Pith. sign in

REVIEW 2 major objections 3 minor 1 cited by

RDIT: Residual-based Diffusion Implicit Models for Probabilistic Time Series Forecasting

T0 review · 2 major / 3 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A two-stage framework that predicts the median, then diffuses only residuals, proves the CRPS-optimal Gaussian width is |y−μ|/√ln2 and beats ten baselines on eight datasets.

desk verdict Solid incremental paper; the Gaussian baseline is the real gem, but the EAE theory doesn't cover the actual algorithm as claimed. read the letter →

arxiv 2509.02341 v1 pith:OBPRHKH4 submitted 2025-09-02 cs.LG cs.AI

Chih-Yu Lai , Yu-Chien Ning , Duane S. Boning This is my paper
classification cs.LGcs.AI
keywords probabilistictimeseriesforecastingresidualdiffusionCRPScoveragecalibrationDDIMbidirectionalMambauncertaintyquantificationvarianceoptimization
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that a strong probabilistic forecaster can be built by keeping point prediction and uncertainty modeling separate: first predict the conditional median, then model only the residuals with diffusion. It proves that for a Gaussian predictive distribution with fixed mean, the CRPS is minimized when the standard deviation equals |y−μ|/√ln2, and it turns this into two algorithms—Error-aware Expansion (EAE) and Coverage Optimization (CO)—that expand or shrink the residual distribution to match that target. On eight multivariate datasets, this two-stage RDIT framework beats ten baselines on CRPS, MAE, MSE, and coverage, including a striking baseline where a point estimator is merely wrapped in a zero-mean Gaussian fitted to training residuals.

What carries the argument

Residual-based conditional diffusion with DDIM accelerated sampling: a point estimator gives ŷ, residuals r = y − ŷ are normalized by the training residual standard deviation σ_trn, diffused, and denoised by a bidirectional Mamba network conditioned on history and ŷ; the final predictive distribution is ŷ plus the denoised residuals, adjusted by EAE and CO. EAE is the CRPS-optimal variance formula σ* = |y−μ|/√ln2 applied per sample through E[|r̂|]; CO performs interval-by-interval expansion factors calibrated on validation data to match target PICP values.

What would settle it

For a fixed point estimator on a dataset with skewed residuals, grid-search the per-point standard deviation that actually minimizes CRPS on a validation set. If the empirically optimal σ deviates substantially from 1.201·E[|r̂|] (the α=1 prediction) in a way that tracks skewness or tail weight, the EAE approximation is falsified. Alternatively, compute CRPS using the ground-truth |y−μ| in place of E[|r̂|] and check whether it reaches the theoretical minimum; any gap under the true error would refute the theorem's applicability.

Watch

Extended reading notes

Core claim

The central claim is that the mismatch between point-error training and probabilistic evaluation can be closed by rescaling the predictive distribution to the CRPS-optimal width. For fixed mean μ and ground truth y, the CRPS of a Gaussian predictive distribution N(μ,σ²) is minimized at σ = |y−μ|/√ln2, roughly 1.2 times the absolute error. Since the true error is unknown at test time, the paper approximates |y−μ| by E[|r̂|], the expected magnitude of diffusion-generated residuals, and sets the variance accordingly (EAE); a second stage (CO) recalibrates quantile coverage on a hold-out set. The diffusion model operates on normalized residuals, so both endpoints of the denoising process have ze

Load-bearing premise

The substitution |y−μ| ≈ α·E[|r̂|] with α=1 in Equation (12) assumes the mean absolute generated residual is a faithful per-point proxy for the true absolute error; if residuals are skewed, heavy-tailed, or miscentered, EAE sets systematically wrong variances and the proven optimality no longer applies to the algorithm actually used.

Editorial extensions

If this is right

  • Any point forecaster can be upgraded to a probabilistic one by diffusing its residuals; the paper shows additive improvements with both SMamba and TimeFilter as the point estimator.
  • The CRPS-optimal width formula gives a parameter-free target for post-hoc variance adjustment, replacing ad-hoc Gaussian scaling with a theoretically grounded one.
  • Because DDIM lets about ten denoising steps stand in for a thousand-step diffusion process, the framework's inference is fast enough for operational use.
  • Coverage can be calibrated without distributional assumptions by quantile-wise expansion factors derived from a validation set, as implemented in CO.
  • The zero-mean Gaussian baseline's strong performance establishes a cheap benchmark: a new probabilistic forecasting method should beat its point estimator wrapped in N(0,σ_trn²).

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If σ* = |y−μ|/√ln2 holds generally, the same variance-calibration rule applies to any Gaussian predictive distribution outside time series—for example, regression or Bayesian neural networks with Gaussian outputs could scale their uncertainty by the observed residual magnitude.
  • The α=1 choice in EAE is a hidden assumption; on skewed or heavy-tailed residual distributions, learning α per dataset or per quantile would likely improve CRPS beyond the fixed setting.
  • The strength of the fitted-Gaussian baseline suggests that future generative PTSF work should be benchmarked against 'point estimator + N(0,σ_trn²)' before claiming gains from distribution modeling.
  • EAE and CO operate on the residual scale and assume a roughly Gaussian final distribution; combining them with a non-Gaussian mixture or copula final layer is a natural extension the paper notes as future work.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 3 minor

Summary. The paper proposes RDIT, a two-stage plug-and-play framework for probabilistic multivariate time series forecasting. A point estimator (MAE-trained) produces conditional-median forecasts; a residual diffusion model with a bidirectional Mamba denoiser models the normalized residuals; DDIM acceleration is used at inference. Two distribution-matching post-processors are introduced: Error-aware Expansion (EAE), which scales the residual samples to a target standard deviation suggested by a CRPS-minimization theorem for Gaussian forecasts, and Coverage Optimization (CO), which calibrates quantile intervals on a validation split. Experiments on eight datasets over several horizons report CRPS, PICP-distance, MAE, and MSE improvements over ten baselines, with ablations showing each component contributes. The appendix contains the proof of Theorem 3.1, implementation details, full result tables, visualizations, and a Limitations section.

Significance. If the empirical results hold, the paper makes a useful practical contribution: decoupling point estimation from residual diffusion with a strong point estimator is an effective recipe, and DDIM acceleration plus validation-based coverage calibration addresses two practical bottlenecks in diffusion-based PTSF. The paper also ships code and provides detailed ablations, which strengthens reproducibility. The CRPS-minimization theorem for a Gaussian predictive distribution is correct and cleanly proved, and it is an instructive motivating result. However, the theoretical result is not actually matched to the EAE algorithm as implemented, because EAE replaces the unknown |y−μ| with α·E[|r̂|] and fixes α=1, and because the final predictive distribution is empirical, not Gaussian. The contribution is therefore primarily empirical; the advertised theoretical backing needs to be substantially qualified.

major comments (2)
  1. [Section 3.3.1, Eq. (12), Algorithm 2] Corollary 3.2 states σ* = |y−μ|/√ln2, but the implemented EAE replaces the unknown |y−μ| with α·E[|r̂|] and sets α=1. This substitution is not derived from the theorem; it assumes per-point absolute error is proportional to the sample mean absolute residual, a condition that is not stated or verified. Appendix J itself reports that the optimal α under a simple expansion search lies in [1,5], so α=1 is not the optimal coefficient by the paper's own experiments. Consequently, the variance produced by EAE is not shown to be the CRPS-optimal variance for the actual algorithm. Since the abstract and Section 1 claim that RDIT 'theoretically proves' CRPS minimization, the claim overstates what is demonstrated. This is load-bearing because EAE is listed as a key contribution and as a source of CRPS improvement.
  2. [Theorem 3.1 and Appendix N] The theorem is proved only under the assumption that the predictive distribution is Gaussian with fixed mean. RDIT's residual diffusion is specifically designed to capture arbitrary residual shapes, and the post-hoc scaling in Algorithm 2 produces an empirical distribution, not a Gaussian. The paper's own Limitations section (Appendix N) concedes that 'EAE currently works in theory only when predictions follow a Gaussian distribution.' Therefore the theoretical result is a motivating special case, not a proof of optimality for RDIT's actual predictive distribution. The manuscript should either restrict the theoretical claim to the Gaussian motivating case, or provide an analysis (even approximate) for the empirical/non-Gaussian setting used by EAE.
minor comments (3)
  1. [Algorithm 2] The expression for λEAE is ambiguous: 'λEAE ← αE[|r̂CO|]/σr̂CO √ln2' should be parenthesized as αE[|r̂CO|] / (σ_{r̂_CO}√ln2) to avoid implying a multiplication by √ln2.
  2. [Table 5] The learning rate entry for ETTh1 appears malformed: '0.0005≤336,0.3720' is not a valid number or notation. Please correct.
  3. [Appendix J] Appendix J states that the location of the minimum lies in [1,5] and that α=1 consistently improves over no expansion, but it does not clarify whether α is chosen per dataset or fixed. Since EAE fixes α=1 for all experiments, a sentence justifying this choice (or reporting sensitivity) would help.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the CRPS-optimal-σ theorem is independently derived; EAE is a transparent heuristic plug-in (Eq. 12), CO calibrates on a separate validation split, and the sole self-citation is non-load-bearing.

full rationale

The paper's central theoretical result (Theorem 3.1 / Corollary 3.2, Appendix E) is a self-contained calculus result: for a Gaussian predictive distribution with fixed mean μ and ground truth y, the CRPS is minimized at σ* = |y−μ|/√ln2. The proof does not use any fitted quantity or the authors' prior work, so it is not circular. The subsequent EAE algorithm (Sec. 3.3.1, Eq. 12) substitutes the unknown |y−μ| with α·E[|r̂|] and sets α=1; this is an approximation/plug-in rather than a defining identity. It may be an unsupported heuristic—especially since Appendix J shows the optimal α can range within [1,5] and Appendix N concedes "EAE currently works in theory only when predictions follow a Gaussian distribution"—but replacing an unknown quantity by an estimator is not circular; it is a standard (if imperfect) statistical approximation. CO (Sec. 3.3.2, Eq. 13) calibrates expansion factors on a hold-out validation set and applies them to the test set; this is an external empirical calibration, not a prediction forced by training data. The only self-citation is reference [34], cited in Sec. 3.3.2 as one of several similar calibration techniques; it is not load-bearing for any derivation or claim. The paper's abstract overstates the reach of the theoretical claim relative to the actual EAE algorithm, but that is a correctness/scope concern, not circularity. Therefore the derivation chain is not circular.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests on two fitted constants (α and the CO λ_i) and on the empirically estimated σ_trn. The EAE proxy assumption is ad hoc and is the main structural weakness; the others are standard domain assumptions.

free parameters (3)
  • α (EAE coefficient) = 1 (Table 5)
    Linear coefficient between |r| and E[|r̂|] in σ* = α E[|r̂|]/√(ln 2); chosen empirically, with Appendix J suggesting a narrow range 1-5.
  • σ_trn (training-residual standard deviation) = estimated per variate and horizon from training residuals
    Used to normalize residuals for diffusion and as the Gaussian baseline; estimated from the training split, so it is data-fitted.
  • λ_i (CO expansion factors) = binary search on validation set for quantiles γ
    Adjusts the width of prediction intervals to match PICP on validation data; fitted to the hold-out set.
assumptions (5)
  • domain assumption The MAE-trained point estimator yields zero-mean residuals (E[r] ≈ 0).
    Used in Equation (11) to set the mean of the predictive distribution to ŷ. Skewed or biased residuals would break the interpretation of z = ŷ + r̂ as N(ŷ, σ²).
  • ad hoc to paper The true absolute error |y−μ| is proportional to the expected absolute residual E[|r̂|], with proportionality α=1.
    This is the key proxy in Equation (12) that converts a per-point oracle theorem into a computable variance. It is a heuristic, not proven.
  • domain assumption Training residuals follow a zero-mean Gaussian distribution.
    Stated in Limitations; used to justify Gaussian augmentation and the σ_trn normalization.
  • domain assumption The validation set is representative of the test set for CO calibration.
    Standard conformal assumption; if distribution shift is large, the fitted λ_i will not transfer.
  • standard math DDIM sampling with W=10 steps approximates the DDPM distribution for residual denoising.
    The paper relies on Song et al. 2020 for the DDIM framework.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RDIT: Residual-based Diffusion Implicit Models for Probabilistic Time Series Forecasting." pith.science (2026). https://pith.science/paper/OBPRHKH4

@misc{pith2026250902341,
  author       = {Pith},
  title        = {Pith review of: RDIT: Residual-based Diffusion Implicit Models for Probabilistic Time Series Forecasting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OBPRHKH4}},
  note         = {Machine review of arXiv:2509.02341}
}
read the original abstract

Probabilistic Time Series Forecasting (PTSF) plays a critical role in domains requiring accurate and uncertainty-aware predictions for decision-making. However, existing methods offer suboptimal distribution modeling and suffer from a mismatch between training and evaluation metrics. Surprisingly, we found that augmenting a strong point estimator with a zero-mean Gaussian, whose standard deviation matches its training error, can yield state-of-the-art performance in PTSF. In this work, we propose RDIT, a plug-and-play framework that combines point estimation and residual-based conditional diffusion with a bidirectional Mamba network. We theoretically prove that the Continuous Ranked Probability Score (CRPS) can be minimized by adjusting to an optimal standard deviation and then derive algorithms to achieve distribution matching. Evaluations on eight multivariate datasets across varied forecasting horizons demonstrate that RDIT achieves lower CRPS, rapid inference, and improved coverage compared to strong baselines.

Figures

Figures reproduced from arXiv: 2509.02341 by the authors.

Figure 1
Figure 1. In the first stage, a point-based estimator ( [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 1
Figure 1. General scheme of this work. The point estimator ( [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. CRPS/|ε| vs σ/|ε| when the predictions z are drawn from a normal distribution with mean µ and standard deviation σ, and the ground truth being y. Let Fyˆ(·) be the CDF of the predictions for the ground truth y. If we condense the predictions to their mean (¯yˆ), making the standard deviation of the distribution 0, then Fyˆ(x) = 1{x ≥ ¯yˆ}, and by the definition of CRPS in (21), CRPS(Fyˆ, y) = Z +∞ −∞ (Fyˆ(x) − 1x≥y)… view at source ↗
Figures from the paper (7 more)
Figure 3
Figure 3. Figure 3: Visualization of prediction intervals for the ETTh1 dataset (variate #2) using SMamba, [PITH_FULL_IMAGE:figures/full_fig_p018_3.png]
Figure 4
Figure 4. Figure 4: Visualization of prediction intervals for the exchange dataset (variate #7) using SMamba, [PITH_FULL_IMAGE:figures/full_fig_p018_4.png]
Figure 5
Figure 5. Figure 5: The ground truth and the 20%, 50%, and 80% prediction intervals plotted against the [PITH_FULL_IMAGE:figures/full_fig_p019_5.png]
Figure 6
Figure 6. Figure 6: PICP plotted against interval percentile for the weather and solar dataset. The prediction [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]
Figure 7
Figure 7. Figure 7: CRPS after applying simple expansion (blue) or EAE (orange) for a prediction length [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]
Figure 8
Figure 8. Figure 8: CRPS during each denoising step across eight datasets. The prediction length is 24. The [PITH_FULL_IMAGE:figures/full_fig_p021_8.png]
Figure 9
Figure 9. Figure 9: Visualization on the Traffic dataset (variate #800) with a prediction length of 96. SMamba is [PITH_FULL_IMAGE:figures/full_fig_p022_9.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Mind the Residual Gap: Probabilistic Downscaling under Real-World Bias

    cs.LG 2026-06 unverdicted novelty 7.0 of 10

    ReMatch corrects train-test residual distribution mismatch in probabilistic downscaling via optimal transport in low-dimensional PCA space, reducing under-dispersion and improving SSR and CRPS on HRRR-ERA5 wind data.

Reference graph

Works this paper leans on

46 extracted references · 27 canonical work pages · cited by 1 Pith paper

  1. [1]

    Financial time series forecasting with deep learning : A systematic literature review: 2005–2019

    Omer Berat Sezer, Mehmet Ugur Gudelek, and Ahmet Murat Ozbayoglu. Financial time series forecasting with deep learning : A systematic literature review: 2005–2019. Applied Soft Computing, 90:106181, 2020

    work page 2005

  2. [2]

    Liu Sheng, and Joseph Dunbar

    Mohammad Amin Morid, Olivia R. Liu Sheng, and Joseph Dunbar. Time series prediction using deep learning methods in healthcare. ACM Trans. Manage. Inf. Syst., 14(1), January 2023

    work page 2023

  3. [3]

    A review of time-series forecasting algorithms for industrial manufacturing systems

    Syeda Sitara Wishal Fatima and Afshin Rahimi. A review of time-series forecasting algorithms for industrial manufacturing systems. Machines, 12(6), 2024

    work page 2024

  4. [4]

    Studies on time series applications in environmental sciences

    Alina B˘arbulescu. Studies on time series applications in environmental sciences. In Intelligent Systems Reference Library, 2016

    work page 2016

  5. [5]

    Review of time series traffic forecasting methods

    Linkai Wang, Jing Chen, Wei Wang, Ruizhuo Song, Zhaochong Zhang, and Guowei Yang. Review of time series traffic forecasting methods. In 2022 4th International Conference on Control and Robotics (ICCR), pages 1–5, 2022

    work page 2022

  6. [6]

    Forecasting e-commerce consumer returns: a systematic literature review

    David Karl. Forecasting e-commerce consumer returns: a systematic literature review. Management Review Quarterly, May 2024

    work page 2024

  7. [7]

    Time Series Analysis for Education: Methods, Applications, and Future Directions

    Shengzhong Mao, Chaoli Zhang, Yichi Song, Jindong Wang, Xiao-Jun Zeng, Zenglin Xu, and Qingsong Wen. Time series analysis for education: Methods, applications, and future directions. arXiv preprint arXiv:2408.13960, 2024

    work page Pith review arXiv 2024

  8. [8]

    A survey of deep learning and foundation models for time series forecasting

    John A Miller, Mohammed Aldosari, Farah Saeed, Nasid Habib Barna, Subas Rana, I Budak Arpinar, and Ninghao Liu. A survey of deep learning and foundation models for time series forecasting. arXiv preprint arXiv:2401.13912, 2024

    arXiv 2024

Show all 46 references
  1. [9]

    Time-series forecasting with deep learning: a survey

    Bryan Lim and Stefan Zohren. Time-series forecasting with deep learning: a survey. Philosophical Trans- actions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 379(2194):20200209, 2021

    2021

  2. [10]

    Torres, Dalil Hadjout, Abderrazak Sebaa, Francisco Martínez-Álvarez, and Alicia Troncoso

    José F. Torres, Dalil Hadjout, Abderrazak Sebaa, Francisco Martínez-Álvarez, and Alicia Troncoso. Deep learning for time series forecasting: A survey. Big Data, 9(1):3–21, 2021. PMID: 33275484

    2021

  3. [11]

    Probabilistic forecasting

    Tilmann Gneiting and Matthias Katzfuss. Probabilistic forecasting. Annual Review of Statistics and Its Application, 1(V olume 1, 2014):125–151, 2014

    2014

  4. [12]

    Bazionis and Pavlos S

    Ioannis K. Bazionis and Pavlos S. Georgilakis. Review of deterministic and probabilistic wind power forecasting: Models, methods, and future research. Electricity, 2(1):13–47, 2021. 10

    2021

  5. [13]

    A review of probabilistic forecasting and prediction with machine learning

    Hristos Tyralis and Georgia Papacharalampous. A review of probabilistic forecasting and prediction with machine learning. arXiv preprint arXiv:2209.08307, 2022

    2022 arXiv

  6. [14]

    Diffusion models for time-series applications: a survey

    Lequan Lin, Zhengkun Li, Ruikun Li, Xuliang Li, and Junbin Gao. Diffusion models for time-series applications: a survey. Frontiers of Information Technology & Electronic Engineering, 25(1):19–41, Jan 2024

    2024

  7. [15]

    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. Advances in Neural Information Processing Systems, 36:2834...

    2023

  8. [16]

    The rise of diffusion models in time-series forecasting

    Caspar Meijer and Lydia Y Chen. The rise of diffusion models in time-series forecasting. arXiv preprint arXiv:2401.03006, 2024

    2024 arXiv

  9. [17]

    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. arXiv preprint arXiv:2211.14730, 2022

    2022 arXiv

  10. [18]

    Revisiting long-term time series forecasting: An investigation on linear mapping

    Zhe Li, Shiyi Qi, Yiduo Li, and Zenglin Xu. Revisiting long-term time series forecasting: An investigation on linear mapping. arXiv preprint arXiv:2305.10721, 2023

    2023 arXiv

  11. [19]

    An analysis of linear time series forecasting models

    William Toner and Luke Darlow. An analysis of linear time series forecasting models. arXiv preprint arXiv:2403.14587, 2024

    2024 arXiv

  12. [20]

    Multivariate probabilistic time series forecasting via conditioned normalizing flows

    Kashif Rasul, Abdul-Saboor Sheikh, Ingmar Schuster, Urs Bergmann, and Roland V ollgraf. Multivariate probabilistic time series forecasting via conditioned normalizing flows. arXiv preprint arXiv:2002.06103, 2020

    2002 arXiv

  13. [21]

    Is mamba effective for time series forecasting? Neurocomputing, 619:129178, 2025

    Zihan Wang, Fanheng Kong, Shi Feng, Ming Wang, Xiaocui Yang, Han Zhao, Daling Wang, and Yifei Zhang. Is mamba effective for time series forecasting? Neurocomputing, 619:129178, 2025

    2025

  14. [22]

    itrans- former: Inverted transformers are effective for time series forecasting

    Yong Liu, Tengge Hu, Haoran Zhang, Haixu Wu, Shiyu Wang, Lintao Ma, and Mingsheng Long. itrans- former: Inverted transformers are effective for time series forecasting. arXiv preprint arXiv:2310.06625, 2023

    2023 arXiv

  15. [23]

    Ada-mshyper: Adaptive multi-scale hypergraph transformer for time series forecasting

    Zongjiang Shang, Ling Chen, Dongliang Cui, et al. Ada-mshyper: Adaptive multi-scale hypergraph transformer for time series forecasting. arXiv preprint arXiv:2410.23992, 2024

    2024 arXiv

  16. [24]

    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. In The Twelfth International Conference on Learning Representations, 2024

    2024

  17. [25]

    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. In The Thirteenth International Conference on Learning Representations, 2025

    2025

  18. [26]

    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. Advances in Neural Information Processing Systems, 35:5775–5787, 2022

    2022

  19. [27]

    Mvg-crps: A robust loss function for multivariate probabilistic forecasting

    Vincent Zhihao Zheng and Lijun Sun. Mvg-crps: A robust loss function for multivariate probabilistic forecasting. arXiv preprint arXiv:2410.09133, 2024

    2024 arXiv

  20. [28]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020

    2010 arXiv

  21. [29]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020

    2020

  22. [30]

    StatProof- Book/StatProofBook.github.io: StatProofBook 2024, 2025

    Joram Soch, The Book of Statistical Proofs, Karahan Sarıta¸ s, Maja, Pietro Monticone, Thomas J Faulkenberry, Osvaldo A Martin, Alex Kipnis, Salvador Balkus, lfkdlfdlk, Carsten Allefeld, Heiner Atze, Adam Knapp, Ciarán D McInerney, Lo4ding, Valeriu Ohan, amvosk, and maxgrozo. ...

    2024

  23. [31]

    Matheson and Robert L

    James E. Matheson and Robert L. Winkler. Scoring rules for continuous probability distributions. Manage- ment Science, 22(10):1087–1096, 1976

    1976

  24. [32]

    Simple and scalable predictive uncertainty estimation using deep ensembles

    Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. In I. Guyon, U. V on Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Informatio...

    2017

  25. [33]

    Accurate uncertainties for deep learning using calibrated regression

    V olodymyr Kuleshov, Nathan Fenner, and Stefano Ermon. Accurate uncertainties for deep learning using calibrated regression. In International conference on machine learning, pages 2796–2804. PMLR, 2018

    2018

  26. [34]

    Nominality score conditioned time series anomaly detection by point/sequential reconstruction

    Chih-Yu (Andrew) Lai, Fan-Keng Sun, Zhengqi Gao, Jeffrey H Lang, and Duane Boning. Nominality score conditioned time series anomaly detection by point/sequential reconstruction. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors, Advances in Neural...

    2023

  27. [35]

    Conformal prediction with temporal quantile adjustments

    Zhen Lin, Shubhendu Trivedi, and Jimeng Sun. Conformal prediction with temporal quantile adjustments. Advances in Neural Information Processing Systems, 35:31017–31030, 2022

    2022

  28. [36]

    Conformalized quantile regression

    Yaniv Romano, Evan Patterson, and Emmanuel Candes. Conformalized quantile regression. Advances in neural information processing systems, 32, 2019

    2019

  29. [37]

    Gaussian error linear units (gelus)

    Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016

    2016 arXiv

  30. [38]

    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. In The 41st international ACM SIGIR conference on research & development in information retrieval, pages 95–104, 2018

    2018

  31. [39]

    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

  32. [40]

    Non-autoregressive conditional diffusion models for time series prediction

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

    2023

  33. [41]

    Timefilter: Patch-specific spatial-temporal graph filtration for time series forecasting

    Yifan Hu, Guibin Zhang, Peiyuan Liu, Disen Lan, Naiqi Li, Dawei Cheng, Tao Dai, Shu-Tao Xia, and Shirui Pan. Timefilter: Patch-specific spatial-temporal graph filtration for time series forecasting. arXiv preprint arXiv:2501.13041, 2025

    2025 arXiv

  34. [42]

    Diffusion-based time series imputation and forecasting with structured state space models

    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

    2022 arXiv

  35. [43]

    Tactis-2: Better, faster, simpler attentional copulas for multivariate time series

    Arjun Ashok, Étienne Marcotte, Valentina Zantedeschi, Nicolas Chapados, and Alexandre Drouin. Tactis-2: Better, faster, simpler attentional copulas for multivariate time series. arXiv preprint arXiv:2310.01327, 2023

    2023 arXiv

  36. [44]

    Chronos: Learning the language of time series

    Abdul Fatir Ansari, Lorenzo Stella, Caner Turkmen, Xiyuan Zhang, Pedro Mercado, Huibin Shen, Olek- sandr Shchur, Syama Sundar Rangapuram, Sebastian Pineda Arango, Shubham Kapoor, et al. Chronos: Learning the language of time series. arXiv preprint arXiv:2403.07815, 2024

    2024 arXiv

  37. [45]

    Difusco: Graph-based diffusion solvers for combinatorial optimization

    Zhiqing Sun and Yiming Yang. Difusco: Graph-based diffusion solvers for combinatorial optimization. Advances in neural information processing systems, 36:3706–3731, 2023. 12 A Probabilistic Time Series Forecasting Time Series Forecasting (TSF) aims to predict future values of ...

    2023

  38. [46]

    (20) It is well known that the MSE is sensitive to outliers. The Continuous Ranked Probability Score (CRPS) is a statistical measure used to evaluate the accuracy of probabilistic forecasts for continuous variables by comparing the cumulative distribution function (CDF) of the...

    2011

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.