Pith. sign in

REVIEW 3 major objections 4 minor 33 references

NeuTSFlow: Modeling Continuous Functions Behind Time Series Forecasting

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read This paper claims that time series forecasting is best modeled as a continuous flow between probability measures over functions, and reports that a neural-operator velocity field achieves the best average rank across eight benchmarks.

desk verdict A promising function-family framing with a load-bearing inconsistency: the conditional velocity is defined as f−h then redefined as f, so the ODE inference yields h+f and the paper's central claim does not follow as written. read the letter →

arxiv 2507.09888 v2 pith:EMR24XU3 submitted 2025-07-14 cs.LG cs.AIstat.ML

classification cs.LGcs.AIstat.ML
keywords timeseriesforecastingflowmatchingneuraloperatorfunctionspaceprobabilitymeasurecontinuousfunctionssuper-resolutioncross-resolutionlearning
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

This paper argues that a time series is best understood not as a discrete sequence but as a noisy sample from a family of continuous functions, and that forecasting is the task of transporting the probability measure over historical functions to the probability measure over future functions. It introduces NeuTSFlow, a flow-matching model whose velocity field is a neural operator, so that solving an ordinary differential equation in function space carries a historical function into a future function. On eight standard datasets, the authors report the best average rank for conventional forecasting and first place on every dataset for time-series super-resolution and cross-resolution learning. The significance of the claim is that modeling function-level structure, rather than point-by-point dependencies, could be the more faithful and transferable basis for forecasting.

What carries the argument

The central object is the marginal velocity field $u_\theta^t: \mathcal{F} \times [0,1] \to \mathcal{F}$, a time-dependent neural operator that defines the flow ODE $\partial_t \phi_t(g) = u_\theta^t(t, \phi_t(g))$ from a historical function $h$ to a future function $f$. A neural operator is a model that maps between infinite-dimensional function spaces rather than between fixed-length vectors. The velocity field is trained by the conditional flow-matching loss $L_{\text{con}}(\theta) = \mathbb{E}_{t, f \sim \mu_F, h \sim \mu_H} \| v^{f,h}_t(g) - u_\theta^t(g) \|^2$, using the straight-line path $g_t = t f + (1-t)h$, and is implemented with a spectral kernel in the Fourier domain that performs the role of the integral operator $\kappa(y,x)$ in neural-operator theory. This construction is what lets the model transport whole distributions of functions instead of predicting isolated future points.

What would settle it

Check the implementation to see whether inference actually time-marches an ODE through the neural operator or produces the forecast in a single forward pass; if no ODE is solved, the flow-matching stage is not the mechanism behind the reported accuracy.

Watch

Extended reading notes

Core claim

The central claim is that the forecasting map $M: \mu_H \to \mu_F$ can be learned by constructing a conditional path of measures $\mu_{f,h}^t = \mathcal{N}(t f + (1-t)h, C_t)$ between paired historical and future functions $h$ and $f$, and by training a neural operator to reproduce the conditional velocity of that path. The paper shows that as $C_t \to 0$ the conditional velocity is $f - h$, the straight-line transport direction, and then reparameterizes the target so the model regresses directly to $f$. The learned marginal velocity field $u_\theta^t$ is built from normalization, spectral decomposition, dimension expansion, and frequency-domain temporal learning, and Proposition 3.2 states that it is a neural operator and a continuous-time version of DLinear. The paper's empirical claim is that this function-family approach ranks first on all eight datasets for super-resolution and cross-resolution tasks and achieves the best average rank on conventional forecasting.

Load-bearing premise

The method depends on the conditional velocity field being defined consistently during training and inference; Section 3.1 first defines it as $f-h$ and then redefines it as $f$, so the claim that solving the ODE from $h$ yields $f$ needs an unstated conversion step to hold.

Editorial extensions

If this is right

  • If discrete time series are noisy samples of continuous functions, then a model trained at one sampling resolution can predict at another, which is what the paper's super-resolution and cross-resolution experiments show.
  • The claim that $u_\theta^t$ is a continuous-time DLinear connects flow matching to classical linear forecasting, implying that linear models can be read as discretizations of a function-space flow.
  • Because the flow transports a full probability measure over functions, the same framework can in principle output a distribution of plausible futures rather than a single point forecast.
  • Solving the flow ODE from an arbitrary historical function gives a path to extrapolate beyond trained forecast horizons without retraining the model.

Reading between the lines

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

  • The paper leaves implicit how the two conditional velocities in Section 3.1, $f-h$ and $f$, are reconciled during ODE integration; a clarifying extension would be to state that the predicted $f$ is converted to $f-h$ before solving the flow.
  • A testable extension is to run inference both with full ODE integration and with a single direct prediction; if the outputs are identical, the flow-matching stage is not the source of the reported gains.
  • The function-family view suggests a natural probabilistic-forecasting extension: sample several functions from the transported measure $\mu_F$ to obtain calibrated prediction intervals, which the paper does not develop.
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

3 major / 4 minor

Summary. The paper proposes NeuTSFlow, a framework that reframes time series forecasting as learning a transition between probability measures over continuous functions, rather than a mapping between discrete sequences. The method constructs a Gaussian measure path between historical and target functions, defines a conditional velocity field, and learns a neural-operator velocity-field model via a flow-matching loss. The authors claim that the resulting model is a continuous generalization of DLinear and report state-of-the-art results on eight benchmark datasets across three tasks: conventional forecasting, time series super-resolution, and cross-resolution temporal learning. The paper also provides propositions about marginalization, loss equivalence, and function-to-function mapping, and it includes ablation studies purporting to show the importance of the neural operator and flow-matching components.

Significance. If the framework were internally consistent, the function-family perspective and the use of neural operators as velocity fields would be a potentially interesting contribution to time series forecasting. The empirical evaluation is broad, covering three tasks and eight datasets, and the paper makes falsifiable claims about the value of the neural-operator and flow-matching components. However, the central flow-matching construction is inconsistent as written: the conditional velocity is defined in two incompatible ways, and the stated inference procedure does not follow from the equations. Because this inconsistency concerns the core mechanism of the method, the significance of the reported results cannot be assessed without substantial revision. The paper also lacks the appendix containing the training and inference details, which further impedes verification.

major comments (3)
  1. [Section 3.1, Eq. (2) and 'Target Reparameterization for Functional Forecasting'] Eq. (2) defines the conditional velocity as v^{f,h}_t(g) = f - h for the straight-line path g_t = t f + (1 - t) h, but the following paragraph redefines it as v^{f,h}_t(g) = f. These definitions are incompatible. If the learned model u_theta approximates v^{f,h}_t(g) = f, then solving the ODE dg_t/dt = u_theta(t, g_t) from g_0 = h yields g_1 = h + f, not f. The claim in Section 3.2 that 'this process continues until t = 1, at which point f is obtained' is therefore not a consequence of the equations as written. The paper does not state any conversion from the model output f to the ODE velocity f - h during inference. This is a load-bearing issue because the entire forecasting mechanism depends on the velocity field used for integration.
  2. [Section 3.2, Eq. (3)] The 'complete prediction process' is written as a single function evaluation, Y = N^{-1}[u_theta(N(H), t, G_t)], with no ODE solver appearing in the equation, while the prose describes solving an ODE from t = 0 to t = 1. The exact inference recipe is therefore ambiguous. This ambiguity also undermines the ablation labeled 'w/o flow matching' in Table 4: if both the full model and the ablation are evaluated by direct evaluation of u_theta, then the reported contribution of flow matching is not actually measured. The referenced Appendix A.3, which is said to contain the training and inference details, is not present in the reviewed text, so the required procedure cannot be verified.
  3. [Section 3.1, Eq. (6), and Table 4] With the reparameterized velocity v^{f,h}_t(g) = f, the conditional loss in Eq. (6) is exactly a direct regression of the model output u_theta(t, g_t) to the target function f at every time t. Because h is fixed and f is independent of g_t, the optimal model can ignore g_t entirely, and the ODE integration contributes nothing beyond the direct prediction of f. The paper provides no experiment that separates the effect of ODE integration from direct functional regression, so the claim in Table 4 that 'flow matching' is the second-most critical component is not supported by the evidence presented. A consistent formulation would require either regressing on f - h and using that as the ODE velocity, or explicitly deriving how the model output f is converted into the velocity field during inference, and then demonstrating that the integration step adds predictive value beyond a single evaluation.
minor comments (4)
  1. [Throughout] The manuscript contains numerous typos and unfinished edits, for example 'to guaranteeensure', 'align witmatch', and 'the predictedion data' in Section 3.2. These should be corrected.
  2. [Eq. (1)] Eq. (1) writes g_t = mu^{f,h}_t = N(tf + (1 - t)h, C_t), but g_t is subsequently used as a function rather than a measure. This notational conflation is confusing and should be clarified.
  3. [References] Reference [3] is cited as Autoformer but the title given, 'Searching transformers for visual recognition', corresponds to a different paper. The citation and bibliography entry need to be corrected.
  4. [Tables 1-3] The experimental tables report only MSE values without standard deviations or statistical significance tests. Given that NeuTSFlow is claimed to rank first on almost every dataset and task, the absence of variance estimates makes it difficult to assess whether the improvements are meaningful.

Circularity Check

1 steps flagged · score 6.0 of 10

Section 3.1 defines the conditional velocity twice, first as f−h and then as f; with v=f the ODE gives h+f, so the claim that solving the ODE yields f reduces to direct regression by construction.

  1. self definitional [Section 3.1, Eq. (2), 'Target Reparameterization for Functional Forecasting'; Section 3.2, Eq. (3)]
    "lim_{Ct→0} v^{f,h}_t(g) = f − h. (2) ... Thus, by defining v^{f,h}_t(g) = f, the velocity field effectively becomes a mapping that predicts the target sequence as a function f. ... By solving the ODE ∂/∂t φ_t(g) = v_t(φ_t(g)) using an ODE solver, the function g at time t can be determined. This process continues until t = 1, at which point f is obtained. ... The complete prediction process is then formulated as: Y = N^{-1}[u^t_θ(N(H), t,G_t)]. (3)"

    The two definitions are not equivalent for the same path g_t = t f + (1−t)h. Eq. (2) gives the straight-line velocity f−h, whose ODE integration from g_0=h reaches f at t=1. The reparameterization replaces it with v^{f,h}_t(g)=f; integrating dg_t/dt=f from g_0=h gives g_1=h+f, not f. Thus 'until t = 1, at which point f is obtained' does not follow from the equations as written. With v=f, the conditional loss in Eq. (6) regresses u^θ_t to f at every time t, and Eq. (3) evaluates the model in one step with no ODE solver. The learned object is therefore a direct functional regressor of f; the flow-matching transport from h to f adds nothing to the prediction by construction, and the measure-path mechanism claimed in the paper is not realized by these equations.

full rationale

The core circularity is definitional: the paper redefines the conditional velocity field to be the prediction target f, which makes the training loss a direct regression to f and the inference formula a single function evaluation. The ODE-solver narrative is inconsistent with this redefinition, so the claimed flow-matching prediction reduces to direct functional regression. I do not treat the self-citation to rectified flow [16] as load-bearing, since that is an external method and the velocity choice is not justified solely by it. The empirical tables are independent benchmark measurements, so the paper is not wholly vacuous; however, the central mechanistic claim is not supported by the paper's own equations. The missing Appendix A.3 prevents verification of any unstated conversion (e.g., predicting f and then using f−h as the ODE velocity) that might repair the inconsistency, but as written the reduction stands. Score 6 reflects partial circularity: the prediction mechanism reduces by construction, while the benchmark comparisons retain independent empirical content.

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

NeuTSFlow rests on four classes of premises: the generative assumption that discrete series are noisy samples of continuous functions (Section 2), the measure-theoretic machinery of Gaussian conditional paths (Section 3.1), the standard flow-matching gradient equivalence (Proposition 3.3), and the Fourier-sparsity inductive bias in the architecture (Eq. 4). No constants are fitted to produce the theoretical claims; the learnable network weights, K, m_max, and C_t are free or hyper parameters that affect the learned velocity field but are not presented with their values. The central theoretical load is carried by the Gaussian-path construction and the unproved identification of the reparameterized velocity with the ODE velocity.

free parameters (3)
  • K (TopK retained Fourier modes) = not stated in main text
    Controls the trend/seasonal split in Eq. (4); the paper does not report the value used in the experiments, and the decomposition changes the input representation the velocity field sees.
  • m_max (frequency truncation length in the spectral kernel) = not stated in main text
    Limits the number of RFFT modes in Eq. (5); truncating low-frequency modes is the kernel-smoothing mechanism of the neural operator, so performance depends on this hyperparameter.
  • C_t (covariance of the conditional Gaussian path) = unspecified; taken to 0 in the limit
    Eq. (1) introduces C_t, and Eq. (2) depends on the limit C_t to 0, but the covariance's form, scale, and limit rate are never specified, even though they determine the absolute continuity properties the construction relies on.
assumptions (4)
  • domain assumption Historical and future time series are noisy observations of continuous functions h ~ μH and f ~ μF, with a shared underlying probability law.
    This generative assumption is the foundation of the reframing in Section 2 and is not tested; all subsequent flow matching is built on it.
  • standard math The Gaussian measures N(t f + (1 − t) h, C_t) are mutually absolutely continuous on F, so the Radon-Nikodym derivatives and marginal velocity exist.
    Invoked in Section 3.1 with a citation to [21]; the paper does not verify the Cameron-Martin-type conditions for its specific C_t, so the measure-theoretic claims rely on this unstated background result.
  • standard math The conditional flow-matching loss has the same gradient as the marginal loss in function space (Proposition 3.3), extending the finite-dimensional result.
    Proposition 3.3 is stated as a theorem with the proof deferred to Appendix A.2; in the provided text it functions as an axiom.
  • domain assumption A small number of Fourier modes (TopK) is sufficient to separate trend and seasonal structure so that the remaining signal is still forecastable.
    Eq. (4) hard-codes this inductive bias in the architecture; no sensitivity analysis of K is provided, and the assumption affects what the neural operator learns.

how reviews work

0 comments
Cite this review

Pith. "Pith review of NeuTSFlow: Modeling Continuous Functions Behind Time Series Forecasting." pith.science (2026). https://pith.science/paper/EMR24XU3

@misc{pith2026250709888,
  author       = {Pith},
  title        = {Pith review of: NeuTSFlow: Modeling Continuous Functions Behind Time Series Forecasting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EMR24XU3}},
  note         = {Machine review of arXiv:2507.09888}
}
read the original abstract

Time series forecasting is a fundamental task with broad applications, yet conventional methods often treat data as discrete sequences, overlooking their origin as noisy samples of continuous processes. Crucially, discrete noisy observations cannot uniquely determine a continuous function; instead, they correspond to a family of plausible functions. Mathematically, time series can be viewed as noisy observations of a continuous function family governed by a shared probability measure. Thus, the forecasting task can be framed as learning the transition from the historical function family to the future function family. This reframing introduces two key challenges: (1) How can we leverage discrete historical and future observations to learn the relationships between their underlying continuous functions? (2) How can we model the transition path in function space from the historical function family to the future function family? To address these challenges, we propose NeuTSFlow, a novel framework that leverages Neural Operators to facilitate flow matching for learning path of measure between historical and future function families. By parameterizing the velocity field of the flow in infinite-dimensional function spaces, NeuTSFlow moves beyond traditional methods that focus on dependencies at discrete points, directly modeling function-level features instead. Experiments on diverse forecasting tasks demonstrate NeuTSFlow's superior accuracy and robustness, validating the effectiveness of the function-family perspective.

Figures

Figures reproduced from arXiv: 2507.09888 by the authors.

Figure 1
Figure 1. Comparison between traditional time series forecasting and the NeuTSFlow paradigm. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Training overview of NeuTSFlow. The model learns a velocity field u θ t that aligns with the conditional velocity v f,h t between sampled functions h ∼ µH and f ∼ µF along the path of measure µt = tf + (1 − t)h. series and those corresponding to target time series. The related work can be found in Appendix A.1, and the proofs for all propositions are provided in Appendix A.2. 3.1 Framework Construction Given the pro… view at source ↗
Figure 3
Figure 3. From left to right: super-resolution results pro [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 17 canonical work pages

  1. [1]

    M. S. Albergo and E. Vanden-Eijnden. Building normalizing flows with stochastic interpolants,

  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

  3. [3]

    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

  4. [4]

    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

  5. [5]

    S. S. W. Fatima and A. Rahimi. A review of time-series forecasting algorithms for industrial manufacturing systems. Machines, 12(6):380, 2024

  6. [6]

    A. B. Galvão and M. Owyang. Forecasting low-frequency macroeconomic events with high- frequency data. Journal of Applied Econometrics, 37(7):1314–1333, 2022

  7. [7]

    F. Gong, D. Raghunathan, A. Gupta, and M. Apostolaki. Super-resolution on network telemetry time series, 2024. URL https://arxiv.org/abs/2403.04165

  8. [8]

    Kaushik, A

    S. Kaushik, A. Choudhury, P. K. Sheron, N. Dasgupta, S. Natarajan, L. A. Pickett, and V . Dutt. Ai in healthcare: time-series forecasting using statistical, neural, and ensemble architectures. Frontiers in big data, 3:4, 2020

Show all 33 references
  1. [9]

    Kerrigan, G

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

  2. [10]

    T. Kim, J. Kim, Y . Tae, C. Park, J.-H. Choi, and J. Choo. Reversible instance normalization for accurate time-series forecasting against distribution shift. In International Conference on Learn- ing Representations, 2022. URL https://openreview.net/forum?id=cGDAkQo1C0p

  3. [11]

    Z. Kong, W. Ping, J. Huang, K. Zhao, and B. Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. arXiv preprint arXiv:2009.09761, 2020

  4. [12]

    Z. Li, N. Kovachki, K. Azizzadenesheli, B. Liu, K. Bhattacharya, A. Stuart, and A. Anandkumar. Neural operator: Graph kernel network for partial differential equations, 2020. URL https: //arxiv.org/abs/2003.03485

  5. [13]

    Z. Li, N. Kovachki, K. Azizzadenesheli, B. Liu, K. Bhattacharya, A. Stuart, and A. Anandkumar. Fourier neural operator for parametric partial differential equations, 2021. URL https:// arxiv.org/abs/2010.08895

  6. [14]

    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. 10

  7. [15]

    S. Liu, H. Yu, C. Liao, J. Li, W. Lin, A. X. Liu, and S. Dustdar. Pyraformer: Low-complexity pyramidal attention for long-range time series modeling and forecasting. In International conference on learning representations, 2021

  8. [16]

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

  9. [17]

    Y . Liu, H. Wu, J. Wang, and M. Long. Non-stationary transformers: Exploring the stationarity in time series forecasting. Advances in Neural Information Processing Systems, 35:9881–9893, 2022

  10. [19]

    Y . Liu, T. Hu, H. Zhang, H. Wu, S. Wang, L. Ma, and M. Long. itransformer: Inverted transformers are effective for time series forecasting, 2024. URL https://arxiv.org/abs/ 2310.06625

  11. [20]

    L. Lu, P. Jin, G. Pang, Z. Zhang, and G. E. Karniadakis. Learning nonlinear operators via deeponet based on the universal approximation theorem of operators. Nature Machine Intelli- gence, 3(3):218–229, Mar. 2021. ISSN 2522-5839. doi: 10.1038/s42256-021-00302-5. URL http://dx....

  12. [21]

    A. R. Nahmod, L. Rey-Bellet, S. Sheffield, and G. Staffilani. Absolute continuity of gaussian measures under certain gauge transformations. arXiv preprint arXiv:1103.4822

  13. [22]

    Y . Nie, N. H. Nguyen, P. Sinthong, and J. Kalagnanam. A time series is worth 64 words: Long-term forecasting with transformers. arXiv preprint arXiv:2211.14730, 2022

  14. [23]

    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

  15. [24]

    S. Ray, S. S. Das, P. Mishra, and A. M. G. Al Khatib. Time series sarima modelling and forecasting of monthly rainfall and temperature in the south asian countries. Earth Systems and Environment, 5:531–546, 2021

  16. [25]

    Sheoran, S

    S. Sheoran, S. Shukla, S. Pasari, R. S. Singh, and R. Kulshrestha. Wind speed forecasting at different time scales using time series and machine learning models. Applied Solar Energy, 58 (5):708–721, 2022

  17. [26]

    A. Tong, K. Fatras, N. Malkin, G. Huguet, Y . Zhang, J. Rector-Brooks, G. Wolf, and Y . Bengio. Improving and generalizing flow-based generative models with minibatch optimal transport,

  18. [27]

    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, 2024. URL https://arxiv. org/abs/2405.14616

  19. [28]

    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(7976):1089–1100, 2023

  20. [29]

    Yuan and Y

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

  21. [30]

    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, 2023

  22. [31]

    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

  23. [32]

    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. 11

  24. [2023]

    URL https://arxiv.org/abs/2209.15571

  25. [2024]

    URL https://arxiv.org/abs/2302.00482

Pith tools

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