PathFM: Fast Generative Forecasting of Multivariate Trajectories
PathFM generates full forecast trajectories with fast flow-matching sampling and leads a new path-score benchmark.
PathFM is a generative time-series model that uses flow matching to sample complete multivariate future paths rather than only marginal quantiles. Distillation to one or two steps and a shared history attention cache let it draw multiple paths in about the time of one quantile-model query, and the sampler supports direct calibration. It matches competitive marginal accuracy on fev-bench and TIME and scores higher than evaluated baselines on the new fps-bench path metrics. On chaotic systems it sustains long-run variability where strong baselines collapse.
- PathFM uses flow matching to generate full future trajectories, not only marginal quantiles.
- Distillation to one or two steps plus a shared attention cache speeds sampling.
- It is competitive on fev-bench and TIME and leads fps-bench path scores.
- On chaotic systems the flow model keeps long-run variability while baselines collapse.
- AI models
- PathFM
From the original article
1 min read at arxiv.orgTime 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.
By Jiachen Yao · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.