PathFM: Fast Generative Forecasting of Multivariate Trajectories
PathFM generates multiple future paths in roughly the time a quantile model needs for one query, the authors report.
The model uses flow matching to capture dependence across a forecast trajectory, which matters for cumulative demand, peak load and event timing. Distilling sampling to one or two steps and reusing the history attention cache across draws reduces generation time. PathFM reports competitive marginal accuracy on fev-bench and TIME, and the highest aggregate path score among evaluated external baselines on the authors’ new fps-bench.
HF Daily Papers' note
The model uses flow matching to capture dependence across a forecast trajectory, which matters for cumulative demand, peak load and event timing. Distilling sampling to one or two steps and reusing the history attention cache across draws reduces generation time. PathFM reports competitive marginal accuracy on fev-bench and TIME, and the highest aggregate path score among evaluated external baselines on the authors’ new fps-bench.
HF Daily Papers' note
score 5