acceptodds
Withdrawn from ICLR 2027

FlowOC: History-Informed Flow Matching with Optimal Control Calibration for Probabilistic Forecasting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Generative models provide a natural framework for probabilistic multivariate time series forecasting by modeling the conditional distribution of future trajectories rather than producing a single deterministic estimate. However, improving the alignment of generated forecasts with future observations while preserving meaningful predictive variability remains challenging. To address this issue, FlowOC adopts a two-stage training procedure. In the first stage, Conditional Flow Matching learns the predictive distribution of future trajectories from informative historical-patch source states. In the second stage, the pretrained flow is frozen, and an additive controller performs optimal control-inspired calibration using ensemble-level accuracy and distribution-shape preservation objectives, thereby improving the predictive center while maintaining the stochastic structure of the generated trajectories. Experiments on six multivariate forecasting benchmarks demonstrate that FlowOC improves point forecasting accuracy while preserving or enhancing probabilistic forecasting quality, thereby achieving a balance between forecast accuracy and faithful uncertainty characterization. Extensive ablation and qualitative analysis further validate the effectiveness of the historical-patch source distribution and the proposed optimal control calibration strategy. Code is available at https://anonymous.4open.science/r/Flow_OC-E8DA.

Primary area:
learning on time series and dynamical systems
Keywords:
Flow matching, Time series forecasting, Optimal control

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_c5NNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a method for probabilistic forecasting, that uses flow-matching forecasting to model the conditional distribution of future trajectories. It does this via two stages: conditional flow matching to learn the predictive distribution of future trajectories, followed by a calibration step, where an optimal control inspired term is used to improve the forecasts. For the first step, an iTransformer-style variate encoder is used, to capture cross channel dependencies. The method uses historical patches, small sampled sequences used as input for the flow-matching forecaster (instead of pure Gaussian samples). This is then trained in the same style as Conditional Flow Matching (CFM). For the second stage, an optimal control problem is formulated by creating an ensemble of K trajectories from different initial conditions, and integrating their flows. The control problem is not solved directly, but a neural controller is trained for gradient matching on the CRPS gradient. This then steers the ensemble members to solutions that improve the quality of the predictive distribution, while penalizing control cost. The proposed method is evaluated on the standard real-world forecasting datasets, and ablates the historical patches and the effect of the calibration step.

Critical strengths

Promising direction: Connecting optimal control-based calibration of probabilistic forecasting is an interesting approach that I could see make a positive impact. Probabilistic forecasting is useful in many real-world forecasting tasks, where the additional probabilistic information is preferred over a deterministic prediction. Providing calibration for these types of models with a similar method to VGG Flow makes for an interesting direction that can improve probabilistic forecasting.

Critical weaknesses
  1. Limited Novelty: The two main novelty claims rely on redesigning of the sampling strategy of flow matching, and the calibration of the model via optimal control refinement strategy. For the first, many of the proposed contributions are already existing concepts: conditional flow matching already introduces the training objective, using transformers for the velocity field is common practice, and the history context encoding follows the iTransformer style of variate encoder. While these are interesting ideas, their specific combination does not form a major contribution on its own. For the second claimed novelty regarding optimal control, the formulation of the optimal control on the particles is quite closely related to VGG Flow (Liu et al., 2025), and also seems to closely align with the OC-Flow method ("Training Free Guided Flow Matching With Optimal control" by Wang et al., ICLR 2025). VGG Flow learns a value gradient, and OC-Flow uses co-states from Pontryagin minimum Principle. While neither of these works exactly address the ensemble setting in forecasting, the novelty remains limited.

  2. Lack of ensemble theory: For a set of ensemble members, the ensemble optimality condition should ideally be applied to the entire set of particles/samples, but currently, each ensemble member has its own optimal control applied without considering the state of the other members. From a theoretical perspective, this seems to be best fitted as a mean field optimal control problem, but it is not linked to this. For probabilistic forecasting, this would concretely mean that the control should consider the joint states of all the ensemble members. It would be good to clarify what the calibrated distribution looks like.

  3. Experimental: lack of error bars in the main comparisons: The main comparisons (table 1 and 2) do not provide error bars, meaning that every reported value could have been lucky initialization. Without the error bars, the main study results cannot be interpreted as evidence for the claimed improvements. For inspiration, related probabilistic flow-matching forecasting works report results over multiple seeds: ("Flow Matching with Gaussian Process Priors for Probabilistic Time Series Forecasting", Kollovieh et al., ICLR 2025). Related, the main result table 1 reports MSE and MAE metrics that in my opinion are not well-suited for a paper that addresses probabilistic forecasting, since these do not reflect the uncertainty of the model.

  4. Experimental: lack of flow-matching baselines: The paper cites several related works (e.g. TSFlow) that also apply flow-matching for probabilistic forecasting, however these are not reflected in the baselines. These would be important additions for a proper claim on improving performance over related works.

As more minor point, figure 2 is missing basic features such as labels on the x and the y axis.

Questions for authors
  1. Are you able to provide error bars for table 2? It would be important to see if the claimed effect can be reproduced over multiple runs, both for your proposed method as well as for the baselines.
  2. Could you include flow-matching baselines for probabilistic forecasting to the main comparison in table 2?
Reviewer_dAbjOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

FlowOC is a largely engineering leaning paper proposing probabilistic forecasting to be a two step process. First, a normal Flow Matching model for probabilistic time series forecasting. Second, small additive (optimal control) to align the forecast and improve calibration with the CRPS as a reward.

Critical strengths
  • Simple, practical receipt with one hyperparameter.
  • The results are competitive with strong deterministic baselines.
Critical weaknesses
  • "However, improving the alignment of generated forecasts with future observations while preserving meaningful predictive variability remains challenging." unclear what the exact reason for wanting this is.
  • This is a simplified VGG-Flow (Liu et al., 2025) with CRPS as the reward.
  • The optimal control is overstated, the implemented method has no value function and no consistency loss, and the controller only acts on each trajectory independently, i.e. see Appendix D (“Comparison with VGG Flow”).
  • The distribution preserving claim is not supported, i.e. control penalty value is effectively of, see ß.
  • Effective guidance strength is very different in the shared hyperparameter setting, e.g. Traffic vs ETTm1 are hundredfold off.
  • The controller only barely improves the model over the base architecture.
  • Calibration is never measured.
  • The historical-patch source has weak evidence and limited novelty. History-informed sources already exist in TSFlow (GP priors) and Lim et al. (2025). Neither is compared to.
Questions for authors

See weaknesses.

Reviewer_3oMzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FlowOC, which adopts a two-stage training framework for conditional time series generation. In the first stage, it trains an initial conditional generative model using conditional flow matching. Specifically, the method randomly samples patches from historical observations and constructs a flow from history-informed source states to future trajectories, thereby learning the corresponding velocity field. In the second stage, FlowOC uses gradients of the CRPS to train an additional control field, which refines the pretrained velocity field. Finally, future trajectories are generated using the calibrated velocity field.

Critical strengths
  1. This paper constructs a new flow using historical information. Specifically, it divides historical observations into patches and randomly samples from these patches, using the sampled historical patches as source states for conditional generation.

  2. This paper uses gradients of the CRPS to fine-tune the velocity field, introducing a relatively novel optimization objective compared with conventional fine-tuning methods.

Critical weaknesses
  1. This paper uses gradients of the CRPS to refine the velocity field. However, the proposed calibration strategy is not fully consistent with the optimal control formulation, and it lacks theoretical guarantees for its effectiveness. Moreover, CRPS gradients may not adequately capture the temporal dependencies and joint distributional structure of time series, which could limit the effectiveness of the calibration. The experimental results also show only marginal improvements over the base model, suggesting that the proposed calibration strategy provides limited benefits.

  2. This paper divides the historical observations into patches and randomly samples from them. However, this sampling strategy may not fully capture the conditional distribution of future trajectories given the complete historical observations. Why would randomly sampling historical patches lead to better performance than using the entire historical sequence for training?

Questions for authors

Please read the weakness

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

Related papers

Discussion (0)

Sign in to comment.