acceptodds
Under review as a conference paper at ICLR 2027

Residual Jump Flow Matching for Extreme-Event Time-Series Generation

Abstract

In linear flow matching, the large increments that define time-series extremes share one transport target with background variation and can be underrepresented by a finite velocity model. We introduce Residual Jump Flow Matching (RJFM), centered on a cumulative-residual probability path that separates large increments from a base trajectory and transports their persistent effect later in generative time. The path preserves both endpoints while changing the regression geometry; we quantify its maximum displacement from linear interpolation and its additional kinetic action. A structured temporal velocity with base and cumulative-increment branches provides an architectural refinement of this path design. Across nine real time series, RJFM improves tail and extreme-increment fidelity over capacity-matched flow baselines and diffusion models.

Primary area:
generative models
Keywords:
Flow Matching, Time-Series Generation, Extreme Events, Heavy-Tailed Data, Residual Probability Paths

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Mo45OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes Residual Jump Flow Matching (RJFM) for improving extreme-event fidelity in time-series generation. It extracts large increments using a training-distribution threshold, accumulates them into a residual, and transports this residual with a delayed schedule while transporting the remaining signal linearly. An optional two-branch temporal network complements the path construction. The paper establishes endpoint and path-geometry properties and evaluates the method against flow-matching controls and diffusion baselines. The motivation is compelling, and the cumulative treatment of increments is well aligned with their effect on sequence values. However, the justification for delayed transport remains incomplete, the practical sampler differs from the field supervised during training, and the evidence against capacity-matched FM controls is less consistent than the broad extreme-event claims suggest.

Critical strengths

1 The paper distinguishes a localized increment from its persistent effect on subsequent sequence values. This provides a clear motivation for a cumulative residual representation.

2 The proposed path preserves the conditional endpoints, recovers linear interpolation, and admits explicit displacement and kinetic-action calculations. Parameter-matched FM and path-only controls help separate capacity, path, and architectural effects. Chronological splits, training-only preprocessing, ten-seed evaluation, and longer Diffusion-TS training are also valuable.

Critical weaknesses

1 A global magnitude threshold cannot distinguish persistent shifts, transient spikes, and noise. Robustness to threshold and gate-temperature choices is insufficiently demonstrated.

2 Training supervises the sum of two non-identifiable branches, but sampling rescales one branch. The theoretical transport guarantee does not directly cover this modified field.

3 Gains over matched FM are inconsistent across primary metrics, and RJFM underperforms on Energy. Winner-count figures obscure effect sizes and uncertainty.

Questions for authors

1 How sensitive is performance to the threshold, gate temperature, and different extreme-event types?

2 What evidence shows that delayed residual transport improves learning, and what explains the weaker ES-95 and Energy results?

Reviewer_6fpoOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a flow-matching framework for improving extreme-event fidelity in time-series generation. The central observation is that a large increment is localized in the difference sequence but can induce a persistent level shift in the original sequence. Based on this observation, the authors decompose a target time series into a base component and a cumulative residual component, and construct a conditional probability path. The authors further introduce a structured temporal velocity network with separate branches for the base and cumulative-increment components. Experiments suggest that the proposed approach improves several extreme-event fidelity metrics over flow-matching baselines.

My expertise is primarily in generative modeling and flow matching, rather than time-series generation. Therefore, my assessment focuses mainly on the methodological novelty, technical soundness, and the extent to which the experiments support the proposed methodological claims. I am less familiar with the standard benchmarks and baseline methods commonly used in the time-series generation literature. Consequently, I am not in the best position to make a definitive judgment about the completeness of the baseline selection or the relative standing of the proposed method within this specific literature.

Critical strengths
  1. The paper addresses the important challenge of generating extreme events in time series. Its key idea is to modify the flow-matching trajectory using a residual representation that emphasizes large temporal increments. This provides an intuitive and interpretable way to incorporate extreme-event structure into the generative process.

  2. The proposed method retains the basic flow-matching training framework while introducing a modified interpolation path and a dual-branch temporal convolutional network. This design is relatively straightforward and does not require a separate generative modeling framework, making the approach conceptually appealing.

Critical weaknesses

It is unclear whether the performance gains primarily come from the proposed trajectory design or the additional inference-time scaling of the jump branch ((\beta=0.5)). An ablation study comparing different (\beta) values, particularly (\beta=1.0), is needed to disentangle their respective contributions.

Questions for authors

NA

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

This paper proposes Residual Jump Flow Matching for generating fixed-length time-series windows with better tail and extreme-increment fidelity. Increments of a training window above a training-set quantile are softly gated and cumulatively summed into a step-shaped residual R, and the remainder B = x1 − R is the base. The conditional path moves B on the linear schedule t and R on t^α with α > 1 (Eq. 4–5), so large increments enter the regression target later while both endpoints are preserved. Prop. 1 gives the maximum displacement from the linear path and the extra kinetic action. A two-head temporal-convolution network is trained on the summed velocity and sampled with a tempered weight β on the increment head.

RJFM is compared with linear-path TCN-FM at two widths, a path-only control, DLPM and Diffusion-TS over ten seeds, on nine real univariate series plus the Stock and Energy benchmarks. The authors conclude that the path is the main source of the improved tail and extreme-increment fidelity.

Critical strengths
  • A simple idea that is easy to adopt. RJFM is a data-dependent time schedule for one cumulative component of the target. It keeps the standard FM regression loss and an ODE sampler, and the closed forms in Prop. 1 are correct.
  • A relevant problem and sensible controls. Tail fidelity in time-series generation matters. The study separates path from capacity with a width-matched linear FM and a path-only model, and uses chronological splits with gaps, training-only normalization and threshold, ten seeds, and paired seed-level intervals.
  • Candid reporting. The paper shows where RJFM does not lead: Energy and DLPM on some MMD statistics. It also says the cumulative head is an inductive bias, not an identified jump estimator.
Critical weaknesses

1. The central claim, that the path accounts for most of the gain, is not supported by the paper's own comparisons. The path-only control (Residual-Path80) is compared only with RJFM (Table 2), never with the linear controls, yet that comparison is what the claims in Sec. 1, Sec. 4.3 and the Conclusion need. The only direct view is Fig. 2(a), where plain TCN-FM64 has more wins (26) than the path-only model (19) over 108 series–metric units. On Stock and Energy (Table 3) the path-only model is within noise of the linear control. Against the flow controls, RJFM has the lower mean ES-95 on only 4/9 series each (no separated interval), and on Energy (Table 3) it is worse than both flow controls on all six metrics. Only tail-99 favours RJFM (8/9 series against each linear control), and the interval separates on only 3 and 2 of them. The claim of better extreme-increment fidelity over capacity-matched flow baselines is therefore not supported.

2. The sampler behind the headline results is not the learned flow. Eq. (7) trains the summed field u^B + u^J on the conditional target, but Eq. (8) samples with u^B + β u^J at a validation-selected β = 0.5 (Sec. 3.3), whereas the population argument of Sec. 3.4 applies to β = 1. Even with exact heads, β = 0.5 would end at B + 0.5R rather than x1, and since only the sum is supervised, the branch split is not identified (as Sec. 3.3 says). β is thus an architecture-dependent knob that can shrink tails, and none of the controls has an analogue. The main comparison should be shown at β = 1, with signed tail errors (the reported errors are unsigned), so readers can tell whether the gain comes from the path or from sampler-side shrinkage.

3. The external baselines are weak, so the large margins say little. DLPM's tail-99 error is roughly 40–50 on every series in Fig. 2(b), more than an order of magnitude above all flow models, which suggests a diverged sampler rather than a competitive heavy-tailed baseline; its α grid {1.5, 1.8, 1.9} also omits the Gaussian limit α = 2 (DDPM). Diffusion-TS is trained for 1,500 updates at batch size 256 (Sec. 4.1), against 10,000 and 25,000 steps at batch size 64 in its original recipe (Yuan & Qiao, 2024, Table 8), and Fig. 3(a) itself shows the gap closing as the budget grows (25/27 primary-metric wins at 1,500 updates, 22/27 at 3,000; Sec. 4.4). In Table 3 the discriminative score is about 0.50 for every model, which, if it is the usual |accuracy − 0.5| score, is its ceiling, whereas Diffusion-TS reports about 0.07 on Stocks and 0.12 on Energy (Yuan & Qiao, 2024, Table 1); Table 3 therefore cannot support sample-quality claims. Closely related methods are also missing: FlowTS (Hu et al., 2024) is a rectified-flow time-series generator evaluated on Stock, Tail-GAN (Cont et al., 2022) trains generators to match tail-risk statistics such as VaR and ES, and a difference-domain FM baseline (generate increments, then cumulate) would show whether the cumulative path adds anything beyond modelling increments.

4. The motivation and mechanism are asserted rather than argued. Sec. 1 says population FM is not limited and a finite network must learn the rare shift, but a sharper known limitation is the source distribution: a Lipschitz transport map applied to a Gaussian source cannot produce heavier tails (Jaini et al., 2020), and Gaussian-prior diffusion and flow-matching models are reported to miss heavy-tailed behaviour, which motivates Student-t priors (Pandey et al., 2025). RJFM keeps a Gaussian source and changes only the schedule, so the paper should say which failure it fixes (increment amplitude or tail index) and compare with a heavy-tailed source. Likewise, under a linear schedule a large step has a high signal-to-noise ratio and is usually resolved early; RJFM delays exactly this component, yet Sec. 3 gives no argument for why delaying, rather than advancing, should help. Prop. 1 is an elementary calculation of distance from the straight path, and Sec. 3.4 does not link the delay to learnability.

Questions for authors
  1. Are the DLPM samples diverging (tail-99 error of roughly 40–50), and what happens at α = 2 (DDPM)?
  2. Does the advantage over Diffusion-TS hold at its published budget of 10,000 to 25,000 steps? How are the discriminative scores in Table 3 computed, and which windows serve as the real set?
  3. Why delay the step component rather than advance it? Does RJFM change the tail index or only the increment amplitude, and how does it compare with a Student-t source ?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.