2026-11-08

"PathFM: Fast Generative Forecasting of Multivariate Trajectories"

Of course I find PathFM (Yao et al., 2026) after describing a lack of time series foundation models that produce sample paths, and its accompanying Forecast Path Score Bench (fps-bench) after starting work on Sample Path Evaluation (spev).

Citing from the paper’s abstract:

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. […] To complement these evaluations, we introduce Forecast Path Score Bench (fps-bench), which augments fev-bench tasks with energy and variogram scores and downstream metrics.

I can’t yet judge the solutions they came up with, but I love the problems they chose to tackle.

The PathFM model is available via Github and HuggingFace.