FlowState: Sampling-Rate-Equivariant Time-Series Forecasting

A new time-series forecasting model called FlowState achieves sampling-rate-equivariant predictions without retraining, according to a paper posted on arXiv. The architecture pairs a state space model encoder with a functional basis decoder, enabling it to generalize across unseen temporal resolutions. The paper, authored by Lars Graf and submitted on 7 August 2025, describes FlowState as a departure from existing time series foundation models, which are often built on transformer variants. Those models, the authors argue, lack adaptability to different sampling rates and struggle to generalize across varying context and target lengths [2]. FlowState addresses these limitations through a unified design that enables continuous-time modeling and dynamic time-scale adjustment [2]. The model can dynamically adjust forecasting horizons without retraining, a capability the authors describe as inherent to its architecture [2]. The submission history shows the paper was revised twice, with the third version posted on 15 June 2026 at a size of 4,872 KB [1]. The initial submission was 211 KB [1]. FlowState is among the smallest time series foundation models measured by parameter count, yet it achieves state-of-the-art results on the GIFT-Eval benchmark [2]. The authors also report superior adaptability to unseen sampling rates, a property they link to the model's sampling-rate-equivariant design [2]. The architecture's functional basis decoder is central to its ability to operate across temporal resolutions. Unlike convolutional neural networks, which rely on sliding filters to achieve translation-equivariant responses but are not generally invariant to translation due to downsampling operations, FlowState's decoder is designed to maintain equivariance across sampling rates [3]. The state space model encoder further supports this by modeling continuous-time dynamics [2]. The paper appears on arXiv, an open-access repository that hosts electronic preprints across disciplines including computer science and statistics [9]. arXiv does not peer-review submissions but moderates them before posting [9]. The repository surpassed two million articles by the end of 2021 and receives roughly 24,000 submissions per month as of November 2024 [9]. The authors also propose an efficient pretraining strategy that they say improves robustness and accelerates training [2]. Detailed analyses in the paper confirm the effectiveness of the model's components, and the authors demonstrate its ability to adapt to varying input sampling rates [2].