PathFM: generating complete time series trajectories through flow matching

Time series foundation models are becoming the standard for universal forecasting, but they leave out critical information about the dependency structure of future trajectories. Most leading solutions — like TimesFM-3 — predict marginal quantiles for each variable separately, without accounting for their joint behavior or temporal correlations. That creates a problem for applied tasks that require a full picture of how things unfold: forecasting aggregate demand, peak network load, or the timing of events.

Autogluon researchers have proposed an alternative approach called PathFM, a generative forecasting model that produces complete multivariate trajectories. Instead of outputting individual points or quantiles, the model uses flow matching to generate realistic paths for how a series develops, preserving all the inherent dependencies and variability.

Architecture and optimization

PathFM relies on flow matching, a method that transforms a simple distribution into a complex one through a reversible process. The key innovation is distillation: the researchers compressed the model down to one or two computation steps and implemented a history attention cache for reuse when generating multiple paths. That made it possible to achieve speeds comparable to a single quantile-model query, but for producing dozens of independent trajectories.

The model is also optimized directly through a calibration loss — a loss function aimed at calibrating uncertainty, something that isn't available to models working with individual quantiles.

Benchmark results

PathFM was evaluated on the standard fev-bench and TIME benchmarks. On fev-bench, the model achieved competitive accuracy for marginal forecasts, placing third. However, its main achievements show up on a new benchmark called Forecast Path Score Bench (fps-bench), which the researchers designed specifically to evaluate trajectory quality.

fps-bench extends fev-bench's tasks with additional metrics: energy score and variogram score, which measure how well a model reproduces the joint distribution and spatial variability. On the aggregated path score, PathFM outperformed every tested solution, including TimesFM-3, by 10.6%.

Speed and practicality

On an NVIDIA L4 GPU, PathFM generates 32 complete paths in 51.6 milliseconds. That makes ensemble methods practical: previously, generative approaches required either expensive computation or compromises on quality.

Behavior on chaotic systems

A dedicated test on chaotic systems showed that the flow-matching approach has a unique property: the model sustains long-term variability 20 times longer or more than baseline methods, while also achieving the lowest average point-forecast error. Other strong models quickly lose randomness and collapse their forecasts toward averaged values.

PathFM opens up a new class of applications: from supply chain management that accounts for the full demand distribution to load management systems in the energy sector, where it's necessary to anticipate not just the average but the entire range of possible scenarios. The authors — Jiachen Yao, Xiyuan Zhang, Annan Yu, Boran Han, Abdul Fatir Ansari, Shuai Zhang, Michael Bohlke-Schneider, and Bernie Wang — promise to release the model weights and source code soon.