Bridging the Last Mile of Prediction: Enhancing Time Series Forecasting with Conditional Guided Flow Matching

Huibo Xu; Likang Wu; Qi Liu; Runlong Yu; Xianquan Wang

REVIEW 4 major objections 4 minor 1 cited by

Bridging the Last Mile of Prediction: Enhancing Time Series Forecasting with Conditional Guided Flow Matching

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that time series forecasting improves when a flow-matching model starts from another model's predictions and learns the probabilistic structure of the residuals between those predictions and the true future.

desk verdict A promising flow-matching refinement layer for forecasting, but the empirical claims are not yet verifiable without code, error bars, and the missing Self-Guided Diffusion comparison. read the letter →

arxiv 2507.07192 v3 pith:VORK54TZ submitted 2025-07-09 cs.LG cs.AI

Huibo Xu , Runlong Yu , Likang Wu , Xianquan Wang , Qi Liu This is my paper

classification cs.LGcs.AI

keywords timeseriesforecastingflowmatchingpredictionresidualsgenerativeconditionalprobabilitypathsprobabilisticresidualdistributionlearningnon-crossingflows

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 tries to establish that time series forecasts can be made more accurate by treating the gap between a base model's predictions and the true future as a learnable distribution, not just an error to minimize. Its proposal, Conditional Guided Flow Matching (CGFM), replaces the usual noise prior in flow matching with the output distribution of an auxiliary forecasting model and transports that distribution to the true future distribution, with both ends conditioned on the same history. The framework is model-agnostic, so the same wrapper can be attached to MLP, transformer, or diffusion forecasters, and the experiments report lower MSE and MAE than the base models on nearly all settings. A sympathetic reader would take the central claim to be that residual distributions carry reusable temporal structure, and that flow paths between predictions and truths can be kept unambiguous by conditioning both endpoints on shared history.

What carries the argument

The central object is the two-sided conditional guided probability path $X_t = \alpha_t X_1 + \beta_t X_0$ between the auxiliary model's prediction distribution $p(x_0|h)$ and the true future distribution $q(x_1|h)$, with both endpoints conditioned on the same history $h$ and with $\alpha_t$, $\beta_t$ smooth affine interpolation coefficients. The load-bearing identity is the marginal velocity field $u_t(x|h) = \mathbb{E}[\dot{\alpha}_t X_1 + \dot{\beta}_t X_0 \mid X_t = x, H = h]$, whose $X_1$-prediction reparameterization gives $u_t(x|h) = \frac{\dot{\beta}_t}{\beta_t} x + \left(\dot{\alpha}_t - \frac{\alpha_t \dot{\beta}_t}{\beta_t}\right) \mathbb{E}[X_1 \mid X_t = x, H = h]$. This field is learned by the conditional guided flow matching loss $L_{CGM}$, and the paper uses the conditional independent coupling $\pi_{0,1|H} = p(x_0|h)q(x_1|h)$ to guarantee non-crossing paths.

What would settle it

Run CGFM with an auxiliary predictor that is deliberately given access to future information during preprocessing; if the reported gains persist rather than collapsing, the conditional-independence coupling is not the mechanism driving the improvement.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that the prediction residuals $\epsilon = x_1 - x_0$, where $x_0$ is an auxiliary model's forecast and $x_1$ the observed future, have a probability law that can be learned by a flow-matching model rather than discarded after supervised training. CGFM sets the source to $p(x_0|h)$, smooths it with Gaussian noise so it satisfies the regularity assumptions, and constructs the two-sided conditional path $X_t = \alpha_t X_1 + \beta_t X_0$ with both endpoints drawn conditionally independently from the same history $h$. The marginal velocity field is $u_t(x|h) = \mathbb{E}[\dot{\alpha}_t X_1 + \dot{\beta}_t X_0 \mid X_t = x, H = h]$, and the paper proves this construction equivalently learns the residual's evolution and yields non-crossing paths, so no intermediate flow state can ambiguously point at two different futures. Reparameterizing the objective as direct $X_1$-prediction is shown empirically to be the best of three mathematically equivalent prediction targets. The reported consequence is consistent improvement over each auxiliary model and across seven benchmark datasets.

Load-bearing premise

The load-bearing premise is that the auxiliary model's prediction and the true future are unrelated once the shared history is fixed, plus smoothness and boundedness conditions on the distributions that the paper assumes without testing on real data.

Editorial extensions

If this is right

Any existing forecasting model can be wrapped by CGFM and refined without changing the base model, with the largest reported gains when the base model is RLinear.
Because CGFM learns the distribution of residuals rather than minimizing point error, systematic bias and variance in the auxiliary predictions are corrected rather than just averaged away.
Conditioning both endpoints on the same history gives non-crossing flow paths, so intermediate sampling states remain unambiguously tied to one future trajectory, reducing information loss during generation.
Using an auxiliary model's prediction distribution as the source shrinks the distance the flow must travel, reducing learning difficulty relative to a Gaussian prior.
The $X_1$-prediction reparameterization of the guided flow matching loss is the empirically preferred objective, despite being mathematically equivalent to the other two forms.

Reading between the lines

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

Not in the paper but testable: if residual structure drives the gains, CGFM's advantage over plain flow matching should grow as residuals become more biased, heteroscedastic, or multimodal; synthetic experiments with controlled residual shapes could isolate the mechanism.
The same two-sided conditional coupling could be carried to other conditional generation settings where a cheap predictor and a shared context exist, such as image restoration or speech enhancement, with the predictor's output serving as the source.
Because only MSE and MAE are reported, the forecasts' distributional quality is unmeasured; if the residual-distribution claim is right, probabilistic metrics such as CRPS should also improve, which would be a direct check of the paper's thesis.

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.

Desk Editor's Note

read the letter

Colleague,

The short version: CGFM is a sensible extension of flow matching for forecasting, and the residual-learning framing is genuinely useful. But the paper currently undersells its own evidence: no code, no error bars, an inconsistent pair of tables, a dangling 'Experiment .' reference, and the closest prior method is cited but never compared. I'd send it to review, but with a request for major revision.

What's new: using the auxiliary model's predictive distribution as the source in flow matching, conditioning both source and target on history, and the affine-path family with a non-crossing argument. The equivalence to learning residual distributions is clean and gives a nice interpretation. The theoretical propositions are mostly standard flow-matching facts; the assumptions are mild, and the Gaussian smoothing step addresses the smoothness requirement. That part is fine.

The experiments show consistent gains over the base models in Table 1, which is a good sign that the method works as advertised. I also appreciate the breadth of datasets and baselines. But the reporting is not up to standard. Ten runs are averaged with no standard deviations or significance tests, so the gains could be noise. More worrying, Table 2 does not match Table 1 for the same settings, and without code I can't tell whether it's a different horizon or a mistake. The biggest omission is the lack of a comparison to Self-Guided Diffusion (Kollovieh et al. 2024a), which also refines auxiliary model predictions via residual learning. The paper cites it, so the authors know it; not benchmarking against it makes the 'outperforms state-of-the-art' claim incomplete.

The theoretical assumptions are not the problem here. The conditional-independence coupling is a design choice, and the smoothness conditions are handled by construction. The real threat to the headline is the missing empirical evidence.

Who should read this: anyone working on generative forecasting or using flow matching as a post-hoc refinement layer. It's a plausible generic component. But before I cite it, I need the code and a clean, reproducible evaluation.

Recommendation: accept for peer review with major revision. The idea is solid enough to deserve referee time, but the current write-up is not ready for acceptance.

Referee Report

4 major / 4 minor

Summary. The paper proposes Conditional Guided Flow Matching (CGFM), a model-agnostic framework for multivariate time series forecasting. Instead of starting from a Gaussian prior, CGFM initializes the flow from the output of an auxiliary predictive model, conditions the probability path and velocity field on the observed history h, and uses general affine interpolants X_t = α_t X1 + β_t X0. The authors argue that this formulation learns the probabilistic structure of prediction residuals and prove several propositions (noise smoothing, marginal velocity representation, non-crossing, and equality of loss gradients). Empirically, the paper reports MSE/MAE on seven datasets with four auxiliary models, claiming consistent improvement over state-of-the-art baselines, together with ablations on path parameterizations and prediction targets.

Significance. If the empirical claims held, CGFM would be a useful and elegant contribution: it is simple, model-agnostic, and using the auxiliary prediction as the source distribution is a natural way to reduce the distributional gap in flow-matching forecasting. The two-sided conditioning and the flexibility of affine paths are reasonable design choices. However, the manuscript's own evidence does not currently establish the headline claim. There are no uncertainty estimates despite 10 repeated runs, the closest prior residual-learning baseline is cited but not compared, and Table 2 cannot be reconciled with Table 1. The theoretical propositions are mostly in the spirit of standard flow matching, but the residual-equivalence and non-crossing claims require stronger justification. The core idea is promising, and the gaps are addressable, so I view this as a revision rather than a rejection.

major comments (4)

[Experiment / Evaluation Metrics; Table 1] The paper states that each experiment was repeated 10 times and the outcomes averaged, but Table 1 reports only point estimates with no standard deviations, confidence intervals, or significance tests. Many improvements are small (e.g., ETTm2 96: Rlinear 0.182 vs. Rlinear+CGFM 0.167; ETTh1 720: iTransformer 0.506 vs. iTransformer+CGFM 0.511, a degradation), so the reader cannot tell whether the claimed consistent gains exceed run-to-run noise. Please report variability measures and paired tests, or temper the claim of consistent improvement.
[Table 2] Table 2's CGFM row cannot be reconciled with Table 1. For ETTh1 it reports 0.373, but none of the four CGFM variants in Table 1 equals this at any horizon (96: 0.363, 0.368, 0.371, 0.365; horizon means: 0.414, 0.429, 0.421, 0.422). Additionally, 'Avg Rank (1)' is computed from mean ranks, which masks per-dataset consistency; CGFM is ranked third on Weather and on Exchange. Please state exactly which configuration and horizon(s) Table 2 refers to, and provide per-dataset rank distributions or per-horizon comparisons.
[Related Work / Baselines] Kollovieh et al. (2024a), self-guided diffusion, is arguably the closest prior method: it also takes an auxiliary predictor's output and learns to refine residuals. It is cited in the related work but never included in the experiments. Without this comparison (or a documented justification for its exclusion), the claim of outperforming state-of-the-art models is incomplete and the novelty claim is not fully differentiated.
[Propositions 0.2 and 0.5; Eq. (6), Eq. (3)] The theoretical contribution is not fully supported. Proposition 0.2 says CGFM 'equivalently learns residual ϵ=x1−x0', but for a general affine path (Eq. 6) X_t = α_t X1 + β_t X0 = α_t ϵ + (α_t + β_t) X0, so the path is coupled to X0 unless α_t + β_t = 1; the paper's parameterizations (e.g., VP, cosine) do not satisfy this identity, and no proof of equivalence is given. Proposition 0.5 asserts non-crossing from conditional independent coupling, but conditioning both endpoints on the same h does not prevent two different pairs (x0,x1) drawn from p(·|h) and q(·|h) from crossing; non-crossing of the marginal ODE requires a uniqueness or Lipschitz argument for the velocity field in Eq. (9), which is not supplied. Proposition 0.4 also assumes bounded support of q, which is not checked or discussed for the benchmark datasets.

minor comments (4)

[Proposition 0.5] There is a typo in the statement: 'there exist not ∈ [0,1)' should read 'there exists no t ∈ [0,1)'.
[Parameterization of the Prediction Target] The sentence 'We further investigate the effects of different parameterizations of α_t and β_t in Experiment .' contains a dangling reference; the actual appendix section should be cited.
[Table 1] Several entries have inconsistent decimal places or missing digits (e.g., 0.48 in the ETTh1 336 row), and the eight repeated metric-pair columns make the table hard to read; consider splitting into separate tables or using a multi-panel layout.
[Figure 4 and reproducibility] The claim that RLinear predictions are 'closely aligned with the ground truth' based on the PCA visualization is qualitative and would benefit from a quantitative smoothness or alignment metric. No code or data splits are provided, so the exact experimental setup cannot be reproduced from the text alone.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: CGFM's derivation is self-contained; residual equivalence is a path-definition restatement, and the empirical gaps are verification issues, not circular reasoning.

full rationale

The paper's construction sets X0 to the auxiliary model's output ("Choice of X0") and defines the two-sided conditional affine paths via Eq. (6), Xt = αtX1 + βtX0. Proposition 0.2 then observes that this is equivalent to learning the residual ε = X1 − X0; this is a mathematical restatement of the chosen path definition, not a fitted result that is recycled as a prediction. Propositions 0.3, 0.4, and 0.6 are standard conditional flow matching identities with stated smoothness assumptions; they do not import the paper's empirical conclusion into the premises. No parameter is fitted to a subset of data and then presented as a prediction of a closely related quantity, and the reference list contains no load-bearing self-citation by the present authors. The unresolved concerns—no error bars, no released code, the missing comparison with self-guided diffusion, and the unverified distributional assumptions on benchmark data—are evidence-quality and correctness issues, not circularity. Even if Proposition 0.5's non-crossing guarantee is too strong, that would be a validity failure, not a circular reduction. The central empirical claim remains externally testable, so the derivation chain is not circular.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical or statistical entities; it relies on existing flow matching machinery and a chosen auxiliary model. The main free choices are the noise scale, the path schedule, and the prediction objective.

free parameters (3)

noise scale sigma = not reported
Gaussian smoothing parameter added to the auxiliary model output (Algorithm 1, line 7). Its value affects the smoothness of the source distribution and is a free hyperparameter.
path schedule alpha_t, beta_t = Poly-n chosen as best in Table 3
The affine path is parameterized by alpha_t and beta_t; different schedules (CondOT, Poly-n, VP, Cosine) are tested and Poly-n is selected. The degree n is a free parameter.
prediction objective gt = X1-prediction
Three equivalent parameterizations (ut, X0, X1) are tested; X1-prediction is chosen based on Table 4, but this choice is an empirical selection.

assumptions (4)

domain assumption Target distribution q has bounded support and source p is C^1 with strictly positive density and finite second moments
Proposition 0.4 relies on this; the paper does not verify it for real datasets, instead invoking noise smoothing to make p smooth.
domain assumption Conditional independent coupling pi_{0,1|H}(x0,x1|h) = p(x0|h) q(x1|h)
Assumed in Eq. (3); it requires that the auxiliary model's prediction and the future are independent given history, which may not hold if the model's errors depend on the input.
domain assumption Auxiliary model Phi is differentiable
Invoked in the analysis of p(x0|h) before Proposition 0.1.
ad hoc to paper Non-crossing guarantee (Proposition 0.5)
Claimed to follow from the affine path and conditional coupling; the proof is not given in the main text and the result appears to be a standard property of well-defined ODE flows.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Bridging the Last Mile of Prediction: Enhancing Time Series Forecasting with Conditional Guided Flow Matching." pith.science (2026). https://pith.science/paper/VORK54TZ

@misc{pith2026250707192,
  author       = {Pith},
  title        = {Pith review of: Bridging the Last Mile of Prediction: Enhancing Time Series Forecasting with Conditional Guided Flow Matching},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VORK54TZ}},
  note         = {Machine review of arXiv:2507.07192}
}

read the original abstract

Existing generative models for time series forecasting often transform simple priors (typically Gaussian) into complex data distributions. However, their sampling initialization, independent of historical data, hinders the capture of temporal dependencies, limiting predictive accuracy. They also treat residuals merely as optimization targets, ignoring that residuals often exhibit meaningful patterns like systematic biases or nontrivial distributional structures. To address these, we propose Conditional Guided Flow Matching (CGFM), a novel model-agnostic framework that extends flow matching by integrating outputs from an auxiliary predictive model. This enables learning from the probabilistic structure of prediction residuals, leveraging the auxiliary model's prediction distribution as a source to reduce learning difficulty and refine forecasts. CGFM incorporates historical data as both conditions and guidance, uses two-sided conditional paths (with source and target conditioned on the same history), and employs affine paths to expand the path space, avoiding path crossing without complex mechanisms, preserving temporal consistency, and strengthening distribution alignment. Experiments across datasets and baselines show CGFM consistently outperforms state-of-the-art models, advancing forecasting.

Figures

Figures reproduced from arXiv: 2507.07192 by the authors.

**Figure 2.** Visualization of CGFM training. The ideal predictor maps the history distribution to the target distribution. However, [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Illustration of Proposition 0.5. Left: Independent [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: PCA visualization of predictions and ground truth, [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 5.** Figure 5: Comparison of results: Left (before CGFM) and [PITH_FULL_IMAGE:figures/full_fig_p007_5.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. Full citation record

Enhancing Irregular Time Series Forecasting with Continuous-Time Modeling Framework
cs.LG 2026-07 conditional novelty 5.0 of 10

WrapFlow combines continuous-time event/gap tokenization with simulation-free residual flow matching on a Transformer to improve irregular multivariate time-series forecasting.

Reference graph

Works this paper leans on

46 extracted references · 24 canonical work pages · cited by 1 Pith paper

[1]

M. S. Albergo and E. Vanden-Eijnden. Building normalizing flows with stochastic interpolants, 2023. URL https://arxiv.org/abs/2209.15571

arXiv 2023
[2]

M. S. Albergo, N. M. Boffi, and E. Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions, 2023. URL https://arxiv.org/abs/2303.08797

arXiv 2023
[3]

J. M. L. Alcaraz and N. Strodthoff. Diffusion-based time series imputation and forecasting with structured state space models. arXiv preprint arXiv:2208.09399, 2022

arXiv 2022
[4]

M. Chen, H. Peng, J. Fu, and H. Ling. Autoformer: Searching transformers for visual recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12270--12280, 2021

work page 2021
[5]

P. Chen, Y. ZHANG, Y. Cheng, Y. Shu, Y. Wang, Q. Wen, B. Yang, and C. Guo. Pathformer: Multi-scale transformers with adaptive pathways for time series forecasting. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=lJkOCMP2aW

work page 2024
[6]

R. T. Q. Chen, Y. Rubanova, J. Bettencourt, and D. Duvenaud. Neural ordinary differential equations, 2019. URL https://arxiv.org/abs/1806.07366

arXiv 2019
[7]

A. Das, W. Kong, A. Leach, S. Mathur, R. Sen, and R. Yu. Long-term forecasting with tide: Time-series dense encoder, 2024. URL https://arxiv.org/abs/2304.08424

arXiv 2024
[8]

C. Doersch. Tutorial on variational autoencoders, 2021. URL https://arxiv.org/abs/1606.05908

arXiv 2021

Show all 46 references

[9]

Y. Du, J. Wang, W. Feng, S. Pan, T. Qin, R. Xu, and C. Wang. Adarnn: Adaptive learning and forecasting of time series. In Proceedings of the 30th ACM international conference on information & knowledge management, pages 402--411, 2021

2021
[10]

Esser, S

P. Esser, S. Kulal, A. Blattmann, R. Entezari, J. Müller, H. Saini, Y. Levi, D. Lorenz, A. Sauer, F. Boesel, D. Podell, T. Dockhorn, Z. English, K. Lacey, A. Goodwin, Y. Marek, and R. Rombach. Scaling rectified flow transformers for high-resolution image synthesis, 2024. URL h...

2024 arXiv
[11]

Franceschi, A

J.-Y. Franceschi, A. Dieuleveut, and M. Jaggi. Unsupervised scalable representation learning for multivariate time series. Advances in neural information processing systems, 32, 2019

2019
[12]

I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio. Generative adversarial networks, 2014. URL https://arxiv.org/abs/1406.2661

2014 arXiv
[13]

J. Ho, A. Jain, and P. Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33: 0 6840--6851, 2020

2020
[14]

V. T. Hu, D. Wu, Y. M. Asano, P. Mettes, B. Fernando, B. Ommer, and C. Snoek. Flow matching for conditional text generation in a few sampling steps. In EACL (2), pages 380--392, 2024. URL https://aclanthology.org/2024.eacl-short.33

2024
[15]

Kerrigan, G

G. Kerrigan, G. Migliorini, and P. Smyth. Functional flow matching, 2023. URL https://arxiv.org/abs/2305.17209

2023 arXiv
[16]

Kollovieh, A

M. Kollovieh, A. F. Ansari, M. Bohlke-Schneider, J. Zschiegner, H. Wang, and Y. B. Wang. Predict, refine, synthesize: Self-guiding diffusion models for probabilistic time series forecasting. Advances in Neural Information Processing Systems, 36, 2024 a

2024
[17]

Kollovieh, M

M. Kollovieh, M. Lienen, D. Lüdke, L. Schwinn, and S. Günnemann. Flow matching with gaussian process priors for probabilistic time series forecasting, 2024 b . URL https://arxiv.org/abs/2410.03024

2024 arXiv
[18]

Y. Li, W. Chen, X. Hu, B. Chen, M. Zhou, et al. Transformer-modulated diffusion models for probabilistic multivariate time series forecasting. In The Twelfth International Conference on Learning Representations
[19]

Y. Li, X. Lu, Y. Wang, and D. Dou. Generative time series forecasting with diffusion, denoise, and disentanglement. Advances in Neural Information Processing Systems, 35: 0 23009--23022, 2022

2022
[20]

Z. Li, S. Qi, Y. Li, and Z. Xu. Revisiting long-term time series forecasting: An investigation on linear mapping, 2023. URL https://arxiv.org/abs/2305.10721

2023 arXiv
[21]

Lim and S

B. Lim and S. Zohren. Time-series forecasting with deep learning: a survey. Philosophical Transactions of the Royal Society A, 379 0 (2194): 0 20200209, 2021

2021
[22]

Lipman, R

Y. Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le. Flow matching for generative modeling, 2023. URL https://arxiv.org/abs/2210.02747

2023 arXiv
[23]

M. Liu, A. Zeng, M. Chen, Z. Xu, Q. Lai, L. Ma, and Q. Xu. Scinet: Time series modeling and forecasting with sample convolution and interaction. Advances in Neural Information Processing Systems, 35: 0 5816--5828, 2022 a

2022
[24]

X. Liu, C. Gong, and Q. Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow, 2022 b . URL https://arxiv.org/abs/2209.03003

2022 arXiv
[25]

Y. Liu, T. Hu, H. Zhang, H. Wu, S. Wang, L. Ma, and M. Long. itransformer: Inverted transformers are effective for time series forecasting. arXiv preprint arXiv:2310.06625, 2023

2023 arXiv
[26]

Y. Liu, T. Hu, H. Zhang, H. Wu, S. Wang, L. Ma, and M. Long. itransformer: Inverted transformers are effective for time series forecasting. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=JePfAI8fah

2024
[27]

Y. Nie, N. H. Nguyen, P. Sinthong, and J. Kalagnanam. A time series is worth 64 words: Long-term forecasting with transformers, 2023. URL https://arxiv.org/abs/2211.14730

2023 arXiv
[28]

Pooladian, H

A.-A. Pooladian, H. Ben-Hamu, C. Domingo-Enrich, B. Amos, Y. Lipman, and R. T. Q. Chen. Multisample flow matching: Straightening flows with minibatch couplings, 2023. URL https://arxiv.org/abs/2304.14772

2023 arXiv
[29]

Rasul, C

K. Rasul, C. Seward, I. Schuster, and R. Vollgraf. Autoregressive denoising diffusion models for multivariate probabilistic time series forecasting. In International Conference on Machine Learning, pages 8857--8868. PMLR, 2021

2021
[30]

D. J. Rezende and S. Mohamed. Variational inference with normalizing flows, 2016. URL https://arxiv.org/abs/1505.05770

2016 arXiv
[31]

Salinas, V

D. Salinas, V. Flunkert, J. Gasthaus, and T. Januschowski. Deepar: Probabilistic forecasting with autoregressive recurrent networks. International journal of forecasting, 36 0 (3): 0 1181--1191, 2020

2020
[32]

Shen and J

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

2023
[33]

L. Shen, W. Chen, and J. Kwok. Multi-resolution diffusion models for time series forecasting. In The Twelfth International Conference on Learning Representations, 2024

2024
[34]

Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole. Score-based generative modeling through stochastic differential equations, 2021. URL https://arxiv.org/abs/2011.13456

2021 arXiv
[35]

Tashiro, J

Y. Tashiro, J. Song, Y. Song, and S. Ermon. Csdi: Conditional score-based diffusion models for probabilistic time series imputation. Advances in Neural Information Processing Systems, 34: 0 24804--24816, 2021

2021
[36]

H. Wang, J. Peng, F. Huang, J. Wang, J. Chen, and Y. Xiao. Micn: Multi-scale local and global context modeling for long-term series forecasting. In The eleventh international conference on learning representations, 2023

2023
[37]

S. Wang, H. Wu, X. Shi, T. Hu, H. Luo, L. Ma, J. Y. Zhang, and J. Zhou. Timemixer: Decomposable multiscale mixing for time series forecasting. arXiv preprint arXiv:2405.14616, 2024

2024 arXiv
[38]

J. L. Watson, D. Juergens, N. R. Bennett, B. L. Trippe, J. Yim, H. E. Eisenach, W. Ahern, A. J. Borst, R. J. Ragotte, L. F. Milles, et al. De novo design of protein structure and function with rfdiffusion. Nature, 620 0 (7976): 0 1089--1100, 2023

2023
[39]

H. Wu, T. Hu, Y. Liu, H. Zhou, J. Wang, and M. Long. Timesnet: Temporal 2d-variation modeling for general time series analysis. arXiv preprint arXiv:2210.02186, 2022

2022 arXiv
[40]

K. Yi, Q. Zhang, W. Fan, S. Wang, P. Wang, H. He, N. An, D. Lian, L. Cao, and Z. Niu. Frequency-domain mlps are more effective learners in time series forecasting. Advances in Neural Information Processing Systems, 36, 2024

2024
[41]

J. Yoon, D. Jarrett, and M. Van der Schaar. Time-series generative adversarial networks. Advances in neural information processing systems, 32, 2019

2019
[42]

Yuan and Y

X. Yuan and Y. Qiao. Diffusion-ts: Interpretable diffusion for general time series generation. arXiv preprint arXiv:2403.01742, 2024

2024 arXiv
[43]

A. Zeng, M. Chen, L. Zhang, and Q. Xu. Are transformers effective for time series forecasting? In Proceedings of the AAAI conference on artificial intelligence, volume 37, pages 11121--11128
[44]

Zhang, Y

X. Zhang, Y. Pu, Y. Kawamura, A. Loza, Y. Bengio, D. Shung, and A. Tong. Trajectory flow matching with applications to clinical time series modelling. NeurIPS, 2024

2024
[45]

H. Zhou, S. Zhang, J. Peng, S. Zhang, J. Li, H. Xiong, and W. 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
[46]

T. Zhou, Z. Ma, Q. Wen, X. Wang, L. Sun, and R. Jin. Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting. In International conference on machine learning, pages 27268--27286. PMLR, 2022

2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

M. S. Albergo and E. Vanden-Eijnden. Building normalizing flows with stochastic interpolants, 2023. URL https://arxiv.org/abs/2209.15571

arXiv 2023

[2] [2]

M. S. Albergo, N. M. Boffi, and E. Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions, 2023. URL https://arxiv.org/abs/2303.08797

arXiv 2023

[3] [3]

J. M. L. Alcaraz and N. Strodthoff. Diffusion-based time series imputation and forecasting with structured state space models. arXiv preprint arXiv:2208.09399, 2022

arXiv 2022

[4] [4]

M. Chen, H. Peng, J. Fu, and H. Ling. Autoformer: Searching transformers for visual recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12270--12280, 2021

work page 2021

[5] [5]

P. Chen, Y. ZHANG, Y. Cheng, Y. Shu, Y. Wang, Q. Wen, B. Yang, and C. Guo. Pathformer: Multi-scale transformers with adaptive pathways for time series forecasting. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=lJkOCMP2aW

work page 2024

[6] [6]

R. T. Q. Chen, Y. Rubanova, J. Bettencourt, and D. Duvenaud. Neural ordinary differential equations, 2019. URL https://arxiv.org/abs/1806.07366

arXiv 2019

[7] [7]

A. Das, W. Kong, A. Leach, S. Mathur, R. Sen, and R. Yu. Long-term forecasting with tide: Time-series dense encoder, 2024. URL https://arxiv.org/abs/2304.08424

arXiv 2024

[8] [8]

C. Doersch. Tutorial on variational autoencoders, 2021. URL https://arxiv.org/abs/1606.05908

arXiv 2021