Forecasting the future is a challenge faced by many industries, from energy management to finance. Traditional models often predict individual points or marginal ranges for future values, but they struggle to capture how these values relate to each other over time. A newly published research paper introduces PathFM, a novel approach that generates entire possible future trajectories of complex time series data quickly and accurately. This advancement could improve decision-making processes that depend on understanding the full range of future scenarios, such as predicting peak demand or timing critical events.
Key Takeaways
- PathFM models the full dependence structure of future time series trajectories, not just individual point forecasts.
- It uses a technique called flow matching to generate multiple complete future paths efficiently, roughly as fast as producing a single quantile forecast.
- The model is distilled to reduce computational steps and shares historical attention information across samples, speeding up the process.
- PathFM outperforms existing forecasting methods on new and established benchmarks, especially in capturing long-term variability in complex systems.
Most existing forecasting models estimate marginal quantiles, which means they predict specific points or ranges for each future time step independently. While this approach provides useful information about likely future values, it does not specify how these values are linked across time — for example, whether a high value at one point makes a high or low value more likely later. This dependence matters greatly when decisions rely on entire trajectories, such as cumulative demand over a day or the timing of a peak event.
PathFM addresses this gap by generating full future paths of multivariate time series data. It does so using flow matching, a technique from generative modeling that learns how to transform simple random noise into realistic data sequences. By training on historical time series, PathFM learns to produce multiple plausible future trajectories that maintain realistic dependencies over time and across variables. To make the model practical for real-world use, the researchers distilled it to require just one or two sampling steps and implemented a shared history attention cache, which reuses information about past data across different generated paths. This dramatically reduces computation time, enabling multiple path samples to be generated as quickly as a single quantile forecast from traditional models.
To evaluate PathFM, the team tested it on established forecasting benchmarks such as fev-bench and TIME, where it delivered competitive accuracy for individual points. Recognizing that pointwise accuracy alone does not capture the quality of full trajectory forecasts, they introduced a new benchmark called Forecast Path Score Bench (fps-bench). This benchmark includes metrics that measure the energy and variogram scores, which assess how well a model captures the overall shape and variability of forecasted paths. PathFM achieved higher aggregate path scores than all other external baseline models on this benchmark. Additional experiments on chaotic systems demonstrated that PathFM sustains realistic long-term variability, whereas other strong models tended to produce overly smooth or collapsed forecasts over time.
The ability to generate multiple realistic future trajectories quickly opens new possibilities for applications that require understanding uncertainty and temporal dependencies in forecasts. For instance, energy grid operators could better predict peak loads and manage resources more effectively. Financial analysts might improve risk assessments by considering a richer set of future market scenarios. While PathFM represents a promising step forward, further research and development will be needed to adapt this approach to various domains and to integrate it with decision-making systems.
Overall, this newly published research highlights how generative modeling techniques like flow matching can make time series forecasting both more comprehensive and practical, enabling a deeper understanding of future dynamics beyond traditional point estimates.
Based on research published on arXiv by Jiachen Yao, Xiyuan Zhang, Annan Yu et al..