TimeFlow: Towards Stochastic-Aware and Efficient Time Series Generation via Flow Matching Modeling
He Panjing, Cheng Mingyue, Li Li, Zhang XiaoHan
TL;DR
TimeFlow introduces a stochastic differential equation-based flow matching method for efficient, high-fidelity time series generation that better models inherent randomness and variability.
Contribution
We propose TimeFlow, a novel SDE-based flow matching framework with an encoder architecture and stochastic components, improving time series generation fidelity and efficiency.
Findings
Outperforms baselines in quality and diversity
Supports both unconditional and conditional generation
Achieves higher efficiency in sampling
Abstract
Generating high-quality time series data has emerged as a critical research topic due to its broad utility in supporting downstream time series mining tasks. A major challenge lies in modeling the intrinsic stochasticity of temporal dynamics, as real-world sequences often exhibit random fluctuations and localized variations. While diffusion models have achieved remarkable success, their generation process is computationally inefficient, often requiring hundreds to thousands of expensive function evaluations per sample. Flow matching has emerged as a more efficient paradigm, yet its conventional ordinary differential equation (ODE)-based formulation fails to explicitly capture stochasticity, thereby limiting the fidelity of generated sequences. By contrast, stochastic differential equation (SDE) are naturally suited for modeling randomness and uncertainty. Motivated by these insights, we…
Peer Reviews
Decision·ICLR 2026 Conference Withdrawn Submission
This paper designed a component-wise decomposed velocity field to capture the multi- faceted structure of time series and seem to contribute to the performance improvement.
This idea to use flow matching has been adopted in FlowTS https://arxiv.org/pdf/2411.07506. The idea is presented much simpler and effective in this paper. So the novelty of this paper is limited. Most importantly, the authors failed to do enough literature review to find the paper published one year ago and did not include any comparisons with this papers. After careful comparing the table 1 of both papers, I think their performances of these two papers are quite similar. Therefore, I think the
- The paper is well written in structure. - The proposed parametrization of the model is reasonable.
In short, the paper lacks novelty and to me seems to misrepresent its contribution. Wrong motivation and falsely claimed contribution: - They argue that Diffusion models (including SDE base score models, where the score matching objective actually motivated the flow matching framework) are inefficient, when compared to the ODE of flow models. Then acknowledge that the ode lacks the capacity to model the stochasticity of time series, leading to their conclusion to turn the ODE into an SDE and s
1. The motivation is clear and the challenges are almost addressed. 2. The paper is easy to follow and less typos. 3. The experimental results show the competitive generation performance and efficiency comparison across the baselines. 4. The process of framework is clear.
1. In the line 266 “Conditional Generation”, there is no observation values as condition in the TimeFlow, how to achieve the training and generation process (Eqn. 8). Can you give the detailed description of Conditional Generation. 2.According to the Figure 5, with the forecasting window longer, the mse is much similar to the baseline, and the growth rate far exceeds the corresponding baseline, which demonstrates the instability of TimeFlow in forecasting task. Can you give some clarification.
1. This paper introduces TimeFlow, an SDE-based FM framework for unconditional time series generation. 2. The paper is well written and easy to follow. 3. Comprehensive experiments are conducted to better understand the performance compared to previous SOTA baselines.
1. The GitHub link to code doesn’t contain the most important files (models, losses, encoder). 2. Based on the ablation study, CA and FD don’t seem to alter the performance so much. Not sure if those modules are necessary. 3. The discussion of FM seems to be lightweight; the story could be more compelling with a detailed comparison between FM and the diffusion model. 4. Minor: In Figure 3a, both pictures have diffusion-ts as legend.
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsTime Series Analysis and Forecasting · Machine Learning in Healthcare · Generative Adversarial Networks and Image Synthesis