acceptodds
Under review as a conference paper at ICLR 2027

ProtoFlow: Prototype-Guided Flow Matching for Multivariate Time Series Forecasting

Abstract

Generative modeling has shown strong promise for multivariate time-series (MTS) forecasting, especially scale to high-dimensional settings. Diffusion-based methods achieve competitive performance but typically require many sampling steps at inference. VAE-based non-iterative forecasting frameworks have therefore emerged as an efficient alternative. Within this line of work, vector quantization (VQ) enables controllable latent space modeling by mapping multivariate series into compact discrete representations. Existing VQ-based forecasting methods, however, typically rely on autoregressive (AR) token generation, which suffers from exposure bias and training-inference mismatch. Flow matching provides an efficient non-autoregressive alternative for latent forecasting, but existing formulations usually initialize transport from a generic Gaussian prior. We instead observe that the trained VQ codebook already captures representative latent prototypes and can thus serve as a more informative prior for flow matching. Based on this insight, we propose ProtoFlow, a forecasting framework that combines vector-quantized autoencoding with Prototype-prior Flow matching. Our method first maps multivariate sequences into a discrete latent space, then constructs a structured prior from the learned codebook, and finally learns a DiT-based rectified flow to transport samples from this prior to future latent representations conditioned on historical observations. By replacing generic noise initialization with a learned prototype prior, ProtoFlow avoids the rollout mismatch of AR token prediction and promotes faster training convergence. Extensive experiments on benchmark datasets show that it consistently achieves superior forecasting performance with efficient inference.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series, Vector-quantization, Flow Matching

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_rekVOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

ProtoFlow combines vector-quantized representations with non-autoregressive flow matching for high-dimensional multivariate time series forecasting. Its central design constructs a source prior from learned codebook prototypes, empirical frequencies, and local variability. Experiments report competitive forecasting performance with few sampling steps.

Critical strengths

(1)The method provides a clear and practical way to reuse learned codebook structure for generative initialization, extending its role beyond quantization.

(2)Joint latent generation avoids teacher-forced autoregressive rollout and enables forecasting with few ODE steps.

(3)Evaluation covers five relevant high-dimensional benchmarks, with ablations examining latent representations, codebook size, and prototype weighting.

Critical weaknesses

(1)The theoretical analysis assumes that the proposed prior is closer to the target and concerns optimal coupling, whereas training uses independent source–target sampling. It therefore does not establish the claimed practical advantages in optimization or few-step generation.The frequency-weighting comparison in Table 3 uses different sampling budgets across variants, limiting attribution of improvements specifically to empirical frequencies. The benefits of prototype geometry versus normalization and scale matching also remain unclear.

(2)Evaluation is limited to two forecast horizons and primarily aggregate-series metrics. This provides limited evidence regarding individual-variable accuracy, joint dependence, and robustness to shifts in prototype frequencies.

Questions for authors

1 Can you connect the proposed prior to the actual independently sampled training objective, or provide direct evidence that it improves velocity-field learnability or sampling trajectories?

2 How does ProtoFlow perform on per-variable metrics, longer forecast horizons, and test periods whose prototype frequencies differ substantially from training?

Reviewer_5xQhOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ProtoFlow is a two-stage framework for high-dimensional multivariate probabilistic time series forecasting. It learns a vector-quantized tokenizer and decoder. It then constructs a source prior from codebook-associated prototype centers, empirical frequencies, and local perturbation statistics. A history-conditioned rectified flow model jointly generates future latent sequences, which are decoded into forecasts. Experiments cover five datasets, two prediction horizons, comparisons with diffusion- and flow-based forecasting methods, and ablations involving latent representations, generation strategies, codebook sizes, prototype weights, and sampling steps.

Critical strengths

The paper extends the role of the VQ codebook beyond quantization by using its associated structure to construct the flow model’s source distribution. This is a concrete design contribution that could be useful for improving latent generative forecasting.

Critical weaknesses
  1. The estimation of the prototype prior is insufficiently specified, limiting reproducibility. Equations (9)–(11) define a sampling procedure, but the estimation of its parameters remains unclear. Algorithm 2 summarizes these operations as estimating the prior parameters, without providing the necessary computational details. Settings such as the perturbation strength γ\gamma are also not fully specified in the hyperparameter tables. Since prior construction is the main contribution, these details should be explicitly documented rather than left to implementation choices by the reader.

  2. The evaluation does not sufficiently address the high-dimensional multivariate forecasting motivation. Jointly encoding multivariate inputs can, in principle, learn cross-variable dependencies implicitly. But, the paper does not adequately analyze whether VQ compression and the prototype prior preserve those dependencies, nor does it provide direct experimental evidence.If the method is claimed to offer particular advantages in high-dimensional settings, accuracy and computational cost as the number of variables increases should also be examined.

  3. The existing comparisons do not adequately separate the benefit of prototype structure from that of matching the source and target scales.The compatibility between continuous flow outputs and the VQ decoder is not directly evaluated.During Stage 1, the decoder is trained on quantized codebook embeddings. At inference, Algorithm 3 feeds continuous flow-generated latent vectors directly into the decoder without re-quantization. This is computationally valid, but the generated endpoints may differ from the latent inputs encountered during decoder training. Two targeted experiments would clarify this issue:

  • Direct decoding versus decoding after re-quantization. Starting from the same flow-generated endpoints, compare direct decoding with first mapping each latent vector to the codebook using the Stage 1 cosine-similarity assignment rule. Evaluate forecasting accuracy, probabilistic forecasting quality, and sample diversity.
  • Distance from generated endpoints to the nearest codebook vector. Report the mean and distribution of these distances and how they change with the number of ODE steps. Given the unit-norm constraint used during training, angular deviations and norm deviations could also be reported separately. These analyses would establish whether few-step flow outputs remain close to the target code representations and whether performance depends on the decoder generalizing to off-codebook inputs. The current manuscript provides no direct evidence on this point.
Questions for authors
  1. Equations (9)–(11) define the prior through a sampling procedure, but the manuscript does not clearly explain how its parameters are obtained. In particular, is the prototype center μk\mu_k set directly to the codebook vector Zk\mathcal Z_k? How are the local directions BkB_k and scales σk\sigma_k calculated?

  2. The comparison with ProtoFlow-G may introduces a substantial scale difference: prototype samples and target codebook vectors are unit-normalized, whereas the stated standard Gaussian prior has expected squared norm nzn_z per latent position. Without normalized or scale-matched Gaussian controls, the experiments cannot adequately separate the benefit of prototype structure from that of source–target scale alignment. Furthermore, there are no targeted ablations removing the local perturbation or varying γ\gamma. These omissions may affect reproducibility and attribution of the paper’s main contribution.

3.How does performance change when the perturbation is removed, i.e., γ=0\gamma=0, and when γ\gamma is varied? A comparison with isotropic local perturbations would further clarify whether the learned local directions contribute beyond prototype-center sampling.

4.Please report the distance from generated endpoints to their nearest codebook vectors, preferably across sampling budgets. Since target embeddings are unit-normalized, separating angular and norm deviations would help interpret the results.

Reviewer_8yJuOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The work proposes ProtoFlow, a time-series forecasting model. First, it maps the continuous representations in a discrete latent space and then learns a rectified flow to generate those. Further, an additional learned prototype prior is demonstrated. In an empirical evaluation, ProtoFlow outperforms other generative models operating in observation space.

Critical strengths
  • The method is interesting and clearly described. While there are some previous quantization works, the combination with flow matching makes it interesting.
  • The empirical evaluation demonstrates clear performance benefits over other previous diffusion approaches in observation space. Further, the most important baselines and datasets are covered.
  • Important design choices, e.g., prior and codebook size, are empirically ablated.
Critical weaknesses
  • The methodological contribution of this work is rather on the simple side, but sufficient for acceptance.
  • The draft does not seem too polished yet. E.g., there are various citation style mismatches (citet vs citep). Also, while the main figure aids in understanding the methodology, it is too crowded.
  • The theoretical motivation for the codebook initialization does not convince me. Is there any theoretical argument why the assumption in Eq. 21 holds? It is also unclear what Wasserstein distance is shown in Figure 4 and what implications those have on the geometry.
  • Some important related works are missing, e.g., the original diffusion work.
  • The algorithms in the appendix are very high-level and do not provide sufficient details.
Questions for authors
  • In L377, it is mentioned that ProtoFlow has a 4-8x larger memory footprint than TSFlow. Could you explain the reason for this?
  • The codebook size ablation in Fig. 3 shows that increasing the size improves results up to 256. Have you done this ablation using more NFEs? Further, I would suggest that the authors present Table 4, similar to Fig. 3, as a line plot.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.