Reproduce “PathFM: Fast Generative Forecasting of Multivariate Trajectories” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jiachen Yao, Xiyuan Zhang, Annan Yu, Boran Han, Abdul Fatir Ansari, Shuai Zhang, Michael Bohlke-Schneider, Bernie Wang
PathFM generates complete multivariate future paths conditioned on observed history and known future information, using parallel-decoding distillation and shared history computation for sampling. Unlike marginal quantiles, these paths support estimates of totals, maxima, and event times. On 94 fev-bench tasks, PathFM-300M scores 0.5458 on FPS versus 0.4871 for TimesFM-3, while TimesFM-3 leads marginal benchmarks; this does not establish universal superiority.
Key insights
- Models with similar marginal forecast scores can differ in their representation of trajectory structure, so marginal metrics alone may not predict performance on downstream path-dependent tasks.
- PathFM’s native path forecasts outperform tested marginal-forecast completions on fps-bench; fitting a temporal copula improves a quantile-model baseline but does not close the reported gap.
- Distillation and stratified sampling reduce the computation or ensemble size needed for competitive scores, though the benefits are demonstrated on the evaluated benchmarks and tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.