acceptodds
Under review as a conference paper at ICLR 2027

OnART: Online Adaptation of Rectified Flow for Probabilistic Time-Series Forecasting with Feedback Delay

Abstract

Probabilistic time-series forecasting is essential for decisions under uncertainty, yet existing generative forecasters are predominantly fixed after offline training, while online methods largely adapt point predictors without updating predictive uncertainty. We address the resulting problem of adapting an entire predictive distribution from sparse, delayed feedback without destabilizing its learned transport. We introduce OnART, an online adaptation framework that turns Rectified Flow into an efficient delayed-feedback probabilistic forecaster. Offline re-flow straightens the conditional transport and enables one-step generation, limiting the trajectory interference induced by online updates. At deployment, a newly observed residual provides a new endpoint but no source from which to update the learned transport. OnART selects, for each variable, the Gaussian source paired with the nearest forecast-time residual endpoint. This choice exactly minimizes the change from the cached displacement target to the new re-flow displacement target among the stored candidates; online training then intentionally moves the selected source toward the new endpoint. Generated replay anchors the cached forecast-time mappings while this update is applied. The point forecaster learns from all causally available values, and the rectified velocity model learns from the assigned pair and generated replay after the feedback delay. Across seven multivariate benchmarks, OnART consistently improves probabilistic forecasting performance while preserving competitive point accuracy. These results establish minimal-shift source assignment and replay as effective mechanisms for online adaptation of probabilistic forecasts.

Primary area:
learning on time series and dynamical systems
Keywords:
Probabilistic Time Series Forecasting, Online Adaptation, Flow Matching

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_MViAOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The authors consider probabilistic time-series forecasting in non-stationary environments with delayed feedback, where existing online approaches mainly adapt point predictions but leave predictive uncertainty fixed. To address this issue, they propose OnART, an online adaptation framework that updates both the point forecaster and a rectified-flow residual model. The method uses proximity-based source assignment and generated replay to stabilize flow adaptation from delayed residual feedback. The reported results are promising, showing consistently improved probabilistic forecasting performance across seven benchmarks.

Critical strengths

1、 The minimal-shift source-assignment idea is intuitive and well motivated for endpoint-only flow adaptation. 2、 The method updates both the point forecast and predictive uncertainty under a causal predict-then-update protocol. 3、 One-step generation makes the approach substantially faster than iterative diffusion-based probabilistic baselines.

Critical weaknesses

1、 The evaluation stream constitutes 75% of several datasets, and the model continually adapts on labels revealed within this stream. This can be valid for online learning, but the protocol and comparison must be stated much more clearly: are all online baselines given exactly the same delayed, overlapping supervision and the same number of update steps? 2、 The theoretical result only proves that proximity assignment minimizes the displacement-target change among cached candidates under its assumptions. It does not show that this criterion minimizes post-update forecasting error or distributional error after model optimization. 3、 The ablation table omits Traffic, although it is included in the main comparison. This weakens the claim that each component consistently helps across all datasets. 4、 The main comparison is largely against offline probabilistic methods and online deterministic methods. Naive online flow matching is useful, but stronger online probabilistic adaptation baselines would better establish the advantage of the proposed source-assignment mechanism. 5、Why were some datasets (ETTh1 and ETTm2) excluded?

Questions for authors

My main questions and concerns are detailed in the Weaknesses section above. I would appreciate the authors’ responses to these concerns.

Reviewer_ZxG9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes OnART, a framework for online adaptation of probabilistic time-series forecasters under delayed feedback. The forecast is decomposed into a deterministic point prediction (PatchTST) and a residual distribution modeled by a conditional Rectified Flow, which is distilled offline into a one-step Euler map via reflow. During online deployment, a residual observed at time tt (for a forecast issued at t−Ht-H) only gives a target endpoint without a matching source sample. To handle this, OnART caches the KK generated source-endpoint pairs in a FIFO buffer, selects for each channel the cached Gaussian source whose generated endpoint is nearest to the observed residual, and updates the velocity model with a combination of this assigned pair and RR replayed generated pairs, conditioned on a frozen vector-quantized (VQ) codebook.

Critical strengths

Clear problem with a practical design: Adapting the full predictive distribution online under delayed feedback is a relevant problem, since most online forecasters only adapt point predictions. Offline re-flow makes this practical with one-step generation (about 40 ms per step on ETTh2), and proximity-based source assignment is a simple and intuitive way to pick a source for a new residual.

Consistent results: OnART achieves the best average CRPSsum\text{CRPS}_{\text{sum}} on all seven datasets with competitive point accuracy, and the ablations show that each component helps.

Critical weaknesses

Channel-separable model evaluated with a multivariate metric: The velocity model (Eq. 16) and the source assignment (Eq. 14) work per channel, so the generated residuals are conditionally independent across channels. Yet the headline metric is CRPSsum\text{CRPS}_{\text{sum}}, which depends on cross-channel covariance, and this is only acknowledged in one sentence in Appendix D. A joint scoring rule (e.g. Energy Score) or a non-separable baseline is needed. In addition, the main gains are reported against offline baselines. Compared with the "Point Update Only" ablation, the gain from adapting the residual flow, which is the actual contribution, is 4.9% on average (vs 10.7% over Offline), and Point Update Only is even worse than Offline on Exchange and Solar.

Limited theoretical contribution: Theorem 1 essentially reduces to (x1new−x0)−(x^1−x0)=x1new−x^1(\mathbf{x}^{\text{new}}_1 - \mathbf{x}_0) - (\hat{\mathbf{x}}_1 - \mathbf{x}_0) = \mathbf{x}^{\text{new}}_1 - \hat{\mathbf{x}}_1 and compares against a counterfactual loss, so it says little about the stability of the actual update. The paper does not report how often the recomputed and stored conditions differ, or how large εc\varepsilon_c gets under regime shifts. Theorem 2 assumes i.i.d. data and a Lipschitz loss, which do not hold for streaming data with a hard quantizer.

Minor: TMDM and NsDiff are orders of magnitude worse than other baselines without explanation; Traffic is missing from the ablations; OnART is not best at every horizon (e.g. the Offline variant is better on ECL H=96H=96 and Solar H=24H=24); MSE is computed from the ensemble mean, which is only stated in Table 8; memory and latency are only reported on ETTh2, while the buffer is about 6.4 GB on Traffic; Naive Flow Matching exceeds three days on ECL and Solar although it only changes the source pairing; and there are no calibration metrics such as interval coverage.

Questions for authors

Q1 Joint dependence. How does OnART perform under a multivariate scoring rule such as the Energy Score, or against a non-separable velocity model? This is the question that would most affect my assessment.

Q2 Fair online baselines. How does OnART compare to the offline probabilistic baselines when their point forecasters are also updated online?

Q3 Size of the online updates. With online learning rates of 10−710^{-7} to 10−610^{-6}, online adaptation reduces CRPSsum\text{CRPS}_{\text{sum}} on ETTm1 from 2.898 to 1.703. How much do the parameters actually change, and is the Offline variant the same checkpoint that OnART starts from?

Q4 Conditioning drift. How often do the recomputed and stored conditions differ during streaming, and what does the distribution of εc\varepsilon_c look like during regime shifts?

Q5 State the potential impact of the findings of this paper

Reviewer_KB8eOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes OnART, a new online probabilistic forecasting model. OnART has two components: a point prediction model (PatchTST) and flow matching to predict the distribution of the residual. The model works in the following steps:

  1. Train both models with initial time steps.
  2. At time tt, keep the information of the probabilistic model for the last t−Ht-H steps: source samples, their projections, discretized context.
  3. At time tt, a new observation arrives. The ground truth residual is available. Take the source sample that gives the closest residual to the ground truth and retrain both the flow model and point prediction model.

Experimental results show that the proposed OnART is better than the baseline offline evaluation models in terms of MSE and CRPS_Sum. Furthermore, point predictions are also competitive with the baseline online point prediction models,

Critical strengths
  1. There is some research on online point forecasting, but online probabilistic forecasting is under-researched. This paper addresses the gap using the OnART model. The model components are well motivated and fit together coherently.
  2. Experimental results show that OnART has better CRPS-sum and MSE compared to offline probabilistic models and online point models. Ablation studies show the importance of the components, such as online adaptation compared to the offline version.
Critical weaknesses
  1. The core of the paper, Section 2.5, is very hard to follow. It is notation-heavy, with low information density. I recommend that the authors move the polished version of the algorithm provided in the appendix to the main text, which would improve the readability of the paper.
  2. The results are insufficient.
  • The main evaluation metric is the CRPS-sum for comparison with the probabilistic baselines. CRPS-sum has little discriminative ability, as it is computed after aggregating the predictions.
  • CRPS is used only in the ablation, but not to compare with the baselines
Questions for authors
  1. Could the authors report the Energy Score and Variogram instead of the CRPS sum?
  2. Could the authors also compare the baselines in terms of CRPS to understand how marginals are?
  3. The MSEs of TMDM and NSDiff are orders of magnitude worse in the Traffic, ECL, and solar datasets. This looks odd even with the model's default hyperparameters after normalization. I briefly reviewed the TMDM paper, and the MSE values look normal, unlike in this paper. Some sanity checks would be helpful.: normalization and denormalization pipeline, reproducing the baseline resutls and tuning hyperparameters.
  4. Since the online point forecasting models provide better MSE, can they be used instead of patchTST?
  5. A simple baseline could strengthen the paper: take the online point forecasting model and predict a simple Gaussian distribution to make it probabilistic. Since the current baselines are only offline, there is no floor for the online probabilistic models.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.