PathFM: Fast Generative Forecasting of Multivariate Trajectories via Flow Matching

Forum topic · 小凯 · 2026-10-17

Summary

PathFM is a trajectory-based time series forecasting model that uses flow matching to generate complete future paths rather than marginal quantiles. Most leading time series foundation models predict marginal quantiles, leaving the dependence structure of future trajectories unspecified—a distinction that matters for quantities like cumulative demand, peak load, and event timing. PathFM addresses this by modeling full trajectory dependencies with flow matching. The model is distilled to one or two sampling steps, and history attention cache is shared across draws, allowing multiple paths to be generated in roughly the time of a single quantile-model query. PathFM achieves competitive marginal accuracy on forecasting benchmarks fev-bench and TIME. A new benchmark, fps-bench (forecasting path score), is introduced, where PathFM achieves higher composite path scores than all evaluated baselines. Experiments on chaotic systems show the flow model maintains long-term variability while strong baselines rapidly collapse. The work by Yao et al. (arXiv:2610.22358) makes generative time series forecasting both competitive and practical.

Overview

Field: Machine Learning Authors: Jiachen Yao, Xiyuan Zhang, Annan Yu, Boran Han, Abdul Fatir Ansari, Shuai Zhang, Michael Bohlke-Schneider, Bernie Wang Published: 2026-10-15 arXiv: 2610.22358

Key Points

Time series foundation models are increasingly popular for general-purpose forecasting, but most leading models predict marginal quantiles, leaving the dependence structure of future trajectories unspecified.
This distinction matters for trajectory-dependent quantities such as cumulative demand, peak load, and event timing; point and marginal scores do not reveal it.
PathFM is a trajectory-based forecasting model that uses flow matching to generate complete future paths, modeling the joint dependence of multivariate trajectories.
The model is distilled to one or two sampling steps, and the history attention cache is shared across draws, enabling generation of multiple paths for a single series in roughly the time of one query to a quantile model.
PathFM achieves competitive marginal accuracy on the forecasting benchmarks fev-bench and TIME.
The authors introduce fps-bench (forecasting path score benchmark), where PathFM attains higher composite path scores than all evaluated baselines.
Experiments on chaotic systems show the flow model maintains long-term variability while strong baselines rapidly collapse.

Original Abstract (excerpt)

Time series foundation models are becoming increasingly popular for general-purpose forecasting. However, most leading models predict marginal quantiles, leaving the dependence structure of future trajectories unspecified. This distinction matters for trajectory-dependent quantities such as cumulative demand, peak load, and event timing, yet point and marginal scores do not reveal it. To model this dependence, we introduce PathFM, a trajectory-based forecasting model that uses flow matching to generate complete future paths. We further distill the model to one or two steps and share history attention cache across draws. Together, these enable PathFM to generate multiple paths for a single series in roughly the time needed for one query to a quantile model. The distilled sampler also makes ...

*Auto-collected on 2026-10-17*