acceptodds
Under review as a conference paper at ICLR 2027

QFLOW: Deterministic Quantile Transport for Probabilistic Time Series Forecasting

Abstract

Recent diffusion forecasters have shown that long-horizon prediction can be reduced to a linear-time recursion over physical time, with noise and scale schedules estimated directly from data. However, these deterministic forecasts do not by themselves provide a principled way to turn the state evolution into a predictive distribution, while resorting to Monte Carlo sampling would reintroduce much of the cost that efficient diffusion forecasting seeks to avoid. We present QFLOW, which extends such forecasters with deterministic quantile transport for probabilistic time series forecasting. We use the Monge–Ampère equation to relate probability density to the local stretching and compression of a vector quantile map. A diagonal approximation of this relation then yields a one-step recursion that transports an entire quantile curve under the same probability-flow field as the underlying state update. At inference, QFLOW advances a dense transport grid, centers the resulting curve on a separately estimated conditional mean, and computes probabilistic forecast metrics directly from its quantiles, without sample ensembles. Across nine long-horizon forecasting datasets, QFLOW ranks among the top three recent diffusion forecasters in 33 of 36 dataset–metric comparisons. It achieves this performance using only a temporal linear layer for the conditional-mean head and either a temporal linear layer or a single cross-attention block for the denoiser.

Primary area:
learning on time series and dynamical systems
Keywords:
Score-based generative models, Diffusion models, Stochastic Differential Equations, Quantiles, Probabilistic time-series forecasting

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_ZHN7OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Building on TEDM's deterministic probability-flow forecasting, the paper starts from the Monge–Ampère relation to move the quantile map along the probability-flow velocity and, via a diagonal Jacobian approximation, obtains a per-coordinate update of quantile spacings, thereby producing, without Monte Carlo sampling, a full quantile curve centered on the characteristic line and shifted onto an independently trained linear mean head, together with a bound on the diagonal-approximation error and a decomposition of the W1W_1 error.

Critical strengths
  1. An elegant line of derivation that proposes a new "deterministic distribution propagation" mechanism. Using the facts that "the probability flow preserves marginal densities" and "the derivative of the quantile function is the reciprocal of the density," the authors obtain a logarithmic update of spacings whose exponential form guarantees positivity and therefore, by construction, ordered quantiles. The idea is natural and concise, and it offers a physical intuition: where the velocity field increases with position, adjacent quantiles are pulled apart. In the direction of sampling-free distributional forecasting that can be queried at arbitrary quantile levels.
  2. Honest statement of the theoretical boundaries. The authors clearly distinguish "transport error" (relative to the marginals of the learned flow) from "model error" (the learned flow relative to the data), explain when the diagonal approximation is exact (separable dynamics plus a diagonal initial map), and acknowledge that Theorem A.1 is only a conditional consistency statement, not a convergence guarantee for the training algorithm.
Critical weaknesses
  1. The diagonal approximation and the "location + shape" splicing undermine the method's theoretical premise: the deployed distribution is neither the marginal of the learned probability flow nor a solution of the Monge–Ampère relation.
    • The method's starting point is that "the quantile map moves with the probability flow and thus preserves the density relation." But the denoiser couples time positions and variables (via cross-attention or shared temporal linear layers). As stated in Appendix A.4, the correct velocity for a coordinate's marginal requires conditional averaging over the other coordinates (vˉt,i\bar v_{t,i}), whereas QFLOW uses the velocity evaluated at the other coordinates of a single characteristic line and keeps only the diagonal of the Jacobian. The transported curve is therefore not, in principle, the marginal quantile of the learned flow, and is correct only in the separable special case. Frameworks like "vector quantile" and "Monge–Ampère" serve a heuristic role in the derivation but are not actually preserved by the method.
    • Location and shape come from two models unaware of each other: the shape is transported centered on the characteristic xtx_t and then shifted wholesale onto the linear head mψm_\psi. This implicitly assumes that the shape of the conditional distribution is independent of the conditional mean, i.e., a pure location family. For heteroscedastic series, or series whose skewness varies with level, this assumption does not hold, and the paper does not discuss this modeling limitation.
    • Variance recalibration (Eq. 22) uses α=1\alpha=1 on most datasets, forcing the curve's variance to match the empirical variance of the rolling window; this is further combined with derivative clipping, quantile clipping, and sorting. This means the final dispersion is mainly determined by heuristic estimates, and the transport equation only determines "the shape up to scaling." The authors themselves acknowledge that these corrections alter the reference dynamics and are thus counted as approximation error. Hence there is a substantial disconnect between the paper's derivation and the deployed forecaster.
  2. The theoretical results place weak constraints on the method, and both the increment over TEDM and the necessity relative to simpler alternatives are insufficiently argued.
    • Prop. 3.1 is essentially the triangle inequality plus "centering at most doubles W1W_1." Theorem A.1 is a Grönwall-type bound with worst-case amplification factors such as exp⁡(LH)\exp(LH) and exp⁡(LˉH)\exp(\bar L H); the velocity discrepancies δt\delta_t and γt\gamma_t in the bound cannot be estimated, and no rates or testable conditions are given. It neither guides design nor explains empirical performance. The authors themselves call it a "conditional consistency statement."
    • The Jacobian Frobenius penalty also penalizes the diagonal entries, which are exactly the quantities the transport relies on. This regularizer is in tension with the method's goal: it pushes ϕt\phi_t toward a constant determined solely by the schedules, causing the shape to degenerate toward a near-Gaussian. Yet the paper describes it as "controlling the neglected coupling" without discussing this side effect.
    • The underlying linear-drift SDE already gives the transition distribution N(stx0,st2Σt)\mathcal{N}(s_tx_0, s_t^2\Sigma_t), and QFLOW uses these empirical schedules for variance recalibration. What, then, does the transport provide beyond the nearly cost-free alternative of "a linear mean head plus schedule-given Gaussian quantiles"? The paper lacks a principled argument.
    • The paper's core motivation is "avoiding sampling cost," but each step requires K⋅NGLK\cdot N_{GL} batched forward passes plus one VJP, which is not necessarily cheaper than few-step sampling; this is not analyzed either.
Questions for authors

Clarifications (answers may affect my rating):

  1. Since the denoiser is coupled, in what sense does QFLOW's scalar curve approximate "the marginal quantiles of the learned flow"? Can you characterize when the diagonal-approximation error is negligible (e.g., a testable condition based on the off-diagonal fraction of the denoiser Jacobian), and relate it to the strength of the Jacobian penalty?
  2. Since the Jacobian Frobenius penalty also shrinks the diagonal entries, does it weaken the effect of the transport itself? Why not penalize only the off-diagonal part (e.g., by subtracting the diagonal estimate from the probe)?
  3. "Shifting the shape onto mψm_\psi" assumes the conditional distribution is a pure location family. How does this assumption affect the method when skewness or dispersion varies with level? Could the shape be made to depend on location?
  4. Given that the SDE transition kernel and the empirical schedules already provide the variance, what exactly does the transport contribute, under α=1, relative to "schedule-given Gaussian quantiles"? Please explain in principle and provide the corresponding simple control.
  5. Do the terms in Theorem A.1 admit estimable upper bounds? Could you give an explicit rate, even in the linear-Gaussian special case, to show that the bound is not vacuous?

Suggestions:

  • A brief experimental suggestion: add wall-clock inference time comparisons with TEDM and NsDiff to support the "avoiding sampling cost" motivation, and explain what the result that linear quantile regression is uniformly better implies for the positioning of the method.
  • Consider block-diagonal or low-rank off-diagonal corrections to partially recover the coupling.
1 comment

Authors have addressed my concerns and i decide to raise my score.

Reviewer_yui4OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes QFLOW, a deterministic quantile transport framework for probabilistic time series forecasting. Its core idea is to propagate quantile curves through a physical-time probability flow, enabling the construction of marginal predictive distributions without Monte Carlo sampling. The authors derive a deterministic quantile update rule from the vector-quantile Monge–Ampère relation using a diagonal approximation. Furthermore, QFLOW adopts a location–shape decomposition, in which a lightweight temporal linear layer estimates the conditional mean, while the transported centered shape determines the predictive uncertainty. The paper also establishes a Wasserstein-distance error bound and introduces Jacobian regularization to stabilize derivative estimates and discourage excessive coordinate coupling. Experiments on nine forecasting benchmarks under the NsDiff protocol show that QFLOW achieves competitive performance against six diffusion-based forecasting methods. However, its advantages over simpler non-diffusion probabilistic forecasting baselines remain unclear, and the claimed computational benefits are not sufficiently validated through end-to-end efficiency comparisons.

Critical strengths

1.A novel deterministic quantile transport framework: Existing diffusion-based forecasting methods often face a trade-off between computationally expensive sampling and efficient deterministic state evolution that does not directly provide a predictive distribution. QFLOW, by leveraging Monge-Ampère optimal transport and Jacobian diagonal approximation, achieves deterministic quantile transport, with the reasoning complexity being O(H), providing a new paradigm for online probabilistic prediction of long time series.

2.Empirical validation of the transport mechanism: The authors not only conducted benchmark comparisons, but also sampled from the same learned probability flow to obtain empirical marginal quantiles, and compared them with the quantile curves after QFLOW transmission. On five datasets, the centralized Wasserstein discrepancy was only 3.1–6.0% of the reference distribution, and the relative CRPS difference was less than 0.5%. This directly indicates that scalar transport approximation can well reproduce the marginal distribution of the learned flow.

Critical weaknesses

1.Insufficient evidence for the effectiveness of quantile transport: Although QFLOW achieves competitive results against diffusion-based models, its proposed quantile transport mechanism does not consistently outperform simpler alternatives. As shown in Section 4.3 and Appendix C.3, a fixed-residual quantile baseline achieves lower CRPS on five of the nine datasets; furthermore, direct linear quantile regression performs better than QFLOW across all nine datasets. These results raise questions about whether the proposed transport mechanism provides sufficient empirical benefits over simpler alternatives to justify its additional computational complexity.

2.Lack of empirical validation of computational efficiency: One of the core motivations of QFLOW is to reduce the computational cost of probability prediction by avoiding Monte Carlo sampling. However, this method requires numerical quadrature and vector–Jacobian products for derivative estimation at each prediction step, introducing additional computational overhead. Although the paper indicates that this method has a linear complexity compared to the prediction time domain, it does not provide a comparison with existing diffusion-based methods in terms of end-to-end running time or memory usage. Therefore, the practical efficiency advantages of QFLOW remain insufficiently substantiated.

Questions for authors

1.The fixed residual quantiles method outperformed QFLOW on all five datasets, and direct linear quantile regression achieved lower CRPS on all nine datasets. What evidence is there to support the necessity of the proposed quantile transmission mechanism in practical applications? Can the author clarify the advantages of this mechanism over simpler alternatives?

2.A central motivation of QFLOW is to avoid the computational overhead of Monte Carlo sampling. However, the method requires numerical quadrature and Jacobian-vector products at each forecasting step. Can the authors provide end-to-end inference time and memory comparisons with representative diffusion-based baselines under comparable hardware and evaluation settings?

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

The paper builds on a physical-time probability-flow diffusion forecaster, where one Euler step of the probability-flow ODE advances the forecast by one time step. It adds a deterministic way to propagate a full predictive quantile curve. Experiments on 9 datasets are conducted to compare against 6 diffusion forecasters.

Critical strengths

The diagonal Monge–Ampère reduction gives a cheap recursion. It keeps quantiles ordered, can be queried at any level without retraining, and its approximation error is written out explicitly

Critical weaknesses

The paper's own baselines show that the transported shape, the core contribution, does not beat trivial uncertainty models. The headline gains come from the linear location head.

Table 6: linear quantile regression beats QFLOW in CRPS on all nine datasets and in QICE on all nine.

Fixed, context-independent validation residuals added to the same mean head beat transport in CRPS on five of nine datasets: ETTm1, ETTm2, ILI, Solar, Traffic.

Table 5: the location swap dominates. For example, Solar CRPS drops from 0.773 (characteristic) to 0.304 (learned head), Traffic from 0.443 to 0.284. The MSE/MAE half of the "33/36" count (Table 2) measures only the temporal-linear head, which is essentially a DLinear-style predictor. Linear models are known to be strong on these benchmarks, so it is not surprising that it beats diffusion baselines on point metrics.

The QR capacity argument (1.07M vs 32–108k parameters) does not apply to the fixed-residual baseline, which adds essentially no parameters. The "new quantile levels without retraining" advantage is shown on only 2 of 5 datasets (Fig. 5), and residual quantiles or spline QR offer the same flexibility. As it stands, the evidence does not show that propagating uncertainty through the denoiser's derivative improves on simple alternatives.

Questions for authors

Please report QFLOW with α=0 and without clipping on all nine datasets, so the raw transport can be evaluated on its own.

Can you give wall-clock inference time and memory against NsDiff and TMDM at H=192, to support the efficiency motivation?

Can you report the main tables with the mean head retrained per seed?

In the fidelity experiment, how do QFLOW and the reference compare before the variance rescaling and clipping are applied?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.