois- ing diffusion models for multivariate probabilistic time series forecasting. InInternational conference on machine learning, pages 8857–8868. PMLR, 2021

    2021

  25. [25]

    Stochastic diffusion: A diffusion based model for stochastic time series forecasting

    Yuansan Liu, Sudanthi Wijewickrema, Dongting Hu, Christofer Bester, Stephen O’Leary, and James Bailey. Stochastic diffusion: A diffusion based model for stochastic time series forecasting. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V . 2, pages 1939–1950, 2025

    1939

  26. [26]

    Csdi: Conditional score-based diffusion models for probabilistic time series imputation

    Yusuke Tashiro, Jiaming Song, Yang Song, and Stefano Ermon. Csdi: Conditional score-based diffusion models for probabilistic time series imputation. volume 34, pages 24804–24816, 2021

    2021

  27. [27]

    Diffusion-based time series imputation and forecasting with structured state space models.arXiv preprint arXiv:2208.09399, 2022

    Juan Miguel Lopez Alcaraz and Nils Strodthoff. Diffusion-based time series imputation and forecasting with structured state space models.arXiv preprint arXiv:2208.09399, 2022

    Pith/arXiv arXiv 2022

  28. [28]

    Non-autoregressive conditional diffusion models for time series prediction

    Lifeng Shen and James Kwok. Non-autoregressive conditional diffusion models for time series prediction. InInternational Conference on Machine Learning, pages 31016–31029. PMLR, 2023

    2023

  29. [29]

    Predict, refine, synthesize: Self-guiding diffusion models for probabilistic time series forecasting

    Marcel Kollovieh, Abdul Fatir Ansari, Michael Bohlke-Schneider, Jasper Zschiegner, Hao Wang, and Yuyang Bernie Wang. Predict, refine, synthesize: Self-guiding diffusion models for probabilistic time series forecasting. volume 36, pages 28341–28364, 2023

    2023

  30. [30]

    Multimodal conditioned diffusive time series forecasting

    Chen Su, Yuanhe Tian, and Yan Song. Multimodal conditioned diffusive time series forecasting. arXiv preprint arXiv:2504.19669, 2025

    Pith/arXiv arXiv 2025

  31. [31]

    Fusing large language models with temporal transformers for time series forecasting.arXiv preprint arXiv:2507.10098, 2025

    Chen Su, Yuanhe Tian, Qinyu Liu, Jun Zhang, and Yan Song. Fusing large language models with temporal transformers for time series forecasting.arXiv preprint arXiv:2507.10098, 2025

    Pith/arXiv arXiv 2025

  32. [32]

    Multi-resolution diffusion models for time series forecasting

    Lifeng Shen, Weiyu Chen, and James Kwok. Multi-resolution diffusion models for time series forecasting. InThe Twelfth International Conference on Learning Representations, 2024

    2024

  33. [33]

    Mg-tsd: Multi-granularity time series diffusion models with guided learning process

    Xinyao Fan, Yueying Wu, Chang Xu, Yuhao Huang, Weiqing Liu, and Jiang Bian. Mg-tsd: Multi-granularity time series diffusion models with guided learning process. 2024

    2024

  34. [34]

    Text reinforcement for multimodal time series forecasting.arXiv preprint arXiv:2509.00687, 2025

    Chen Su, Yuanhe Tian, Yan Song, and Yongdong Zhang. Text reinforcement for multimodal time series forecasting.arXiv preprint arXiv:2509.00687, 2025. 11

    Pith/arXiv arXiv 2025

  35. [35]

    Diffusion-ts: Interpretable diffusion for general time series genera- tion

    Xinyu Yuan and Yan Qiao. Diffusion-ts: Interpretable diffusion for general time series genera- tion. 2024

    2024

  36. [36]

    Generative time series forecasting with diffusion, denoise, and disentanglement

    Yan Li, Xinjiang Lu, Yaqing Wang, and Dejing Dou. Generative time series forecasting with diffusion, denoise, and disentanglement. volume 35, pages 23009–23022, 2022

    2022

  37. [37]

    Retrieval-augmented diffusion models for time series forecasting

    Jingwei Liu, Ling Yang, Hongyan Li, and Shenda Hong. Retrieval-augmented diffusion models for time series forecasting. volume 37, pages 2766–2786, 2024

    2024

  38. [38]

    Retrieval- augmented diffusion models

    Andreas Blattmann, Robin Rombach, Kaan Oktay, Jonas Müller, and Björn Ommer. Retrieval- augmented diffusion models. volume 35, pages 15309–15324, 2022

    2022

  39. [39]

    Retrieval based time series forecasting.arXiv preprint arXiv:2209.13525, 2022

    Baoyu Jing, Si Zhang, Yada Zhu, Bin Peng, Kaiyu Guan, Andrew Margenot, and Hanghang Tong. Retrieval based time series forecasting.arXiv preprint arXiv:2209.13525, 2022

    Pith/arXiv arXiv 2022

  40. [40]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. InInternational conference on machine learning, pages 8748–8763. PmLR, 2021

    2021

  41. [41]

    Flamingo: a visual language model for few-shot learning.Advances in neural information processing systems, 35: 23716–23736, 2022

    Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning.Advances in neural information processing systems, 35: 23716–23736, 2022

    2022

  42. [42]

    Reinforced context augmentation for multimodal emotion analysis.IEEE Transactions on Multimedia, 2026

    Ruyi Gan, Yuanhe Tian, Kunhao Pan, Yan Song, and Yongdong Zhang. Reinforced context augmentation for multimodal emotion analysis.IEEE Transactions on Multimedia, 2026

    2026

  43. [43]

    Learning shared sentiment prototypes for adaptive multimodal sentiment analysis.arXiv preprint arXiv:2604.05873, 2026

    Chen Su, Yuanhe Tian, and Yan Song. Learning shared sentiment prototypes for adaptive multimodal sentiment analysis.arXiv preprint arXiv:2604.05873, 2026

    Pith/arXiv arXiv 2026

  44. [44]

    A time series is worth 64 words: Long-term forecasting with transformers

    Yuqi Nie, Nam H Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. A time series is worth 64 words: Long-term forecasting with transformers. 2022

    2022

  45. [45]

    itransformer: Inverted transformers are effective for time series forecasting

    Yong Liu, Tengge Hu, Haoran Zhang, Haixu Wu, Shiyu Wang, Lintao Ma, and Mingsheng Long. itransformer: Inverted transformers are effective for time series forecasting. 2023

    2023

  46. [46]

    Timesnet: Temporal 2d-variation modeling for general time series analysis

    Haixu Wu, Tengge Hu, Yong Liu, Hang Zhou, Jianmin Wang, and Mingsheng Long. Timesnet: Temporal 2d-variation modeling for general time series analysis. 2022

    2022

  47. [47]

    N-beats: Neural basis expansion analysis for interpretable time series forecasting

    Boris N Oreshkin, Dmitri Carpov, Nicolas Chapados, and Yoshua Bengio. N-beats: Neural basis expansion analysis for interpretable time series forecasting. 2019

    2019

  48. [48]

    Non-stationary transformers: Exploring the stationarity in time series forecasting

    Yong Liu, Haixu Wu, Jianmin Wang, and Mingsheng Long. Non-stationary transformers: Exploring the stationarity in time series forecasting. volume 35, pages 9881–9893, 2022

    2022

  49. [49]

    Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting

    Tian Zhou, Ziqing Ma, Qingsong Wen, Xue Wang, Liang Sun, and Rong Jin. Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting. InInternational conference on machine learning, pages 27268–27286. PMLR, 2022

    2022

  50. [50]

    Crossformer: Transformer utilizing cross-dimension dependency for multivariate time series forecasting

    Yunhao Zhang and Junchi Yan. Crossformer: Transformer utilizing cross-dimension dependency for multivariate time series forecasting. InThe eleventh international conference on learning representations, 2023

    2023

  51. [51]

    Pyraformer: Low-complexity pyramidal attention for long-range time series modeling and forecasting

    Shizhan Liu, Hang Yu, Cong Liao, Jianguo Li, Weiyao Lin, Alex X Liu, and Schahram Dustdar. Pyraformer: Low-complexity pyramidal attention for long-range time series modeling and forecasting. InInternational conference on learning representations, 2021

    2021

  52. [52]

    Blurring diffusion models

    Emiel Hoogeboom and Tim Salimans. Blurring diffusion models. 2022

    2022

  53. [53]

    Generative modelling with inverse heat dissipation

    Severi Rissanen, Markus Heinonen, and Arno Solin. Generative modelling with inverse heat dissipation. 2022. 12

    2022

  54. [54]

    Improving radiology report generation with multi-grained abnormality prediction.Neurocomputing, 600:128122, 2024

    Yuda Jin, Weidong Chen, Yuanhe Tian, Yan Song, and Chenggang Yan. Improving radiology report generation with multi-grained abnormality prediction.Neurocomputing, 600:128122, 2024

    2024

  55. [55]

    Soft diffusion: Score matching for general corruptions.arXiv preprint arXiv:2209.05442, 2022

    Giannis Daras, Mauricio Delbracio, Hossein Talebi, Alexandros G Dimakis, and Peyman Milan- far. Soft diffusion: Score matching for general corruptions.arXiv preprint arXiv:2209.05442, 2022

    Pith/arXiv arXiv 2022

  56. [56]

    Cold diffusion: Inverting arbitrary image transforms without noise

    Arpit Bansal, Eitan Borgnia, Hong-Min Chu, Jie Li, Hamid Kazemi, Furong Huang, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Cold diffusion: Inverting arbitrary image transforms without noise. volume 36, pages 41259–41282, 2023

    2023

  57. [57]

    Whitened score diffusion: A structured prior for imaging inverse problems.arXiv preprint arXiv:2505.10311, 2025

    Jeffrey Alido, Tongyu Li, Yu Sun, and Lei Tian. Whitened score diffusion: A structured prior for imaging inverse problems.arXiv preprint arXiv:2505.10311, 2025

    arXiv 2025

  58. [58]

    Improved denoising diffusion probabilistic models

    Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. InInternational conference on machine learning, pages 8162–8171. PMLR, 2021

    2021

  59. [59]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. 2020

    2020

  60. [60]

    Reversible instance normalization for accurate time-series forecasting against distribution shift

    Taesung Kim, Jinhee Kim, Yunwon Tae, Cheonbok Park, Jang-Ho Choi, and Jaegul Choo. Reversible instance normalization for accurate time-series forecasting against distribution shift. InInternational conference on learning representations, 2021

    2021

  61. [61]

    Informer: Beyond efficient transformer for long sequence time-series forecasting

    Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang. Informer: Beyond efficient transformer for long sequence time-series forecasting. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 11106–11115, 2021

    2021

  62. [62]

    Modeling long-and short-term temporal patterns with deep neural networks

    Guokun Lai, Wei-Cheng Chang, Yiming Yang, and Hanxiao Liu. Modeling long-and short-term temporal patterns with deep neural networks. InThe 41st international ACM SIGIR conference on research & development in information retrieval, pages 95–104, 2018

    2018

  63. [63]

    Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting

    Haixu Wu, Jiehui Xu, Jianmin Wang, and Mingsheng Long. Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting. volume 34, pages 22419– 22430, 2021

    2021

  64. [64]

    Strictly proper scoring rules, prediction, and estimation

    Tilmann Gneiting and Adrian E Raftery. Strictly proper scoring rules, prediction, and estimation. Journal of the American statistical Association, 102(477):359–378, 2007

    2007

  65. [65]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. 2014

    2014

  66. [66]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. volume 30, 2017

    2017

  67. [67]

    Statistical comparisons of classifiers over multiple data sets.Journal of Machine learning research, 7(Jan):1–30, 2006

    Janez Demšar. Statistical comparisons of classifiers over multiple data sets.Journal of Machine learning research, 7(Jan):1–30, 2006

    2006

  68. [68]

    Temporal query network for efficient multivariate time series forecasting.arXiv preprint arXiv:2505.12917, 2025

    Shengsheng Lin, Haojun Chen, Haijie Wu, Chunyun Qiu, and Weiwei Lin. Temporal query network for efficient multivariate time series forecasting.arXiv preprint arXiv:2505.12917, 2025

    Pith/arXiv arXiv 2025

  69. [69]

    Times2d: Multi-period decom- position and derivative mapping for general time series forecasting

    Reza Nematirad, Anil Pahwa, and Balasubramaniam Natarajan. Times2d: Multi-period decom- position and derivative mapping for general time series forecasting. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 19651–19658, 2025

    2025

  70. [70]

    Pathformer: Multi-scale transformers with adaptive pathways for time series forecasting.arXiv preprint arXiv:2402.05956, 2024

    Peng Chen, Yingying Zhang, Yunyao Cheng, Yang Shu, Yihang Wang, Qingsong Wen, Bin Yang, and Chenjuan Guo. Pathformer: Multi-scale transformers with adaptive pathways for time series forecasting.arXiv preprint arXiv:2402.05956, 2024. 13 A Diffusion Model Background This appendix reviews the standard denoising diffusion framework that underpins the DIFFDIFF...

    Pith/arXiv arXiv 2024

  71. [71]

    At DC (ω= 0):SNR eff,DD /SNReff,std = (1−λ t)2 <1forλ t ∈(0,1)

  72. [72]

    At Nyquist (ω=π):SNR eff,DD /SNReff,std = (1 + 3λt)2 >1forλ t >0

  73. [73]

    19 Proof.Items (1) and (2) follow directly from Corollary 1

    For λt ∈(0,1) , there exists a unique crossover frequency ω∗ ∈(0, π)where |HAt (ω∗)|2 = 1. 19 Proof.Items (1) and (2) follow directly from Corollary 1. For item (1), (1−λ t)2 <1 holds whenever λt ∈(0,1) (equivalently 0<1−λ t <1 ). For item (3), we have|HAt (0)|2 = (1−λt)2 <1 and |HAt (π)|2 = (1 + 3λt)2 >1 for λt ∈(0,1) . Since |HAt (ω)|2 is a continuous f...

  74. [74]

    The future-only target satisfies [D(H) 2 xfut]1 = 0 and [D(H) 2 xfut]2 =x f 2 −x f 1, encoding only differences within the future window

  75. [75]

    boundary anchoring

    The extended target satisfies [D(S) 2 xext]Ll+1 =x f 1 −2x 0 +x −1, encoding the second-order difference across the history–future boundary. Consequently, the differencing forward process applied to the extended target preserves cross- boundary curvature information at positions near the junction, which is lost when the target is restricted to the future ...

    arXiv 1925