acceptodds
Under review as a conference paper at ICLR 2027

Flow Matching in Hierarchical Residual Coordinates for Non-Stationary Probabilistic Forecasting

Abstract

Probabilistic forecasting for non-stationary time series is difficult because the conditional target distribution entangles drifting location-scale statistics with residual stochastic dynamics. This entanglement forces generative models to learn both macroscopic distribution drift across contexts and microscopic stochastic dynamics within each context in a single transport map. We propose a novel extensible framework of hierarchical residual coordinates: an order-indexed composition of context-dependent invertible maps removes predictable distributional components from the full future trajectory before generation, a residual generator learns the remaining joint law while retaining temporal and cross-variate dependence, and the inverse coordinate map transports the learned residual flow back to reconstruct forecasts in the original space. We establish a conditional pushforward consistency guarantee for this construction. As a practical second-order instantiation, LS-Flow combines location-scale residualization with Conditional Flow Matching. Across eight real-world non-stationary multivariate benchmarks, LS-Flow consistently improves probabilistic forecasting over state-of-the-art generative baselines; order-wise ablations and a third-order pilot examine the framework's extensibility.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series, Non-Stationary, Probabilistic Forecasting, Flow Matching

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_d91mOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The authors propose that lower order moments hinder generative models from modelling stochastically drifting higher order statistics. They show that decoupling the lower- and higher order moments is possible without conflicting pushforward consistency during flow matching. Empirical results show improvements over chosen baselines when separately modeling the first two moments using an iTransformer and MLPs and the residual using conditional flow matching.

Critical strengths
  • The paper demonstrates a clear improvement on modeling lower and higher order statistics (Figure 2b).
  • The method is simple, fast, and shows good performance against the chosen baselines.
Critical weaknesses

While the central idea is intuitive, the paper hides that behind complicated and dense phrasing.

  • The introduction uses non-standard terminology compared to the field, but makes no attempt at introducing it, or guiding the reader into the terminology used throughout.
  • Section 4.3 seems disconnected from the rest of the paper and has no clear embedding in the story.
  • A discussion of the method and its limitations is only partially included, but would help to put the method into perspective.

The categories of macroscopic drifts and microscopic stochasic dynamics seems strange. The distinction suggests differences in their drifting behaviour, but is claimed not to be different in lines 144/145. What are then the differences and why are those differences important for the method at all?

Section 4.3: The proof sketch is really hard to parse given that the terminology is poorly introduced, and a full proof would be greatly beneficial.

Minor:

  • Lines 298, 301: WpW_p is not introduced before its use. Connection to Wasserstein distance can be made only because it is mentioned once in line 74.
  • Eq.3/6 : ⊘\oslash is used in Eq. 3, but only introduced after Eq. 6, ⊙\odot is not introduced at all
  • Line 122/123: refering to the shape of a distribution as its topology is imprecise
  • Given the main contribution of this work, I would expect a more detailed description of how to train the network predicting the mean and scale without consulting the codebase.
  • Formatting below table 2
  • Formatting appendix F - it appears to be empty
  • Line 826: D$^3$U is not bold
Questions for authors
  • Theorem 1 begins with "For almost every context x", but never clarifies which contexts are not covered by the theorem. Is this a typo, or can you clarify for which contexts this theorem holds?
Reviewer_TSSAOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes LS-Flow, a probabilistic time-series forecasting framework that predicts future location and scale, applies Conditional Flow Matching in a standardized residual space, and reconstructs forecasts through an inverse coordinate transformation. The authors provide theoretical results on conditional transformations and evaluate the method on eight real-world benchmarks.

Critical strengths
  1. Strong empirical performance across multiple benchmarks. LS-Flow achieves competitive probabilistic forecasting performance across eight real-world datasets, outperforming existing generative baselines on most benchmarks in terms of CRPS and CRPSsum. The results suggest that incorporating predicted location-scale statistics into generative forecasting can be practically beneficial.

  2. Intuitive and effective integration of location-scale decomposition with Flow Matching. The proposed framework provides a straightforward way to separate predictable location-scale variations from residual stochastic dynamics. The use of invertible transformations preserves the joint predictive distribution, including temporal and cross-variate dependencies. The ablation studies further demonstrate that the decomposition can improve forecasting performance over alternative configurations.

Critical weaknesses
  1. Limited methodological and theoretical novelty. The proposed framework primarily combines existing techniques, including future location-scale estimation, residual-based generative modeling, and Conditional Flow Matching. Related approaches such as D3U and NsDiff have already explored separating deterministic components from residual uncertainty or incorporating location-scale information into probabilistic forecasting. Furthermore, the main theoretical results in Section 4.3 largely follow from standard change-of-variables identities and the Lipschitz stability of Wasserstein distances. In particular, for the affine transformation used in LS-Flow, the proposed vector-field pushforward is mathematically equivalent to integrating the flow in residual space and applying inverse normalization afterward. Consequently, the extent to which the paper introduces a fundamentally new methodological or theoretical contribution remains unclear.

  2. Insufficient attribution of empirical improvements to the proposed framework. Although LS-Flow achieves strong forecasting performance, the experiments do not sufficiently isolate the contributions of the deterministic backbone, predictive Gaussian prior, location-scale decomposition, and Flow Matching. In Table 13, the Gaussian prior alone achieves an ECL CRPS of 0.197, already outperforming all reported baselines, while adding Flow Matching improves CRPS to 0.189. Furthermore, Table 12 shows that replacing iTransformer with DLinear substantially reduces performance on several datasets, suggesting considerable dependence on backbone quality. While Flow Matching provides meaningful improvements in some settings, the reported results do not fully establish which components account for the overall superiority of LS-Flow. Additionally, the Wasserstein-distance comparisons in Figure 2(c) evaluate distributions in differently scaled coordinate spaces, making it difficult to attribute the smaller distances directly to improved transport geometry or optimization.

Questions for authors
  1. How much of LS-Flow's performance improvement is attributable to Flow Matching rather than the backbone and predictive Gaussian prior (Tables 12–13)?

  2. What additional benefit does the affine vector-field pushforward provide over simple inverse normalization, and how can the Wasserstein comparisons in Figure 2(c) be justified across differently scaled spaces?

Reviewer_oRHwOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a flow matching model for time series forecasting built on a hierarchical decomposition. First, the location and scale components are removed, and the generative model then operates in a residual space to better capture non-stationary time series. In an empirical evaluation, the authors demonstrate that they outperform selected diffusion-based models.

Critical strengths
  • The appendix has a lot of analysis and additional information. Also, the code is attached, allowing us to look up unclear details.
Critical weaknesses
  • The literature research is not up to date. There is no paper from this year and barely any from last year. Many time series generative models are missing from the related work section and in the empirical evaluation, including diffusion and flow matching works.
  • The model itself is not particularly novel and is mostly a reparametrization and normalization. I do not think that the insights gained by the work are sufficient for acceptance.
  • The empirical evaluation does not answer how LS-Flow compares to other standard normalization techniques. Also, it is unclear on what kind of data, e.g., stationary vs. non-stationary, the model performs best.
  • It is really hard to read and understand the paper, given that the method results in a simple model. Especially, the methodology section is hard to follow. Also, there are some layout violations, e.g., table 1.
Questions for authors

See weaknesses, and could you also elaborate on how/whether the selected datasets are non-stationary?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.