PathFM Enables Rapid Generative Forecasting of Multivariate Trajectories
PathFM generates full multivariate paths, improves dependence and path accuracy, ranks high on fev-bench, enables fast ensembles; code coming soon.
Key Takeaways
- PathFM generates complete multivariate trajectories instead of marginal quantiles, better capturing multimodal behavior and temporal or cross-variable dependence.
- It ranks third on fev-bench for marginal accuracy but leads path-forecast accuracy with a 10.6% advantage over TimesFM-3.
- On an NVIDIA L4, PathFM can generate 32 probabilistic paths in just 51.6 milliseconds, enabling fast ensemble forecasting.
- For chaotic systems, it sustains realistic long-run variability 20+ times longer than baselines while achieving the lowest average point-forecast error.
- Model weights and code are coming soon, making PathFM a promising next step for probabilistic time-series foundation models.
Meta description: PathFM generates complete multivariate trajectories, ranking third on fev-bench marginal accuracy and leading path forecasts with a 10.6% advantage.
Tags: PathFM, time-series foundation models, probabilistic forecasting, fev-bench, NVIDIA L4
Background: Time-Series Forecasts Need Full Paths
Existing time-series foundation models commonly predict marginal quantiles. Those outputs do not capture multi-mode behavior and temporal or cross-variable dependence well. PathFM instead generates complete multivariate trajectories, enabling more agile probabilistic forecasts.
Development Process: Benchmarking Accuracy and Speed
PathFM's reported results highlight three benchmark dimensions. For marginal accuracy, it holds 3rd place on fev-bench as of today. For path accuracy, it has the highest path score, with a 10.6% advantage over the strongest baseline, TimesFM-3. For fast ensembles, it can generate 32 paths in just 51.6 milliseconds on an NVIDIA L4.
Latest Progress: Chaotic Systems and Long-Run Variability
On chaotic systems, PathFM sustains long-run variability 20+ times longer than baselines while achieving the lowest average point-forecast error. The comparison points to a practical weakness of marginal-quantile forecasting: realistic dependence across time and variables can matter more than isolated marginal accuracy.
Current Stage: Weights and Code Coming Soon
Model weights and code are coming soon, and the announcement invites thoughts and feedback. The next step to watch is the release itself and any future benchmark detail beyond the current fev-bench ranking and TimesFM-3 comparison.
Comments
No comments yet. Be the first to comment.