FlowState: Sampling-Rate‑Equivariant Time‑Series Forecasting

Existing time series foundation models (TSFMs), often based on transformer variants, lack adaptability to different sampling rates, struggle with generalization across varying context and target lengths and are computationally inefficient. We introduce FlowState, a novel TSFM architecture that achieves sampling-rate–equivariant forecasting through a unified design that pairs a state space model (SSM) encoder with a functional basis decoder (FBD). This design enables continuous-time modeling and dynamic time-scale adjustment, allowing FlowState to inherently generalize across all possible temporal resolutions, and dynamically adjust the forecasting horizons without retraining. We further propose an efficient pretraining strategy that improves robustness and accelerates training. Despite being one of the smallest TSFMs, FlowState achieves state-of-the-art results on the widely used GIFT-Eval benchmark, while demonstrating superior adaptability to unseen sampling rates. Our detailed analyses confirm the effectiveness of its components, and we demonstrate its unique ability to adapt to varying input sampling rates.

Time series data, such as energy consumption, traffic flows, or stock prices, is collected at different time intervals (e.g., seconds, minutes, or days). Thus, a crucial capability for broadly applicable time series foundation models (TSFMs) aiming to achieve high performance in real-world tasks is to seamlessly adjust to these different sampling rates. However, most current TSFMs struggle with this and often need to be retrained or specifically adapted, limiting their applicability. In this work, we introduce FlowState, a new type of TSFM designed to handle data across different sampling rates without retraining, making it better suited for real-world tasks. The key idea is to represent the time series as continuous signals instead of a fixed sequences. FlowState uses a combination of a state space model (SSM), which efficiently processes time series with various sampling rates, and a novel functional basis decoder (FBD), which produces a continuous forecast that can be sampled at any desired rate. Along with the model, we also propose a training strategy that exposes the model to many different context lengths simultaneously, improving robustness and efficiency. Across a large benchmark of 97 datasets, FlowState achieves state-of-the-art performance while using significantly fewer parameters than other state-of-the-art models. In addition, it shows strong robustness when evaluated on unseen sampling rates, where other models typically struggle. Overall, our results suggest that explicitly modeling time as a continuous quantity can lead to more efficient, adaptable, and general-purpose forecasting models.