Whole-Route Forecasts: Better Plans for Linked Events
Original title: PathFM: Fast Generative Forecasting of Multivariate Trajectories
Read by Roger, a synthetic voice
PathFM forecasts whole future paths instead of separate points. It draws complete linked futures in short passes. The authors report it matches usual accuracy on standard tests, leads a new whole-path test, and keeps long-run variation better in hard tests.
Our take: The evidence is fairly strong across standard and new tests, but the path test is new and comes from the authors. It matters where plans depend on runs of events, such as energy or supplies.
- Most leading forecasters give separate future odds, leaving links between times unclear.
- PathFM draws complete future paths for linked measures and uses short passes to do so quickly.
- The authors report top whole-path scores, competitive usual accuracy, and steadier long-run swings in hard tests.
12 claims checked against the source Read the paper on arXiv PDF
Checked against the source
Every statement in this brief had to match a word-for-word passage from the source; code checks each match, and a second AI model reviews the whole brief for errors. A brief that fails is not published. These are the 12 passages it rests on.
Show the 12 passages
Many leading forecasting models give separate future odds and leave the links between future times unspecified.
However, most leading models predict marginal quantiles, leaving the dependence structure of future trajectories unspecified.
The lack of links matters for total demand, peak load, and event timing, and usual scores do not reveal it.
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.
The paper's title is PathFM: Fast Generative Forecasting of Multivariate Trajectories.
PATHFM: FASTGENERATIVEFORECASTING OF MULTIVARIATETRAJECTORIES
PathFM uses flow matching to generate complete future paths.
we introduce PathFM, a trajectory-based forecasting model that uses flow matching to generate complete future paths.
PathFM draws complete future paths for linked measures.
The model learns temporal and cross-variable dependence through conditional flow matching (Lipman et al., 2022), generating complete future paths conditioned on history and known future covariates.
PathFM uses short passes to generate multiple paths quickly.
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.
PathFM stays close to usual separate-time accuracy on established benchmarks.
PathFM achieves competitive marginal accuracy on the established forecasting benchmarks fev-bench and TIME.
The authors introduce a new benchmark for path quality.
we introduce Forecast Path Score Bench (fps-bench), which augments fev-bench tasks with energy and variogram scores and downstream metrics.
PathFM had the best overall whole-path score among the evaluated external baselines.
PathFM achieves a higher aggregate path score on fps-bench than all evaluated external baselines.
In chaotic-system tests, PathFM keeps long-run variability while strong baselines collapse.
Experiments on chaotic systems further show that our flow model sustains long-run variability, while strong baselines rapidly collapse.
PathFM does not yet accept non-number inputs such as text.
PathFM does not yet support conditioning on modalities beyond time series, such as text.
PathFM is meant to support practical downstream applications.
PathFM makes generative time series forecasting both competitive and practical, enabling diverse downstream applications.